Forward backend options during streaming speech generation (#1389) (#1390)

Switch SQLite ANN connections when load() is given a new path

.github/workflows/build.yml

@@ -10,42 +10,43 @@ jobs:
       matrix:
         os: [ubuntu-latest, macos-latest, windows-latest]
 
-    timeout-minutes: 60
+    timeout-minutes: 75
     steps:
       - name: Checkout code
-        uses: actions/checkout@v3
+        uses: actions/checkout@v6
 
       - name: Install Python
-        uses: actions/setup-python@v4
+        uses: actions/setup-python@v6
         with:
-          python-version: 3.9
+          python-version: "3.11"
 
       - name: Install Java
-        uses: actions/setup-java@v3
+        uses: actions/setup-java@v5
         with:
           distribution: "zulu"
-          java-version: 17
+          java-version: 21
 
       - name: Install dependencies - Linux
-        run: sudo apt-get update && sudo apt-get install libportaudio2 libsndfile1
+        run: |
+          sudo apt-get update
+          sudo apt-get install libportaudio2 libsndfile1 libegl1 libgles2
+          sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc
         if: matrix.os == 'ubuntu-latest'
 
       - name: Install dependencies - macOS
         run: |
-          echo "OMP_NUM_THREADS=1" >> $GITHUB_ENV
           echo "PYTORCH_MPS_DISABLE=1" >> $GITHUB_ENV
           echo "LLAMA_NO_METAL=1" >> $GITHUB_ENV
           echo "TIKA_STARTUP_SLEEP=30" >> $GITHUB_ENV
           echo "TIKA_STARTUP_MAX_RETRY=10" >> $GITHUB_ENV
           brew install portaudio
+          pip install torch==2.12.1 torchvision==0.27.1
         if: matrix.os == 'macos-latest'
 
       - name: Install dependencies - Windows
         run: |
           "PYTHONIOENCODING=utf-8" >> $env:GITHUB_ENV
           choco install wget
-          pip install -U wheel
-          pip install llama-cpp-python==0.3.2 --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cpu
         if: matrix.os == 'windows-latest'
 
       - name: Build
@@ -56,12 +57,20 @@ jobs:
           python -c "import nltk; nltk.download(['punkt', 'punkt_tab', 'averaged_perceptron_tagger_eng'])"
           python --version
           make data coverage
+        env:
+          HF_HUB_ETAG_TIMEOUT: 100
+          HF_HUB_DOWNLOAD_TIMEOUT: 100
+          HF_XET_CHUNK_CACHE_SIZE_BYTES: 0
+          DOCLING_DEVICE: "cpu"
+          DOCLING_INFERENCE_COMPILE_TORCH_MODELS: false
 
-      - uses: pre-commit/[email protected].0
+      - uses: pre-commit/[email protected].1
         if: matrix.os == 'ubuntu-latest'
 
       - name: Test Coverage
-        run: coveralls --service=github
+        run: |
+          coverage report --fail-under=100
+          coveralls --service=github
         if: matrix.os == 'ubuntu-latest'
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/docs.yml

@@ -7,12 +7,19 @@ jobs:
   deploy:
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v3
-      - uses: actions/setup-python@v4
+      - name: Checkout code
+        uses: actions/checkout@v6
+
+      - name: Install Python
+        uses: actions/setup-python@v6
         with:
-          python-version: 3.9
-      - run: |
+          python-version: "3.11"
+
+      - name: Install txtai
+        run: |
           pip install -U pip wheel
           pip install .[all,dev]
-      - run: mkdocs gh-deploy --force
+
+      - name: Deploy documentation
+        run: mkdocs gh-deploy --force
  

.github/workflows/minimal.yml

@@ -0,0 +1,22 @@
+# GitHub Actions minimal build
+name: minimal
+
+on: ["push", "pull_request"]
+
+jobs:
+  deploy:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v6
+
+      - name: Install Python
+        uses: actions/setup-python@v6
+        with:
+          python-version: "3.11"
+
+      - name: Minimal install
+        run: |
+          pip install -U pip wheel
+          MINIMAL=1 pip install .
+          python -c "import txtai"

.pylintrc

@@ -12,6 +12,8 @@ max-locals=40
 max-returns=10
 max-attributes=20
 min-public-methods=0
+max-public-methods=80
 
 [FORMAT]
 max-line-length=150
+max-module-lines=1500

CITATION.cff

@@ -1,8 +1,8 @@
 cff-version: 1.2.0
 date-released: 2020-08-11
 message: "If you use this software, please cite it as below."
-title: "txtai: the all-in-one embeddings database"
-abstract: "txtai is an all-in-one open-source embeddings database for semantic search, LLM orchestration and language model workflows"
+title: "txtai: the all-in-one AI framework"
+abstract: "txtai is an all-in-one open-source AI framework for semantic search, LLM orchestration and language model workflows"
 url: "https://github.com/neuml/txtai"
 authors:
 - family-names: "Mezzetti"

Makefile

@@ -42,6 +42,7 @@ coverage:
 	coverage run -m unittest discover -v -k testdatabase -s ${TEST_DIR}
 	coverage run -m unittest discover -v -k testembeddings -s ${TEST_DIR}
 	coverage run -m unittest discover -v -k testgraph -s ${TEST_DIR}
+	coverage run -m unittest discover -v -k testlibrary -s ${TEST_DIR}
 	coverage run -m unittest discover -v -k testmodels -s ${TEST_DIR}
 	coverage run -m unittest discover -v -k testoptional -s ${TEST_DIR}
 	coverage run -m unittest discover -v -k testpipeline.testaudio -s ${TEST_DIR}

README.md

@@ -3,7 +3,7 @@
 </p>
 
 <p align="center">
-    <b>All-in-one embeddings database</b>
+    <b>All-in-one AI framework</b>
 </p>
 
 <p align="center">
@@ -16,7 +16,7 @@
     <a href="https://github.com/neuml/txtai/issues">
         <img src="https://img.shields.io/github/issues/neuml/txtai.svg?style=flat&color=success" alt="GitHub issues"/>
     </a>
-    <a href="https://join.slack.com/t/txtai/shared_invite/zt-1cagya4yf-DQeuZbd~aMwH5pckBU4vPg">
+    <a href="https://join.slack.com/t/txtai/shared_invite/zt-37c1zfijp-Y57wMty6YOx_hyIHEQvQJA">
         <img src="https://img.shields.io/badge/slack-join-blue?style=flat&logo=slack&logocolor=white" alt="Join Slack"/>
     </a>
     <a href="https://github.com/neuml/txtai/actions?query=workflow%3Abuild">
@@ -27,12 +27,12 @@
     </a>
 </p>
 
-txtai is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.
+txtai is an all-in-one AI framework for semantic search, LLM orchestration and language model workflows.
 
 ![architecture](https://raw.githubusercontent.com/neuml/txtai/master/docs/images/architecture.png#gh-light-mode-only)
 ![architecture](https://raw.githubusercontent.com/neuml/txtai/master/docs/images/architecture-dark.png#gh-dark-mode-only)
 
-Embeddings databases are a union of vector indexes (sparse and dense), graph networks and relational databases.
+The key component of txtai is an embeddings database, which is a union of vector indexes (sparse and dense), graph networks and relational databases.
 
 This foundation enables vector search and/or serves as a powerful knowledge source for large language model (LLM) applications.
 
@@ -45,13 +45,17 @@ Summary of txtai features:
 - 💡 Pipelines powered by language models that run LLM prompts, question-answering, labeling, transcription, translation, summarization and more
 - ↪️️ Workflows to join pipelines together and aggregate business logic. txtai processes can be simple microservices or multi-model workflows.
 - 🤖 Agents that intelligently connect embeddings, pipelines, workflows and other agents together to autonomously solve complex problems
-- ⚙️ Build with Python or YAML. API bindings available for [JavaScript](https://github.com/neuml/txtai.js), [Java](https://github.com/neuml/txtai.java), [Rust](https://github.com/neuml/txtai.rs) and [Go](https://github.com/neuml/txtai.go).
+- ⚙️ Web and Model Context Protocol (MCP) APIs. Bindings available for [JavaScript](https://github.com/neuml/txtai.js), [Java](https://github.com/neuml/txtai.java), [Rust](https://github.com/neuml/txtai.rs) and [Go](https://github.com/neuml/txtai.go).
 - 🔋 Batteries included with defaults to get up and running fast
 - ☁️ Run local or scale out with container orchestration
 
-txtai is built with Python 3.9+, [Hugging Face Transformers](https://github.com/huggingface/transformers), [Sentence Transformers](https://github.com/UKPLab/sentence-transformers) and [FastAPI](https://github.com/tiangolo/fastapi). txtai is open-source under an Apache 2.0 license.
+txtai is built with Python 3.11+, [Hugging Face Transformers](https://github.com/huggingface/transformers), [Sentence Transformers](https://github.com/UKPLab/sentence-transformers) and [FastAPI](https://github.com/tiangolo/fastapi). txtai is open-source under an Apache 2.0 license.
 
-*Interested in an easy and secure way to run hosted txtai applications? Then join the [txtai.cloud](https://txtai.cloud) preview to learn more.*
+> [!NOTE]
+>
+> [NeuML](https://neuml.com) is the company behind txtai and we provide AI consulting services around our stack. [Schedule a meeting](https://cal.com/neuml/intro) or [send a message](mailto:[email protected]) to learn more.
+>
+> We're also building an easy and secure way to run hosted txtai applications with [txtai.cloud](https://txtai.cloud).
 
 ## Why txtai?
 
@@ -87,7 +91,7 @@ curl -X GET "http://localhost:8000/search?query=positive"
 
 ## Use Cases
 
-The following sections introduce common txtai use cases. A comprehensive set of over 60 [example notebooks and applications](https://neuml.github.io/txtai/examples) are also available.
+The following sections introduce common txtai use cases. A comprehensive set of over 70 [example notebooks and applications](https://neuml.github.io/txtai/examples) are also available.
 
 ### Semantic Search
 
@@ -130,14 +134,15 @@ Agents connect embeddings, pipelines, workflows and other agents together to aut
 
 ![agent](https://raw.githubusercontent.com/neuml/txtai/master/docs/images/agent.png)
 
-txtai agents are built on top of the Transformers Agent framework. This supports all LLMs txtai supports (Hugging Face, llama.cpp, OpenAI / Claude / AWS Bedrock via LiteLLM).
+txtai agents are built on top of the [smolagents](https://github.com/huggingface/smolagents) framework. This supports all LLMs txtai supports (Hugging Face, llama.cpp, OpenAI / Claude / AWS Bedrock via LiteLLM). Agent prompting with [`agents.md`](https://github.com/agentsmd/agents.md) and [`skill.md`](https://agentskills.io/specification) are also supported.
 
-See the link below to learn more.
+Check out this [Agent Quickstart Example](https://github.com/neuml/txtai/blob/master/examples/agent_quickstart.py). Additional examples are listed below.
 
 | Notebook  | Description  |       |
 |:----------|:-------------|------:|
-| [Analyzing Hugging Face Posts with Graphs and Agents](https://github.com/neuml/txtai/blob/master/examples/68_Analyzing_Hugging_Face_Posts_with_Graphs_and_Agents.ipynb) | Explore a rich dataset with Graph Analysis and Agents | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/68_Analyzing_Hugging_Face_Posts_with_Graphs_and_Agents.ipynb) |
 | [Granting autonomy to agents](https://github.com/neuml/txtai/blob/master/examples/69_Granting_autonomy_to_agents.ipynb) | Agents that iteratively solve problems as they see fit | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/69_Granting_autonomy_to_agents.ipynb) |
+| [TxtAI got skills](https://github.com/neuml/txtai/blob/master/examples/83_TxtAI_got_skills.ipynb) | Integrate skill.md files with your agent | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/83_TxtAI_got_skills.ipynb) |
+| [Agent Tools](https://github.com/neuml/txtai/blob/master/examples/84_Agent_Tools.ipynb) [▶️](https://www.youtube.com/watch?v=RDNaFXQy3GQ) | Learn about the txtai agent toolkit | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/84_Agent_Tools.ipynb) |
 | [Analyzing LinkedIn Company Posts with Graphs and Agents](https://github.com/neuml/txtai/blob/master/examples/71_Analyzing_LinkedIn_Company_Posts_with_Graphs_and_Agents.ipynb) | Exploring how to improve social media engagement with AI | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/71_Analyzing_LinkedIn_Company_Posts_with_Graphs_and_Agents.ipynb) |
 
 #### Retrieval augmented generation
@@ -147,13 +152,13 @@ Retrieval augmented generation (RAG) reduces the risk of LLM hallucinations by c
 ![rag](https://raw.githubusercontent.com/neuml/txtai/master/docs/images/rag.png#gh-light-mode-only)
 ![rag](https://raw.githubusercontent.com/neuml/txtai/master/docs/images/rag-dark.png#gh-dark-mode-only)
 
-A novel feature of txtai is that it can provide both an answer and source citation.
+Check out this [RAG Quickstart Example](https://github.com/neuml/txtai/blob/master/examples/rag_quickstart.py). Additional examples are listed below.
 
 | Notebook  | Description  |       |
 |:----------|:-------------|------:|
-| [Build RAG pipelines with txtai](https://github.com/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) | Guide on retrieval augmented generation including how to create citations | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) |
-| [Chunking your data for RAG](https://github.com/neuml/txtai/blob/master/examples/73_Chunking_your_data_for_RAG.ipynb) | Extract, chunk and index content for effective retrieval | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/73_Chunking_your_data_for_RAG.ipynb) |
-| [Advanced RAG with graph path traversal](https://github.com/neuml/txtai/blob/master/examples/58_Advanced_RAG_with_graph_path_traversal.ipynb) | Graph path traversal to collect complex sets of data for advanced RAG | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/58_Advanced_RAG_with_graph_path_traversal.ipynb) |
+| [Build RAG pipelines with txtai](https://github.com/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) [▶️](https://www.youtube.com/watch?v=t_OeAc8NVfQ) | Guide on retrieval augmented generation including how to create citations | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) |
+| [RAG is more than Vector Search](https://github.com/neuml/txtai/blob/master/examples/79_RAG_is_more_than_Vector_Search.ipynb) | Context retrieval via Web, SQL and other sources | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/79_RAG_is_more_than_Vector_Search.ipynb) |
+| [GraphRAG with Wikipedia and GPT OSS](https://github.com/neuml/txtai/blob/master/examples/77_GraphRAG_with_Wikipedia_and_GPT_OSS.ipynb) | Deep graph search powered RAG | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/77_GraphRAG_with_Wikipedia_and_GPT_OSS.ipynb) |
 | [Speech to Speech RAG](https://github.com/neuml/txtai/blob/master/examples/65_Speech_to_Speech_RAG.ipynb) [▶️](https://www.youtube.com/watch?v=tH8QWwkVMKA) | Full cycle speech to speech workflow with RAG | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/65_Speech_to_Speech_RAG.ipynb) |
 
 ### Language Model Workflows
@@ -165,6 +170,8 @@ Language model workflows, also known as semantic workflows, connect language mod
 
 While LLMs are powerful, there are plenty of smaller, more specialized models that work better and faster for specific tasks. This includes models for extractive question-answering, automatic summarization, text-to-speech, transcription and translation.
 
+Check out this [Workflow Quickstart Example](https://github.com/neuml/txtai/blob/master/examples/workflow_quickstart.py). Additional examples are listed below.
+
 | Notebook  | Description  |       |
 |:----------|:-------------|------:|
 | [Run pipeline workflows](https://github.com/neuml/txtai/blob/master/examples/14_Run_pipeline_workflows.ipynb) [▶️](https://www.youtube.com/watch?v=UBMPDCn1gEU) | Simple yet powerful constructs to efficiently process data | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/14_Run_pipeline_workflows.ipynb) |
@@ -183,9 +190,9 @@ The easiest way to install is via pip and PyPI
 pip install txtai
 ```
 
-Python 3.9+ is supported. Using a Python [virtual environment](https://docs.python.org/3/library/venv.html) is recommended.
+Python 3.11+ is supported. Using a Python [virtual environment](https://docs.python.org/3/library/venv.html) is recommended.
 
-See the detailed [install instructions](https://neuml.github.io/txtai/install) for more information covering [optional dependencies](https://neuml.github.io/txtai/install/#optional-dependencies), [environment specific prerequisites](https://neuml.github.io/txtai/install/#environment-specific-prerequisites), [installing from source](https://neuml.github.io/txtai/install/#install-from-source), [conda support](https://neuml.github.io/txtai/install/#conda) and how to [run with containers](https://neuml.github.io/txtai/cloud).
+See the detailed [install instructions](https://neuml.github.io/txtai/install) for more information covering [optional dependencies](https://neuml.github.io/txtai/install/#optional-dependencies), [environment specific prerequisites](https://neuml.github.io/txtai/install/#environment-specific-prerequisites), [installing from source](https://neuml.github.io/txtai/install/#install-from-source), [conda support](https://neuml.github.io/txtai/install/#conda), [lightweight minimal installation](https://neuml.github.io/txtai/install/#minimal-install) and how to [run with containers](https://neuml.github.io/txtai/cloud).
 
 ## Model guide
 
@@ -197,9 +204,9 @@ See the table below for the current recommended models. These models all allow c
 | ----------------------------------------------------------------------------- | ------------------------------------------------------------------------ |
 | [Embeddings](https://neuml.github.io/txtai/embeddings)                        | [all-MiniLM-L6-v2](https://hf.co/sentence-transformers/all-MiniLM-L6-v2) | 
 | [Image Captions](https://neuml.github.io/txtai/pipeline/image/caption)        | [BLIP](https://hf.co/Salesforce/blip-image-captioning-base)              |
-| [Labels - Zero Shot](https://neuml.github.io/txtai/pipeline/text/labels)      | [BART-Large-MNLI](https://hf.co/facebook/bart-large)                     |
+| [Labels - Zero Shot](https://neuml.github.io/txtai/pipeline/text/labels)      | [DeBERTa v3 Zeroshot](https://hf.co/MoritzLaurer/deberta-v3-base-zeroshot-v2.0-c) |
 | [Labels - Fixed](https://neuml.github.io/txtai/pipeline/text/labels)          | Fine-tune with [training pipeline](https://neuml.github.io/txtai/pipeline/train/trainer)          |
-| [Large Language Model (LLM)](https://neuml.github.io/txtai/pipeline/text/llm) | [Llama 3.1 Instruct](https://hf.co/meta-llama/Llama-3.1-8B-Instruct)     |
+| [Large Language Model (LLM)](https://neuml.github.io/txtai/pipeline/text/llm) | [Gemma 4 31B](https://hf.co/google/gemma-4-31B) |
 | [Summarization](https://neuml.github.io/txtai/pipeline/text/summary)          | [DistilBART](https://hf.co/sshleifer/distilbart-cnn-12-6)                |
 | [Text-to-Speech](https://neuml.github.io/txtai/pipeline/audio/texttospeech)   | [ESPnet JETS](https://hf.co/NeuML/ljspeech-jets-onnx)                    |
 | [Transcription](https://neuml.github.io/txtai/pipeline/audio/transcription)   | [Whisper](https://hf.co/openai/whisper-base)                             | 
@@ -211,9 +218,8 @@ See the following links to learn more.
 
 - [Hugging Face Tasks](https://hf.co/tasks)
 - [Hugging Face Model Hub](https://hf.co/models)
-- [MTEB Leaderboard](https://hf.co/spaces/mteb/leaderboard)
-- [LMSYS LLM Leaderboard](https://chat.lmsys.org/?leaderboard)
-- [Open LLM Leaderboard](https://hf.co/spaces/HuggingFaceH4/open_llm_leaderboard)
+- [Embeddings Leaderboard](https://hf.co/spaces/mteb/leaderboard)
+- [LLM Leaderboard](https://hf.co/spaces/lmarena-ai/arena-leaderboard)
 
 ## Powered by txtai
 
@@ -224,8 +230,8 @@ The following applications are powered by txtai.
 | Application  | Description  |
 |:------------ |:-------------|
 | [rag](https://github.com/neuml/rag) | Retrieval Augmented Generation (RAG) application |
-| [ragdata](https://github.com/neuml/ragdata) | Build knowledge bases for RAG |
-| [paperai](https://github.com/neuml/paperai) | Semantic search and workflows for medical/scientific papers |
+| [ncoder](https://github.com/neuml/ncoder) | Open-Source AI coding agent |
+| [paperai](https://github.com/neuml/paperai) | AI for medical and scientific papers |
 | [annotateai](https://github.com/neuml/annotateai) | Automatically annotate papers with LLMs |
 
 In addition to this list, there are also many other [open-source projects](https://github.com/neuml/txtai/network/dependents), [published research](https://scholar.google.com/scholar?q=txtai&hl=en&as_ylo=2022) and closed proprietary/commercial projects that have built on txtai in production.
@@ -235,9 +241,10 @@ In addition to this list, there are also many other [open-source projects](https
 ![further](https://raw.githubusercontent.com/neuml/txtai/master/docs/images/further.png#gh-light-mode-only)
 ![further](https://raw.githubusercontent.com/neuml/txtai/master/docs/images/further-ghdark.png#gh-dark-mode-only)
 
-- [Introducing txtai, the all-in-one embeddings database](https://medium.com/neuml/introducing-txtai-the-all-in-one-embeddings-database-c721f4ff91ad)
+- [Introducing txtai, the all-in-one AI framework](https://medium.com/neuml/introducing-txtai-the-all-in-one-ai-framework-0660ecfc39d7)
+- [txtai: An All-in-One AI Framework for Semantic Search and LLM Workflows](https://github.com/neuml/papers/blob/master/txtai/txtai.pdf)
 - [Tutorial series on Hashnode](https://neuml.hashnode.dev/series/txtai-tutorial) | [dev.to](https://dev.to/neuml/tutorial-series-on-txtai-ibg)
-- [What's new in txtai 8.0](https://medium.com/neuml/whats-new-in-txtai-8-0-2d7d0ab4506b) | [7.0](https://medium.com/neuml/whats-new-in-txtai-7-0-855ad6a55440) | [6.0](https://medium.com/neuml/whats-new-in-txtai-6-0-7d93eeedf804) | [5.0](https://medium.com/neuml/whats-new-in-txtai-5-0-e5c75a13b101) | [4.0](https://medium.com/neuml/whats-new-in-txtai-4-0-bbc3a65c3d1c)
+- [What's new in txtai 9.0](https://medium.com/neuml/whats-new-in-txtai-9-0-d522bb150afa) | [8.0](https://medium.com/neuml/whats-new-in-txtai-8-0-2d7d0ab4506b) | [7.0](https://medium.com/neuml/whats-new-in-txtai-7-0-855ad6a55440) | [6.0](https://medium.com/neuml/whats-new-in-txtai-6-0-7d93eeedf804) | [5.0](https://medium.com/neuml/whats-new-in-txtai-5-0-e5c75a13b101) | [4.0](https://medium.com/neuml/whats-new-in-txtai-4-0-bbc3a65c3d1c)
 - [Getting started with semantic search](https://medium.com/neuml/getting-started-with-semantic-search-a9fd9d8a48cf) | [workflows](https://medium.com/neuml/getting-started-with-semantic-workflows-2fefda6165d9) | [rag](https://medium.com/neuml/getting-started-with-rag-9a0cca75f748)
 - [Running txtai at scale](https://medium.com/neuml/running-at-scale-with-txtai-71196cdd99f9)
 - [Vector search & RAG Landscape: A review with txtai](https://medium.com/neuml/vector-search-rag-landscape-a-review-with-txtai-a7f37ad0e187)

docker/base/Dockerfile

@@ -1,5 +1,5 @@
 # Set base image
-ARG BASE_IMAGE=python:3.9-slim
+ARG BASE_IMAGE=python:3.11-slim
 FROM $BASE_IMAGE
 
 # Install GPU-enabled version of PyTorch if set
@@ -8,7 +8,7 @@ ARG GPU
 # Target CPU architecture
 ARG TARGETARCH
 
-# Set Python version (i.e. 3, 3.9)
+# Set Python version (i.e. 3, 3.11)
 ARG PYTHON_VERSION=3
 
 # List of txtai components to install
@@ -21,18 +21,18 @@ ENV LANG=C.UTF-8
 RUN \
     # Install required packages
     apt-get update && \
-    apt-get -y --no-install-recommends install libgomp1 libportaudio2 libsndfile1 gcc g++ python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python3-pip && \
+    apt-get -y --no-install-recommends install libgomp1 libportaudio2 libsndfile1 libegl1 libgles2 libvulkan1 git gcc g++ python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python3-pip && \
     rm -rf /var/lib/apt/lists && \
     \
     # Install txtai project and dependencies
     ln -s /usr/bin/python${PYTHON_VERSION} /usr/bin/python && \
     python -m pip install --no-cache-dir -U pip wheel setuptools && \
-    if [ -z ${GPU} ] && { [ -z ${TARGETARCH} ] || [ ${TARGETARCH} = "amd64" ] ;}; then pip install --no-cache-dir torch==2.6.0+cpu torchvision==0.21.0+cpu -f https://download.pytorch.org/whl/torch -f https://download.pytorch.org/whl/torchvision; fi && \
+    if [ -z ${GPU} ]; then pip install --no-cache-dir torch==2.14.1+cpu torchvision==0.29.1+cpu -f https://download.pytorch.org/whl/torch -f https://download.pytorch.org/whl/torchvision; fi && \
     python -m pip install --no-cache-dir txtai${COMPONENTS} && \
     python -c "import sys, importlib.util as util; 1 if util.find_spec('nltk') else sys.exit(); import nltk; nltk.download(['punkt', 'punkt_tab', 'averaged_perceptron_tagger_eng'])" && \
     \
     # Cleanup build packages
-    apt-get -y purge gcc g++ python${PYTHON_VERSION}-dev && apt-get -y autoremove
+    apt-get -y purge git gcc g++ python${PYTHON_VERSION}-dev && apt-get -y autoremove
 
 # Set default working directory
 WORKDIR /app

docker/minimal/Dockerfile

@@ -0,0 +1,20 @@
+# Set base image
+ARG BASE_IMAGE=python:3.11-slim
+FROM $BASE_IMAGE
+
+# Target CPU architecture
+ARG TARGETARCH
+
+# Set Python version (i.e. 3, 3.11)
+ARG PYTHON_VERSION=3
+
+# Locale environment variables
+ENV LC_ALL=C.UTF-8
+ENV LANG=C.UTF-8
+
+RUN \
+    # Install minimal txtai project and dependencies
+    MINIMAL=1 python -m pip install --no-cache-dir txtai_minimal
+
+# Set default working directory
+WORKDIR /app

docs/agent/configuration.md

@@ -2,7 +2,7 @@
 
 An agent takes two main arguments, an LLM and a list of tools.
 
-The txtai agent framework is built with [Transformers Agents](https://huggingface.co/docs/transformers/en/agents) and additional options can be directly passed in the `Agent` constructor.
+The txtai agent framework is built with [smolagents](https://github.com/huggingface/smolagents). Additional options can be passed in the `Agent` constructor.
 
 ```python
 from datetime import datetime
@@ -34,18 +34,20 @@ def today() -> str:
     return datetime.today().isoformat()
 
 agent = Agent(
+    model="Qwen/Qwen3-4B-Instruct-2507",
     tools=[today, wikipedia, arxiv, "websearch"],
-    llm="hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
 )
 ```
 
-## llm
+## model
 
 ```yaml
-llm: string|llm instance
+model: string|llm instance
 ```
 
-LLM path or LLM pipeline instance. See the [LLM pipeline](../../pipeline/text/llm) for more information.
+LLM model path or LLM pipeline instance. The `llm` parameter is also supported for backwards compatibility.
+
+See the [LLM pipeline](../../pipeline/text/llm) for more information.
 
 ## tools
 
@@ -77,18 +79,63 @@ Embeddings indexes have built-in support. Provide the following dictionary confi
 | description | embeddings index description               | 
 | **kwargs    | Parameters to pass to [embeddings.load](../../embeddings/methods/#txtai.embeddings.Embeddings.load) |
 
-### transformers
+### tool
 
-A Transformers tool instance can be provided. Additionally, the following strings load tools directly from Transformers.
+The following shortcut strings load tools directly. Passing a Tool instance is also supported.
 
 | Tool        | Description                                               |
 |:------------|:----------------------------------------------------------|
-| websearch   | Runs a websearch using built-in Transformers Agent tool   |
+| bash        | Runs a shell command through subprocess                   |
+| defaults    | Loads all of these tools as the default toolkit           |
+| edit        | Edits a file in place and returns a diff                  |
+| glob        | Finds matching file patterns in a directory               |
+| grep        | Finds matching file content in a directory                |
+| http.*      | HTTP Path to a [Model Context Protocol (MCP)](https://modelcontextprotocol.io/docs/getting-started/intro) server |
+| python      | Runs a Python action                                      |
+| read        | Reads file or url content, supports text extraction       |
+| todowrite   | Generates a task list to organize complex tasks           |
+| websearch   | Runs a websearch using the built-in websearch tool        |
+| webview     | Extracts content from a web page. Alias for `read` tool   |
+| write       | Writes content to file                                    |
+| *.md        | Loads a [`skill.md`](https://agentskills.io/specification) file |
+
+## instructions
+
+```yaml
+instructions: string|path
+```
+
+Supports loading an `agents.md` file. Can be provided directly as a string or as a path to a file.
+
+[Read more about agents.md here](https://github.com/agentsmd/agents.md)
+
+## template
+
+```yaml
+template: string
+```
+
+Customize the prompt template used by this agent. Supports Jinja templates. Uses a default template when this parameter is not provided.
+Must include `{{ text }}` and `{{ memory }}` placeholders.
+
+## memory
+
+```yaml
+memory: int
+```
+
+Keeps a rolling window of `memory` inputs and outputs. These are added to future prompts and serve as "agent memory".
+
+Supports storing memory by `session` to enable multiple conversation threads. Defaults to shared memory when not set. See the [method documentation](../methods#txtai.agent.base.Agent.__call__) for more information.
 
 ## method
 
 ```yaml
-method: reactjson|reactcode|code
+method: code|tool
 ```
 
-Sets the agent method. Defaults to `reactjson`. [Read more on this here](https://huggingface.co/docs/transformers/en/agents#types-of-agents).
+Sets the agent method. Supports either a `code` or `tool` (default) calling agent. A code agent generates Python code and executes that. A tool calling agent generates JSON blocks and calls the agents within those blocks.
+
+Additional options can be directly passed. See [CodeAgent](https://huggingface.co/docs/smolagents/main/en/reference/agents#smolagents.CodeAgent) or [ToolCallingAgent](https://huggingface.co/docs/smolagents/main/en/reference/agents#smolagents.ToolCallingAgent) for a list of parameters.
+
+[Read more here](https://huggingface.co/docs/smolagents/main/en/guided_tour).

docs/agent/index.md

@@ -5,6 +5,8 @@
 An agent automatically creates workflows to answer multi-faceted user requests. Agents iteratively prompt and/or interface with tools to
 step through a process and ultimately come to an answer for a request.
 
+Agent prompting with [`agents.md`](https://github.com/agentsmd/agents.md) and [`skill.md`](https://agentskills.io/specification) are also supported. [Read the configuration](./configuration/#tool) for more on how to setup those up.
+
 Agents excel at complex tasks where multiple tools and/or methods are required. They incorporate a level of randomness similar to different
 people working on the same task. When the request is simple and/or there is a rule-based process, other methods such as RAG and Workflows
 should be explored.
@@ -41,9 +43,9 @@ def today() -> str:
     return datetime.today().isoformat()
 
 agent = Agent(
+    model="Qwen/Qwen3-4B-Instruct-2507",
     tools=[today, wikipedia, arxiv, "websearch"],
-    llm="hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-    max_iterations=10,
+    max_steps=10,
 )
 ```
 
@@ -83,34 +85,35 @@ Agents can also be tools. This enables the concept of building "Agent Teams" to
 ```python
 from txtai import Agent, LLM
 
-llm = LLM("hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4")
+llm = LLM("Qwen/Qwen3-4B-Instruct-2507")
 
 websearcher = Agent(
+    model=llm,
     tools=["websearch"],
-    llm=llm,
 )
 
 wikiman = Agent(
+    model=llm,
     tools=[{
         "name": "wikipedia",
         "description": "Searches a Wikipedia database",
         "provider": "huggingface-hub",
         "container": "neuml/txtai-wikipedia"
     }],
-    llm=llm,
 )
 
 researcher = Agent(
+    model=llm,
     tools=[{
         "name": "arxiv",
         "description": "Searches a database of scientific papers",
         "provider": "huggingface-hub",
         "container": "neuml/txtai-arxiv"
     }],
-    llm=llm,
 )
 
 agent = Agent(
+    model=llm,
     tools=[{
         "name": "websearcher",
         "description": "I run web searches, there is no answer a web search can't solve!",
@@ -124,8 +127,7 @@ agent = Agent(
         "description": "I'm a science guy. I search arXiv to get all my answers.",
         "target": researcher
     }],
-    llm=llm,
-    max_iterations=10
+    max_steps=10
 )
 ```
 
@@ -133,15 +135,14 @@ This provides another level of intelligence to the process. Instead of just a si
 
 ```python
 agent("""
-Work with your team and build a comprehensive report on fundamental
-concepts about Signal Processing.
+Research fundamental concepts about Signal Processing and build a comprehensive report.
 Write the output in Markdown.
 """)
 ```
 
 # More examples
 
-See the link below to learn more.
+Check out this [Agent Quickstart Example](https://github.com/neuml/txtai/blob/master/examples/agent_quickstart.py). Additional examples are listed below.
 
 | Notebook  | Description  |       |
 |:----------|:-------------|------:|
@@ -149,3 +150,7 @@ See the link below to learn more.
 | [Analyzing Hugging Face Posts with Graphs and Agents](https://github.com/neuml/txtai/blob/master/examples/68_Analyzing_Hugging_Face_Posts_with_Graphs_and_Agents.ipynb) | Explore a rich dataset with Graph Analysis and Agents | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/68_Analyzing_Hugging_Face_Posts_with_Graphs_and_Agents.ipynb) |
 | [Granting autonomy to agents](https://github.com/neuml/txtai/blob/master/examples/69_Granting_autonomy_to_agents.ipynb) | Agents that iteratively solve problems as they see fit | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/69_Granting_autonomy_to_agents.ipynb) |
 | [Analyzing LinkedIn Company Posts with Graphs and Agents](https://github.com/neuml/txtai/blob/master/examples/71_Analyzing_LinkedIn_Company_Posts_with_Graphs_and_Agents.ipynb) | Exploring how to improve social media engagement with AI | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/71_Analyzing_LinkedIn_Company_Posts_with_Graphs_and_Agents.ipynb) |
+| [Parsing the stars with txtai](https://github.com/neuml/txtai/blob/master/examples/72_Parsing_the_stars_with_txtai.ipynb) | Explore an astronomical knowledge graph of known stars, planets, galaxies | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/72_Parsing_the_stars_with_txtai.ipynb) |
+| [Agentic College Search](https://github.com/neuml/txtai/blob/master/examples/82_Agentic_College_Search.ipynb) | Identify list of strong engineering colleges | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/82_Agentic_College_Search.ipynb) |
+| [TxtAI got skills](https://github.com/neuml/txtai/blob/master/examples/83_TxtAI_got_skills.ipynb) | Integrate skill.md files with your agent | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/83_TxtAI_got_skills.ipynb) |
+| [Agent Tools](https://github.com/neuml/txtai/blob/master/examples/84_Agent_Tools.ipynb) [▶️](https://www.youtube.com/watch?v=RDNaFXQy3GQ) | Learn about the txtai agent toolkit | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/84_Agent_Tools.ipynb) |

docs/agent/methods.md

@@ -1,4 +1,4 @@
 # Methods
 
-## ::: txtai.agent.Agent.__init__
-## ::: txtai.agent.Agent.__call__
+## ::: txtai.agent.base.Agent.__init__
+## ::: txtai.agent.base.Agent.__call__

docs/api/configuration.md

@@ -33,7 +33,16 @@ Path to save and load the embeddings index. Each API instance can only access a
 writable: boolean
 ```
 
-Determines if the input embeddings index is writable (true) or read-only (false). This allows serving a read-only index.
+Sets this embeddings index as writable (true) or read-only (false). This allows serving a read-only index. Defaults to read-only.
+
+### reindex
+```yaml
+reindex: boolean
+```
+
+Enables re-indexing for this embeddings index (defaults to disabled).
+
+Re-indexing accepts new configuration and should only be enabled when the configuration comes from trusted sources.
 
 ### cloud
 [Cloud storage settings](../../embeddings/configuration/cloud) can be set under a `cloud` top level configuration group.
@@ -51,7 +60,7 @@ agent:
             - websearch
 
 llm:
-    path: hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4
+    path: Qwen/Qwen3-4B-Instruct-2507
 ```
 
 ## Pipeline

docs/api/index.md

@@ -120,7 +120,7 @@ The following programming languages have bindings with the txtai API:
 - [Rust](https://github.com/neuml/txtai.rs)
 - [Go](https://github.com/neuml/txtai.go)
 
-The API also supports hosting [OpenAI-compatible endpoints](./openai).
+The API also supports hosting [OpenAI-compatible](./openai) and [Model Context Protocol (MCP)](./mcp) endpoints.
 
 See the links below for detailed examples covering the API.
 

docs/api/mcp.md

@@ -0,0 +1,34 @@
+# Model Context Protocol
+
+The [Model Context Protocol (MCP)](https://modelcontextprotocol.io/introduction) is an open standard that enables developers to build secure, two-way connections between their data sources and AI-powered tools.
+
+The API can be configured to handle MCP requests. All enabled endpoints set in the API configuration are automatically added as MCP tools.
+
+```yaml
+mcp: boolean|dict
+```
+
+When `mcp` is a boolean, default arguments are used. When `mcp` is a dictionary it supports the following options.
+
+```yaml
+mcp:
+  clientargs: http client options (dict)
+  mcpargs: FastApiMCP options (dict)
+```
+
+See the following links for details on the options.
+
+- [HTTP Client Arguments](https://www.python-httpx.org/api/#asyncclient)
+- [FastApiMCP Arguments](https://github.com/tadata-org/fastapi_mcp/blob/main/fastapi_mcp/server.py#L22)
+
+Once this configuration option is added, a new route is added to the application `/mcp`. 
+
+The [Model Context Protocol Inspector tool](https://www.npmjs.com/package/@modelcontextprotocol/inspector) is a quick way to explore how the MCP tools are exported through this interface.
+
+Run the following and go to the local URL specified.
+
+```
+npx @modelcontextprotocol/inspector node build/index.js
+```
+
+Enter `http://localhost:8000/mcp` to see the full list of tools available.

docs/api/security.md

@@ -6,9 +6,9 @@ The default implementation of an API service runs via HTTP and is fully open. If
 
 The default API service command starts a Uvicorn server as a HTTP service on port 8000. To run a HTTPS service, consider the following options.
 
-- [TLS Proxy Server](https://fastapi.tiangolo.com/deployment/https/). *Recommended choice*. With this configuration, the txtai API service runs as a HTTP service only accessible on the localhost/local network. The proxy server handles all encryption and redirects requests to local services. See this [example configuration](https://www.uvicorn.org/deployment/#running-behind-nginx) for more.
+- [TLS Proxy Server](https://fastapi.tiangolo.com/deployment/https/). *Recommended choice*. With this configuration, the txtai API service runs as a HTTP service only accessible on the localhost/local network. The proxy server handles all encryption and redirects requests to local services. See this [example configuration](https://www.uvicorn.dev/deployment/#running-behind-nginx) for more.
 
-- [Uvicorn SSL Certificate](https://www.uvicorn.org/deployment/). Another option is setting the SSL certificate on the Uvicorn service. This works in simple situations but gets complex when hosting multiple txtai or other related services.
+- [Uvicorn SSL Certificate](https://www.uvicorn.dev/deployment/). Another option is setting the SSL certificate on the Uvicorn service. This works in simple situations but gets complex when hosting multiple txtai or other related services.
 
 ## Authorization
 
@@ -31,3 +31,44 @@ curl \
 It's important to note that HTTPS **must** be enabled using one of the methods mentioned above. Otherwise, tokens will be exchanged as clear text. 
 
 Authentication and Authorization can be fully customized. See the [dependencies](../customization#dependencies) section for more.
+
+## Safe Open
+
+Many components in txtai work with local files and have no restrictions on what files and urls can be accessed. The [textractor pipeline](../../pipeline/data/textractor) and [retrieve](../../workflow/task/retrieve) tasks both have a built-in option called `safeopen` which is enabled by default when running the API and applications. It enforces the following rules:
+
+- Local files must be in the safeopen directory (defaults to temp dir)
+- URLs must be public URLs (also has checks for redirects and DNS rebind attempts)
+
+The following workflow only allows accessing files stored in a specific directory along with public urls.
+
+```yaml
+workflow:
+  textract:
+    tasks:
+      - action: textractor
+```
+
+A number of of pipelines support reading local files. For these it's recommended to wrap those pipelines as workflows with a `retrieve` task. Note that pipelines can be created in-line and don't need to be exposed via the API.
+
+```yaml
+workflow:
+  caption:
+    tasks:
+      - task: retrieve
+        action: caption
+```
+
+Additionally, internal pipelines can be created but not exposed via the API.
+
+```yaml
+routes:
+  caption: False
+
+caption:
+
+workflow:
+  caption:
+    tasks:
+      - task: retrieve
+        action: caption
+```

docs/cloud.md

@@ -8,7 +8,7 @@ Scalable cloud-native applications can be built with txtai. The following cloud
 - Container Orchestration Systems (i.e. Kubernetes)
 - Docker Engine
 - Serverless Compute
-- txtai.cloud (coming in 2024)
+- txtai.cloud (planned for future)
 
 Images for txtai are available on Docker Hub for [CPU](https://hub.docker.com/r/neuml/txtai-cpu) and [GPU](https://hub.docker.com/r/neuml/txtai-gpu) installs. The CPU install is recommended when GPUs aren't available given the image is significantly smaller.
 
@@ -30,8 +30,8 @@ Examples build commands below.
 # Get Dockerfile
 wget https://raw.githubusercontent.com/neuml/txtai/master/docker/base/Dockerfile
 
-# Build Ubuntu 22.04 image running Python 3.10
-docker build -t txtai --build-arg BASE_IMAGE=ubuntu:22.04 --build-arg PYTHON_VERSION=3.10 .
+# Build Ubuntu 22.04 image running Python 3.11
+docker build -t txtai --build-arg BASE_IMAGE=ubuntu:22.04 --build-arg PYTHON_VERSION=3.11 .
 
 # Build image with GPU support
 docker build -t txtai --build-arg GPU=1 .

docs/embeddings/configuration/ann.md

@@ -4,7 +4,7 @@ Approximate Nearest Neighbor (ANN) index configuration for storing vector embedd
 
 ## backend
 ```yaml
-backend: faiss|hnsw|annoy|numpy|torch|pgvector|sqlite|custom
+backend: faiss|hnsw|annoy|ggml|milvus|numpy|rabitq|torch|turbovec|zvec|pgvector|sqlite|custom
 ```
 
 Sets the ANN backend. Defaults to `faiss`. Additional backends are available via the [ann](../../../install/#ann) extras package. Set custom backends via setting this parameter to the fully resolvable class string.
@@ -62,11 +62,89 @@ annoy:
 
 See [Annoy documentation](https://github.com/spotify/annoy#full-python-api) for more information on these parameters. Note that annoy indexes can not be modified after creation, upserts/deletes and other modifications are not supported.
 
+### ggml
+```yaml
+ggml:
+    gpu: enable GPU - defaults to True
+    quantize: sets the tensor quantization - defaults to F32
+    querysize: query buffer size - defaults to 64
+```
+
+The [GGML](https://github.com/ggml-org/ggml) backend is a k-nearest neighbors backend. It stores tensors using GGML and [GGUF](https://huggingface.co/docs/hub/en/gguf). It supports GPU-enabled operations and supports quantization. GGML is the framework used by [llama.cpp](https://github.com/ggml-org/llama.cpp).
+
+[See this](https://github.com/ggml-org/ggml/blob/master/include/ggml.h#L379) for a list of quantization types.
+
+### milvus
+
+```yaml
+milvus:
+    m: number of HNSW links per element (int) - defaults to 50
+```
+
 ### numpy
 
-The NumPy backend is a k-nearest neighbors backend. It's designed for simplicity and works well with smaller datasets.
+```yaml
+numpy:
+    safetensors: stores vectors using the safetensors format
+                 defaults to NumPy array storage
+```
 
-The `torch` backend supports the same options. The only difference is that the vectors can be search using GPUs.
+The NumPy backend is a k-nearest neighbors backend. It's designed for simplicity and works well with smaller datasets that fit into memory.
+
+### rabitq
+
+```yaml
+rabitq:
+    mode: index mode (ivf or hnsw) - defaults to "ivf"
+    nbits: number of quantization bits per dimension (int) - defaults to 1,
+           supports 1 to 9 in both modes, ivf mode also supports 32 which
+           keeps raw vectors for reranking
+    clusters: number of IVF clusters (int) - defaults to
+              max(1, min(round(4 * sqrt(embeddings count)), embeddings count))
+    nprobe: search probe setting for ivf mode (int) - defaults to
+            max(1, round(num_clusters/16))
+    m: M param for hnsw mode (int) - defaults to 16
+    efconstruction: ef_construction param for hnsw mode (int) - defaults to 200
+    efsearch: ef search param for hnsw mode (int) - defaults to None and not set
+    randomseed: random-seed param for hnsw mode (int) - defaults to 100
+```
+
+The [rabitq](https://github.com/VectorDB-NTU/RaBitQ-Library) backend is a quantized index powered by the [RaBitQ algorithm](https://doi.org/10.1145/3725413). It supports ivf and hnsw search modes. Vectors are stored with 1-bit quantization by default. More bits increase accuracy at the cost of memory. The index is stored as a single file and no vectors are kept outside of it. With nbits set to 32, the vectors are stored in the index for reranking. The ivf mode supports upserts and deletes. Deleted rows keep their storage. Rows added after the first index are assigned to the clusters trained when the index was built, so recall can drop if the new data differs from the original data. Rebuild the index to train new clusters. Note that hnsw indexes can not be modified after creation, upserts/deletes and other modifications are not supported.
+
+### torch
+
+```yaml
+torch:
+    safetensors: stores vectors using the safetensors format - defaults
+                 to NumPy array storage if quantization is disabled
+    quantize:
+        type: quantization type (fp4, nf4, int8)
+        blocksize: quantization block size parameter
+```
+
+The Torch backend is a k-nearest neighbors backend like NumPy. It supports GPU-enabled operations. It also has support for quantization which enables fitting larger arrays into GPU memory.
+
+When quantization is enabled, vectors are _always_ stored in safetensors. _Note that macOS support for quantization is limited._
+
+### turbovec
+
+```yaml
+turbovec:
+    bitwidth: number of bits to store each vector dimension as. Supports 2, 3 or 4.
+```
+
+The [turbovec](https://github.com/RyanCodrai/turbovec) backend is a k-nearest neighbors backend powered by the [TurboQuant algorithm](https://arxiv.org/abs/2504.19874).
+
+### zvec
+
+```yaml
+zvec:
+    efconstruction: ef_construction param for HnswIndexParam (int) - defaults to 200
+    m: number of HNSW links per element (int) - defaults to 50
+    efsearch: ef search param for HnswQueryParam (int) - defaults to 300
+```
+
+The [zvec](https://github.com/alibaba/zvec) backend is an embedded, path-based vector index.
 
 ### pgvector
 ```yaml

docs/embeddings/configuration/database.md

@@ -50,14 +50,27 @@ Object encoding options are:
 functions: list
 ```
 
-List of functions with user-defined SQL functions, only used when [content](#content) is enabled. Each list element must be one of the following:
+List of functions with user-defined SQL functions. Each list element must be one of the following:
 
 - function
 - callable object
-- dict with fields for name, argcount and function
+- dict with fields for name, argcount, function and deterministic
 
 [An example can be found here](../../query#custom-sql-functions).
 
+## expressions
+```yaml
+expressions: list
+```
+
+List of expression shortcuts. Each list element must be a dict with the following fields.
+
+- `name`: name of the expression
+- `expression`: SQL expression, defaults to `name` when empty
+- `index`: if this expression should have a database index, defaults to False when not provided
+
+The expression can be a json data column, sql function or anything that can be run as a SQL snippet.
+
 ## query
 ```yaml
 query:

docs/embeddings/configuration/general.md

@@ -1,15 +1,35 @@
 # General
 
-General configuration options that don't fit elsewhere.
+General configuration options.
 
 ## keyword
 ```yaml
-keyword: boolean
+keyword: boolean|string
 ```
 
 Enables sparse keyword indexing for this embeddings.
 
-When enabled, this parameter creates a BM25 index for full text search. It also implicitly disables the [defaults](../vectors/#defaults) setting for vector search.
+When set to a boolean, this parameter creates a BM25 index for full text search. When set to a string, it expects a [keyword method](../scoring#method).
+
+It also implicitly disables the [defaults](#defaults) setting for vector search.
+
+## sparse
+```yaml
+sparse: boolean|path
+```
+
+Enables sparse vector indexing for this embeddings.
+
+When set to `True`, this parameter creates a sparse vector index using the [default sparse index model](https://huggingface.co/prithivida/Splade_PP_en_v2). When set to a string, it expects a local or Hugging Face model path.
+
+It also implicitly disables the [defaults](#defaults) setting for vector search.
+
+## dense
+```yaml
+dense: boolean|string
+```
+
+Alias for the [vector model path](../vectors/#path). When set to `True`, the [default transformers vector model](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) is used.
 
 ## hybrid
 ```yaml
@@ -18,7 +38,14 @@ hybrid: boolean
 
 Enables hybrid (sparse + dense) indexing for this embeddings.
 
-When enabled, this parameter creates a BM25 index for full text search. It has no effect on the [defaults](../vectors/#defaults) or [path](../vectors/#path) settings.
+When enabled, this parameter creates a BM25 index for full text search. It has no effect on the [defaults](#defaults) or [path](../vectors/#path) settings.
+
+## defaults
+```yaml
+defaults: boolean
+```
+
+Uses default vector model path when enabled (default setting is True) and `path` is not provided. See [this link](../) for an example.
 
 ## indexes
 ```yaml
@@ -39,13 +66,18 @@ Sets the auto id generation method. When this is not set, an autogenerated numer
 columns:
     text: name of the text column
     object: name of the object column
+    store: limit json data fields to this list of columns
 ```
 
 Sets the `text` and `object` column names. Defaults to `text` and `object` if not provided.
 
+`store` sets a list of columns to store in the JSON data field. When this isn't provided, all columns are stored (default). When `store` is set to `None`, no JSON columns are stored. This is useful is a field is only needed at indexing time but not search time.
+
 ## format
 ```yaml
 format: json|pickle
 ```
 
 Sets the configuration storage format. Defaults to `json`.
+
+Note that `pickle` configuration is deprecated and will raise an error with the default settings. Its only kept for reading legacy indexes and requires setting the `ALLOW_PICKLE` environment variable. Only enable for local and/or trusted sources.

docs/embeddings/configuration/scoring.md

@@ -12,7 +12,7 @@ The following covers the available options.
 
 ## method
 ```yaml
-method: bm25|tfidf|sif|pgtext|custom
+method: bm25|tfidf|sif|pgtext|sparse|custom
 ```
 
 Sets the scoring method. Add custom scoring via setting this parameter to the fully resolvable class string.
@@ -25,6 +25,57 @@ schema: database schema to store keyword index - defaults to being
 
 Additional settings for Postgres full-text keyword indexes.
 
+### sparse
+```yaml
+path: sparse vector model path
+vectormethod: vector embeddings method
+vectornormalize: enable vector embeddings normalization (boolean)
+gpu: boolean|int|string|device
+normalize: enable score normalization (boolean|float|string|dict)
+batch: Sets the transform batch size
+encodebatch: Sets the encode batch size
+vectors: additional model init args
+encodeargs: additional encode() args
+backend: ivfsparse|pgsparse|zvecsparse
+```
+
+Sparse vector scoring options. The sparse scoring instance combines a sparse vector model with a sparse approximate nearest neighbor index (ANN). This method supports both vector normalization and score normalization.
+
+Vector normalization normalizes all vectors to have a magnitude of 1. By extension, all generated scores will be 0 to 1.
+
+Score normalization scales the output between 0 and 1. This setting supports:
+
+- `True` for default scale normalization
+- `float` normalize using this as the scale factor
+- `"bayes"` for Bayesian normalization using dynamic candidate score statistics
+- `{method: "bayes", alpha: 1.0, beta: null}` for Bayesian normalization with optional custom parameters
+
+#### ivfsparse
+```yaml
+ivfsparse:
+  sample: percent of data to use for model training (0.0 - 1.0)
+  nfeatures: top n features to use for model training, defaults to all features (int)
+  nlist: desired number of clusters (int)
+  nprobe: search probe setting (int)
+  minpoints: minimum number of points for a cluster (int)
+```
+
+Inverted file (IVF) index with flat vector file storage and sparse array support.
+
+#### pgsparse
+
+Sparse ANN backed by Postgres. Supports same options as the [pgvector](../ann/#pgvector) ANN.
+
+#### zvecsparse
+```yaml
+zvecsparse:
+  efconstruction: ef_construction param for HnswIndexParam (int) - defaults to 200
+  m: number of HNSW links per element (int) - defaults to 50
+  efsearch: ef search param for HnswQueryParam (int) - defaults to 300
+```
+
+Sparse ANN backed by [zvec](https://github.com/alibaba/zvec). Stores sparse vectors in an embedded, path-based vector index. Requires the [ann](../../../install/#ann) extras package.
+
 ## terms
 ```yaml
 terms: boolean|dict
@@ -40,9 +91,38 @@ Supports a `dict` with the parameters `cachelimit` and `cutoff`.
 
 When `terms` is set to `True`, default parameters are used for the `cachelimit` and `cutoff`. Normally, these defaults are sufficient.
 
+### Wildcard queries
+
+Keyword queries containing `*` expand matching terms using SQL `LIKE`. Within these
+expressions, `%` matches any sequence and `_` matches one character. To match a literal
+percent sign, underscore or backslash, use `\%`, `\_` or `\\`, respectively.
+For example, `report\_*` matches `report_2026`, while `report_*` also matches `reportX2026`.
+Use a tokenizer that preserves these characters, such as `tokenizer: {whitespace: true}`,
+or pass pre-tokenized query terms. Queries without `*` retain exact-term matching.
+
 ## normalize
 ```yaml
-normalize: boolean
+normalize: boolean|str|dict
 ```
 
-Enables normalized scoring (ranging from 0 to 1). When enabled, statistics from the index will be used to calculate normalized scores.
+Enables normalized scoring (ranging from 0 to 1). This setting supports:
+
+- `True` for standard score normalization
+- `"bayes"` | `"bb25"` for Bayesian normalization using dynamic candidate score statistics
+- `{method: "bayes", alpha: 1.0, beta: null}` for Bayesian normalization with optional custom parameters
+
+When standard normalization is enabled, statistics from the index are used to calculate normalized scores.
+When Bayesian/BB25 normalization is enabled, it uses positive-score candidates, dynamic `beta=median(scores)`, adaptive
+`alpha_eff=alpha/std(scores)` and a sigmoid transform (likelihood-only variant with flat prior) to map scores to `[0, 1]`.
+
+Bayesian normalization references:
+
+- [https://github.com/instructkr/bb25](https://github.com/instructkr/bb25)
+- [https://github.com/cognica-io/bayesian-bm25](https://github.com/cognica-io/bayesian-bm25)
+
+## tokenizer
+```yaml
+tokenizer: dict
+```
+
+Set tokenization rules. Passes these arguments to the underlying [Tokenization pipeline](../../../pipeline/data/tokenizer#txtai.pipeline.Tokenizer.__init__).

docs/embeddings/configuration/vectors.md

@@ -10,16 +10,10 @@ path: string
 Sets the path for a vectors model. When using a transformers/sentence-transformers model, this can be any model on the
 [Hugging Face Hub](https://huggingface.co/models) or a local file path. Otherwise, it must be a local file path to a word embeddings model.
 
-## defaults
-```yaml
-defaults: boolean
-```
-
-Uses default vector model path when enabled (default setting is True) and `path` is not provided. See [this link](../) for an example.
-
 ## method
 ```yaml
-method: transformers|sentence-transformers|llama.cpp|litellm|model2vec|external|words
+method: transformers|sentence-transformers|llama.cpp|litellm|model2vec|external|
+        words
 ```
 
 Embeddings method to use. If the method is not provided, it is inferred using the `path`.
@@ -31,11 +25,12 @@ Embeddings method to use. If the method is not provided, it is inferred using th
 Builds embeddings using a transformers model. While this can be any transformers model, it works best with
 [models trained](https://huggingface.co/models?pipeline_tag=sentence-similarity) to build embeddings.
 
-Both `mean` and `cls` pooling are supported and automatically inferred from the model. The pooling method can be overwritten by changing the method
-from `transformers` to `meanpooling` or `clspooling` respectively.
+`mean`, `cls`, `last`, `late` and `max` pooling are supported and automatically inferred from the model. The pooling method can be overwritten by changing the method from `transformers` to `meanpooling`, `clspooling`, `lastpooling`, `latepooling` or `maxpooling` respectively.
 
 Setting `maxlength` to `True` enables truncating inputs to the `max_seq_length`. Setting `maxlength` to an integer will truncate inputs to that value. When omitted (default), the `maxlength` will be set to either the model or tokenizer maxlength.
 
+Supports loading ONNX models directly. [See this section for more](../../../examples/#model-training).
+
 ### sentence-transformers
 
 Same as transformers but loads models with the [sentence-transformers](https://github.com/UKPLab/sentence-transformers) library.
@@ -72,7 +67,9 @@ Embeddings are created via an external model or API. Requires setting the [trans
 transform: function
 ```
 
-When method is `external`, this function transforms input content into embeddings. The input to this function is a list of data. This method must return either a numpy array or list of numpy arrays.
+When method is `external`, this function transforms input content into embeddings. The input to this function is a list of data. This method must return either a numpy array, list of numpy arrays or list of float arrays.
+
+Note that this method is deprecated and will raise an error with the default settings. Adding a custom `Vectors` implementation is a cleaner approach. The `ALLOW_RESOLVE_TRANSFORM` environment variable must be set to enable. Only enable for trusted and/or reviewed code.
 
 ## gpu
 ```yaml
@@ -111,7 +108,7 @@ quantize: int|boolean
 
 Enables scalar vector quantization at the specified precision. Supports 1-bit through 8-bit quantization. Scalar quantization transforms continuous floating point values to discrete unsigned integers. The `faiss`, `pgvector`, `numpy` and `torch` ANN backends support storing these vectors.
 
-This parameter supports booleans for backwards compatability. When set to true/false, this flag sets [faiss.quantize](../ann/#faiss).
+This parameter supports booleans for backwards compatibility. When set to true/false, this flag sets [faiss.quantize](../ann/#faiss).
 
 In addition to vector-level quantization, some ANN backends have the ability to quantize vectors at the storage layer. See the [ANN](../ann) configuration options for more.
 
@@ -124,7 +121,7 @@ instructions:
 
 Instruction-based models use prefixes to modify how embeddings are computed. This is especially useful with asymmetric search, which is when the query and indexed data are of vastly different lengths. In other words, short queries with long documents.
 
-[E5-base](https://huggingface.co/intfloat/e5-base) is an example of a model that accepts instructions. It takes `query: ` and `passage: ` prefixes and uses those to generate embeddings that work well for asymmetric search.
+`txtai` automatically loads prompts stored in `config_sentence_transformers.json` except if this parameter is set. For some older models such as [E5-base](https://huggingface.co/intfloat/e5-base), instructions still need to be provided via this parameter.
 
 ## models
 ```yaml
@@ -146,3 +143,58 @@ vectors: dict
 ```
 
 Passes these additional parameters to the underlying vector model.
+
+### center
+```yaml
+center: boolean | dict
+```
+
+Centers normalized late-interaction token vectors and normalizes them again before MUVERA or LEMUR encoding. Centering is applied to
+both query and document vectors. `true` uses batch scope and `false` disables centering. An explicit setting always takes precedence.
+When omitted, a LEMUR artifact with a stored collection mean uses collection scope. Other LEMUR artifacts and late-interaction models
+default to batch scope only when they load more than one `torch.nn.Linear` layer; models with zero or one loaded linear layer keep their
+previous output.
+
+Dictionary configuration supports the following scopes:
+
+```yaml
+center:
+    scope: document | batch | collection
+```
+
+`document` subtracts the mean of each document or query independently. `batch` subtracts the mean of all real token rows in the current
+model batch. Explicit `collection` scope requires a caller-provided, one-dimensional mean vector through either `mean` (an array-like
+value) or `path` (a safetensors file containing a `center.mean` tensor), but not both. LEMUR artifacts can provide this mean
+automatically. Zero-padding rows are never centered and remain zero.
+
+### muvera
+```yaml
+muvera:
+    repetitions: defaults 20
+    hashes: defaults to 5
+    projection: defaults 16
+```
+
+Settings to control the size of MUVERA fixed dimensional outputs. Default is 20 * 2^5 * 16 = 10,240 dimensions.
+Set `muvera` to `false` to disable it or `true` to use the default settings. `false` only makes sense when the multi-vector output
+is consumed directly, for example through `PoolingFactory`. Embeddings vectors require `muvera` or `lemur`.
+
+### lemur
+```yaml
+lemur:
+    path: local artifact directory or Hugging Face Hub path
+```
+
+Loads a trained [LEMUR](../../../pipeline/train/lemur) fixed dimensional encoder. LEMUR artifacts are corpus-specific and must be
+created with `LemurTrainer` before configuring an embeddings index. A bare string is treated as the artifact path. Query vectors are
+summed learned features and document vectors are ordinary least squares weights over the artifact's stored token sample. Each artifact
+contains `config.json` and `model.safetensors`; newly trained artifacts also store the collection token mean as `lemur.center`. Loading that mean automatically
+selects collection centering unless `center` is explicitly configured. Artifacts without `lemur.center` retain the previous default.
+The Safetensors file contains inference state only. See the trainer page for score filtering and Faiss exact/IVF search behavior.
+
+### trust_remote_code
+```yaml
+trust_remote_code: boolean
+```
+
+Parameter for trusting the code from Hugging Face models with custom implementations.

docs/embeddings/indexing.md

@@ -23,6 +23,8 @@ As mentioned above, computed vectors are stored in an ANN. There are various ind
 
 Embeddings indexes can optionally [store content](../configuration/database#content). When this is enabled, the input content is saved in a database alongside the computed vectors. This enables filtering on additional fields and content retrieval.
 
+The columns used for text, object and JSON data storage are set via [column configuration](../configuration/general#columns).
+
 ## Index vs Upsert
 
 Data is loaded into an index with either an [index](../methods#txtai.embeddings.base.Embeddings.index) or [upsert](../methods#txtai.embeddings.base.Embeddings.upsert) call.
@@ -112,7 +114,7 @@ Graphs are persisted alongside an embeddings index. Each save and load will also
 
 ## Sparse vectors
 
-Scoring instances can create a standalone [keyword](../configuration/general#keyword) or sparse index (BM25, TF-IDF). This enables [hybrid](../configuration/general/#hybrid) search when there is an available dense vector index.
+Scoring instances can create a standalone [sparse keyword indexes](../configuration/general#keyword) (BM25, TF-IDF) and [sparse vector indexes](../configuration/general#sparse) (SPLADE). This enables [hybrid](../configuration/general/#hybrid) search when there is an available dense vector index.
 
 The terms `sparse vector index`, `keyword index`, `terms index` and `scoring index` are used interchangeably throughout txtai's documentation.
 

docs/embeddings/query.md

@@ -54,6 +54,8 @@ The number of candidates should be larger than the desired number of results whe
 - For a single query filter clause, the default is the query limit
 - With multiple filtering clauses, the default is 10x the query limit
 
+When a query orders results by score, its `offset` adds to the query limit.
+
 The index name is only applicable when [subindexes](../configuration/general/#indexes) are enabled. This specifies the index to use for the query.
 
 Weights sets the hybrid score weights when an index has both a sparse and dense index.
@@ -132,8 +134,11 @@ embeddings = Embeddings(content=True, objects=True)
 request = open("demo.gif", "rb")
 
 # Insert record
-embeddings.index([("txtai", {"text": "txtai executes machine-learning workflows.",
-                             "object": request.read()})])
+embeddings.index([(
+  "txtai",
+  {"text": "txtai executes machine-learning workflows.",
+   "object": request.read()}
+)])
 
 # Query txtai and get associated object
 query = "SELECT object FROM txtai WHERE similar('machine learning') LIMIT 1"
@@ -169,6 +174,26 @@ LIMIT 1
 """)
 ```
 
+## Expressions
+
+Expression shortcuts expand into more complex SQL snippets. This is useful for making SQL queries more concise. Indexing is also available on expressions as a performance improvement.
+
+The following example indexes a json extraction field (`filepath`) and the length of each field.
+
+```python
+# Create embeddings index
+embeddings = Embeddings(
+  path="sentence-transformers/nli-mpnet-base-v2",
+  content=True,
+  expressions=[
+    {"name": "filepath", "index": True},
+    {"name": "textlength", "expression": "length(text)", "index": True}
+  ]
+)
+
+embeddings.search("SELECT textlength, filepath FROM txtai LIMIT 1")
+```
+
 ## Query translation
 
 Natural language queries with filters can be converted to txtai-compatible SQL statements with query translation. For example:
@@ -192,7 +217,9 @@ When an embeddings database has both a sparse and dense index, both indexes will
 
 ```python
 embeddings.search("query", weights=0.5)
-embeddings.search("SELECT id, text, score FROM txtai WHERE similar('query', 0.5)")
+embeddings.search(
+  "SELECT id, text, score FROM txtai WHERE similar('query', 0.5)"
+)
 ```
 
 ## Graph search
@@ -203,11 +230,20 @@ Additionally, standard embeddings searches can be returned as graphs.
 
 ```python
 # Find all paths between id: 0 and id: 5 between 1 and 3 hops away
-embeddings.graph.search("""
+embeddings.search("""
 MATCH P=({id: 0})-[*1..3]->({id: 5})
 RETURN P
 """)
 
+# Find related nodes for query matches
+embeddings.search("""
+  MATCH P=(A)-[]->(B)
+  WHERE SIMILAR(A, "query")
+  RETURN B
+  ORDER BY A.score DESC
+  LIMIT 10
+""")
+
 # Standard embeddings search as graph
 embeddings.search("query", graph=True)
 ```
@@ -217,13 +253,28 @@ embeddings.search("query", graph=True)
 Subindexes can be queried as follows:
 
 ```python
+# Build index with subindexes
+embeddings = Embeddings(
+  content=True,
+  defaults=False,
+  indexes={
+    "keyword": {
+      "keyword": True
+    },
+    "dense":{
+      "dense": True
+    }
+  }
+)
+embeddings.index(stream())
+
 # Query with index parameter
-embeddings.search("query", index="subindex1")
+embeddings.search("query", index="keyword")
 
 # Specify with SQL
 embeddings.search("""
 SELECT id, text, score FROM txtai
-WHERE similar('query', 'subindex1')
+WHERE similar('query', 'keyword')
 """)
 ```
 

docs/examples.md

@@ -31,7 +31,7 @@ Autonomous agents, retrieval augmented generation (RAG), chat with your data, pi
 |:----------|:-------------|------:|
 | [Prompt-driven search with LLMs](https://github.com/neuml/txtai/blob/master/examples/42_Prompt_driven_search_with_LLMs.ipynb) | Embeddings-guided and Prompt-driven search with Large Language Models (LLMs) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/42_Prompt_driven_search_with_LLMs.ipynb) |
 | [Prompt templates and task chains](https://github.com/neuml/txtai/blob/master/examples/44_Prompt_templates_and_task_chains.ipynb) | Build model prompts and connect tasks together with workflows | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/44_Prompt_templates_and_task_chains.ipynb) |
-| [Build RAG pipelines with txtai](https://github.com/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) | Guide on retrieval augmented generation including how to create citations | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) |
+| [Build RAG pipelines with txtai](https://github.com/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) [▶️](https://www.youtube.com/watch?v=t_OeAc8NVfQ) | Guide on retrieval augmented generation including how to create citations | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) |
 | [Integrate LLM frameworks](https://github.com/neuml/txtai/blob/master/examples/53_Integrate_LLM_Frameworks.ipynb) | Integrate llama.cpp, LiteLLM and custom generation frameworks | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/53_Integrate_LLM_Frameworks.ipynb) |
 | [Generate knowledge with Semantic Graphs and RAG](https://github.com/neuml/txtai/blob/master/examples/55_Generate_knowledge_with_Semantic_Graphs_and_RAG.ipynb) | Knowledge exploration and discovery with Semantic Graphs and RAG | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/55_Generate_knowledge_with_Semantic_Graphs_and_RAG.ipynb) |
 | [Build knowledge graphs with LLMs](https://github.com/neuml/txtai/blob/master/examples/57_Build_knowledge_graphs_with_LLM_driven_entity_extraction.ipynb) | Build knowledge graphs with LLM-driven entity extraction | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/57_Build_knowledge_graphs_with_LLM_driven_entity_extraction.ipynb) |
@@ -46,6 +46,13 @@ Autonomous agents, retrieval augmented generation (RAG), chat with your data, pi
 | [Analyzing LinkedIn Company Posts with Graphs and Agents](https://github.com/neuml/txtai/blob/master/examples/71_Analyzing_LinkedIn_Company_Posts_with_Graphs_and_Agents.ipynb) | Exploring how to improve social media engagement with AI | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/71_Analyzing_LinkedIn_Company_Posts_with_Graphs_and_Agents.ipynb) |
 | [Parsing the stars with txtai](https://github.com/neuml/txtai/blob/master/examples/72_Parsing_the_stars_with_txtai.ipynb) | Explore an astronomical knowledge graph of known stars, planets, galaxies | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/72_Parsing_the_stars_with_txtai.ipynb) |
 | [Chunking your data for RAG](https://github.com/neuml/txtai/blob/master/examples/73_Chunking_your_data_for_RAG.ipynb) | Extract, chunk and index content for effective retrieval | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/73_Chunking_your_data_for_RAG.ipynb) |
+| [Medical RAG Research with txtai](https://github.com/neuml/txtai/blob/master/examples/75_Medical_RAG_Research_with_txtai.ipynb) | Analyze PubMed article metadata with RAG | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/75_Medical_RAG_Research_with_txtai.ipynb) |
+| [GraphRAG with Wikipedia and GPT OSS](https://github.com/neuml/txtai/blob/master/examples/77_GraphRAG_with_Wikipedia_and_GPT_OSS.ipynb) | Deep graph search powered RAG | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/77_GraphRAG_with_Wikipedia_and_GPT_OSS.ipynb) |
+| [RAG is more than Vector Search](https://github.com/neuml/txtai/blob/master/examples/79_RAG_is_more_than_Vector_Search.ipynb) | Context retrieval via Web, SQL and other sources | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/79_RAG_is_more_than_Vector_Search.ipynb) |
+| [OpenCode as a txtai LLM](https://github.com/neuml/txtai/blob/master/examples/81_OpenCode_as_a_txtai_LLM.ipynb) | Integrate OpenCode with the txtai ecosystem | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/81_OpenCode_as_a_txtai_LLM.ipynb) |
+| [Agentic College Search](https://github.com/neuml/txtai/blob/master/examples/82_Agentic_College_Search.ipynb) | Identify a list of strong engineering colleges | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/82_Agentic_College_Search.ipynb) |
+| [TxtAI got skills](https://github.com/neuml/txtai/blob/master/examples/83_TxtAI_got_skills.ipynb) | Integrate skill.md files with your agent | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/83_TxtAI_got_skills.ipynb) |
+| [Agent Tools](https://github.com/neuml/txtai/blob/master/examples/84_Agent_Tools.ipynb) [▶️](https://www.youtube.com/watch?v=RDNaFXQy3GQ) | Learn about the txtai agent toolkit | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/84_Agent_Tools.ipynb) |
 
 ## Pipelines
 
@@ -83,7 +90,7 @@ Efficiently process data at scale.
 
 ## Model Training
 
-Train NLP models.
+Train, distill, fine-tune and export models.
 
 | Notebook  | Description  |       |
 |:----------|:-------------|------:|
@@ -91,6 +98,8 @@ Train NLP models.
 | [Train without labels](https://github.com/neuml/txtai/blob/master/examples/17_Train_without_labels.ipynb) | Use zero-shot classifiers to train new models | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/17_Train_without_labels.ipynb) |
 | [Train a QA model](https://github.com/neuml/txtai/blob/master/examples/19_Train_a_QA_model.ipynb) | Build and fine-tune question-answering models | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/19_Train_a_QA_model.ipynb) |
 | [Train a language model from scratch](https://github.com/neuml/txtai/blob/master/examples/41_Train_a_language_model_from_scratch.ipynb) | Build new language models | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/41_Train_a_language_model_from_scratch.ipynb) |
+| [Distilling Knowledge into Tiny LLMs](https://github.com/neuml/txtai/blob/master/examples/80_Distilling_Knowledge_into_Tiny_LLMs.ipynb) [▶️](https://www.youtube.com/watch?v=Ol560ktgkf0) | Finetune tiny LLMs to enable inference using less resources | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/80_Distilling_Knowledge_into_Tiny_LLMs.ipynb) |
+| [Progressive Distillation](https://github.com/neuml/txtai/blob/master/examples/85_Progressive_Distillation.ipynb) | Incrementally compress knowledge from a series of teachers | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/85_Progressive_Distillation.ipynb) |
 | [Export and run models with ONNX](https://github.com/neuml/txtai/blob/master/examples/18_Export_and_run_models_with_ONNX.ipynb) | Export models with ONNX, run natively in JavaScript, Java and Rust | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/18_Export_and_run_models_with_ONNX.ipynb) |
 | [Export and run other machine learning models](https://github.com/neuml/txtai/blob/master/examples/21_Export_and_run_other_machine_learning_models.ipynb) | Export and run models from scikit-learn, PyTorch and more | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/21_Export_and_run_other_machine_learning_models.ipynb) |
 
@@ -116,13 +125,14 @@ Project architecture, data formats, external integrations, scale to production,
 | [Anatomy of a txtai index](https://github.com/neuml/txtai/blob/master/examples/29_Anatomy_of_a_txtai_index.ipynb) | Deep dive into the file formats behind a txtai embeddings index | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/29_Anatomy_of_a_txtai_index.ipynb) |
 | [Embeddings components](https://github.com/neuml/txtai/blob/master/examples/37_Embeddings_index_components.ipynb) | Composable search with vector, SQL and scoring components | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/37_Embeddings_index_components.ipynb) |
 | [Customize your own embeddings database](https://github.com/neuml/txtai/blob/master/examples/45_Customize_your_own_embeddings_database.ipynb) | Ways to combine vector indexes with relational databases | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/45_Customize_your_own_embeddings_database.ipynb) |
-| [Building an efficient sparse keyword index in Python](https://github.com/neuml/txtai/blob/master/examples/47_Building_an_efficient_sparse_keyword_index_in_Python.ipynb) | Fast and accurate sparse keyword indexing | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/48_Benefits_of_hybrid_search.ipynb) |
+| [Building an efficient sparse keyword index in Python](https://github.com/neuml/txtai/blob/master/examples/47_Building_an_efficient_sparse_keyword_index_in_Python.ipynb) | Fast and accurate sparse keyword indexing | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/47_Building_an_efficient_sparse_keyword_index_in_Python.ipynb) |
 | [Benefits of hybrid search](https://github.com/neuml/txtai/blob/master/examples/48_Benefits_of_hybrid_search.ipynb) | Improve accuracy with a combination of semantic and keyword search | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/48_Benefits_of_hybrid_search.ipynb) |
 | [External database integration](https://github.com/neuml/txtai/blob/master/examples/49_External_database_integration.ipynb) | Store metadata in PostgreSQL, MariaDB, MySQL and more | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/49_External_database_integration.ipynb) |
 | [All about vector quantization](https://github.com/neuml/txtai/blob/master/examples/50_All_about_vector_quantization.ipynb) | Benchmarking scalar and product quantization methods | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/50_All_about_vector_quantization.ipynb) |
 | [External vectorization](https://github.com/neuml/txtai/blob/master/examples/56_External_vectorization.ipynb) | Vectorization with precomputed embeddings datasets and APIs | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/56_External_vectorization.ipynb) |
 | [Integrate txtai with Postgres](https://github.com/neuml/txtai/blob/master/examples/61_Integrate_txtai_with_Postgres.ipynb) | Persist content, vectors and graph data in Postgres | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/61_Integrate_txtai_with_Postgres.ipynb) |
 | [Embeddings index format for open data access](https://github.com/neuml/txtai/blob/master/examples/64_Embeddings_index_format_for_open_data_access.ipynb) | Platform and programming language independent data storage with txtai | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/64_Embeddings_index_format_for_open_data_access.ipynb) |
+| [Accessing Low Level Vector APIs](https://github.com/neuml/txtai/blob/master/examples/78_Accessing_Low_Level_Vector_APIs.ipynb) | Build a vector database using txtai's low-level APIs | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/78_Accessing_Low_Level_Vector_APIs.ipynb) |
 
 ## Releases
 
@@ -134,6 +144,7 @@ New functionality added in major releases.
 | [What's new in txtai 6.0](https://github.com/neuml/txtai/blob/master/examples/46_Whats_new_in_txtai_6_0.ipynb) | Sparse, hybrid and subindexes for embeddings, LLM improvements | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/46_Whats_new_in_txtai_6_0.ipynb) |
 | [What's new in txtai 7.0](https://github.com/neuml/txtai/blob/master/examples/59_Whats_new_in_txtai_7_0.ipynb) | Semantic graph 2.0, LoRA/QLoRA training and binary API support | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/59_Whats_new_in_txtai_7_0.ipynb) |
 | [What's new in txtai 8.0](https://github.com/neuml/txtai/blob/master/examples/67_Whats_new_in_txtai_8_0.ipynb) | Agents with txtai | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/67_Whats_new_in_txtai_8_0.ipynb) |
+| [What's new in txtai 9.0](https://github.com/neuml/txtai/blob/master/examples/76_Whats_new_in_txtai_9_0.ipynb) | Learned sparse vectors, late interaction models and rerankers | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/76_Whats_new_in_txtai_9_0.ipynb) |
 
 ## Applications
 

docs/faq.md

@@ -68,11 +68,11 @@ import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer
 from txtai import LLM
 
-# Load Phi 3.5-mini
-path = "microsoft/Phi-3.5-mini-instruct"
+# Load Qwen3 0.6B
+path = "Qwen/Qwen3-0.6B"
 model = AutoModelForCausalLM.from_pretrained(
   path,
-  torch_dtype=torch.bfloat16,
+  dtype=torch.bfloat16,
 )
 tokenizer = AutoTokenizer.from_pretrained(path)
 
@@ -105,11 +105,12 @@ __Solution__
 
 Set the following environment parameters.
 
-- OpenMP threading is handled internally on macOS platforms but it can be disabled via `export OMP_NUM_THREADS=1`
+- Disable OpenMP multithreading via `export OMP_NUM_THREADS=1`
+- Workaround `OMP: Error #15` errors via `export KMP_DUPLICATE_LIB_OK=TRUE`
 - Disable PyTorch MPS device via `export PYTORCH_MPS_DISABLE=1`
 - Disable llama.cpp metal via `export LLAMA_NO_METAL=1`
 
-For more details, refer to [this issue on GitHub](https://github.com/kyamagu/faiss-wheels/issues/100).
+For more details, refer to [this issue on GitHub](https://github.com/kyamagu/faiss-wheels/issues/73).
 
 ----------
 

docs/further.md

@@ -3,9 +3,10 @@
 ![further](images/further.png#only-light)
 ![further](images/further-dark.png#only-dark)
 
-- [Introducing txtai, the all-in-one embeddings database](https://medium.com/neuml/introducing-txtai-the-all-in-one-embeddings-database-c721f4ff91ad)
+- [Introducing txtai, the all-in-one AI framework](https://medium.com/neuml/introducing-txtai-the-all-in-one-ai-framework-0660ecfc39d7)
+- [txtai: An All-in-One AI Framework for Semantic Search and LLM Workflows](https://github.com/neuml/papers/blob/master/txtai/txtai.pdf)
 - [Tutorial series on Hashnode](https://neuml.hashnode.dev/series/txtai-tutorial) | [dev.to](https://dev.to/neuml/tutorial-series-on-txtai-ibg)
-- [What's new in txtai 8.0](https://medium.com/neuml/whats-new-in-txtai-8-0-2d7d0ab4506b) | [7.0](https://medium.com/neuml/whats-new-in-txtai-7-0-855ad6a55440) | [6.0](https://medium.com/neuml/whats-new-in-txtai-6-0-7d93eeedf804) | [5.0](https://medium.com/neuml/whats-new-in-txtai-5-0-e5c75a13b101) | [4.0](https://medium.com/neuml/whats-new-in-txtai-4-0-bbc3a65c3d1c)
+- [What's new in txtai 9.0](https://medium.com/neuml/whats-new-in-txtai-9-0-d522bb150afa) | [8.0](https://medium.com/neuml/whats-new-in-txtai-8-0-2d7d0ab4506b) | [7.0](https://medium.com/neuml/whats-new-in-txtai-7-0-855ad6a55440) | [6.0](https://medium.com/neuml/whats-new-in-txtai-6-0-7d93eeedf804) | [5.0](https://medium.com/neuml/whats-new-in-txtai-5-0-e5c75a13b101) | [4.0](https://medium.com/neuml/whats-new-in-txtai-4-0-bbc3a65c3d1c)
 - [Getting started with semantic search](https://medium.com/neuml/getting-started-with-semantic-search-a9fd9d8a48cf) | [workflows](https://medium.com/neuml/getting-started-with-semantic-workflows-2fefda6165d9) | [rag](https://medium.com/neuml/getting-started-with-rag-9a0cca75f748)
 - [Running txtai at scale](https://medium.com/neuml/running-at-scale-with-txtai-71196cdd99f9)
 - [Vector search & RAG Landscape: A review with txtai](https://medium.com/neuml/vector-search-rag-landscape-a-review-with-txtai-a7f37ad0e187)

docs/index.md

@@ -5,7 +5,7 @@
 </p>
 
 <p align="center">
-    <b>All-in-one embeddings database</b>
+    <b>All-in-one AI framework</b>
 </p>
 
 <p align="center">
@@ -18,7 +18,7 @@
     <a href="https://github.com/neuml/txtai/issues">
         <img src="https://img.shields.io/github/issues/neuml/txtai.svg?style=flat&color=success" alt="GitHub issues"/>
     </a>
-    <a href="https://join.slack.com/t/txtai/shared_invite/zt-1cagya4yf-DQeuZbd~aMwH5pckBU4vPg">
+    <a href="https://join.slack.com/t/txtai/shared_invite/zt-37c1zfijp-Y57wMty6YOx_hyIHEQvQJA">
         <img src="https://img.shields.io/badge/slack-join-blue?style=flat&logo=slack&logocolor=white" alt="Join Slack"/>
     </a>
     <a href="https://github.com/neuml/txtai/actions?query=workflow%3Abuild">
@@ -29,12 +29,12 @@
     </a>
 </p>
 
-txtai is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.
+txtai is an all-in-one AI framework for semantic search, LLM orchestration and language model workflows.
 
 ![architecture](images/architecture.png#gh-light-mode-only)
 ![architecture](images/architecture-dark.png#gh-dark-mode-only)
 
-Embeddings databases are a union of vector indexes (sparse and dense), graph networks and relational databases.
+The key component of txtai is an embeddings database, which is a union of vector indexes (sparse and dense), graph networks and relational databases.
 
 This foundation enables vector search and/or serves as a powerful knowledge source for large language model (LLM) applications.
 
@@ -47,10 +47,14 @@ Summary of txtai features:
 - 💡 Pipelines powered by language models that run LLM prompts, question-answering, labeling, transcription, translation, summarization and more
 - ↪️️ Workflows to join pipelines together and aggregate business logic. txtai processes can be simple microservices or multi-model workflows.
 - 🤖 Agents that intelligently connect embeddings, pipelines, workflows and other agents together to autonomously solve complex problems
-- ⚙️ Build with Python or YAML. API bindings available for [JavaScript](https://github.com/neuml/txtai.js), [Java](https://github.com/neuml/txtai.java), [Rust](https://github.com/neuml/txtai.rs) and [Go](https://github.com/neuml/txtai.go).
+- ⚙️ Web and Model Context Protocol (MCP) APIs. Bindings available for [JavaScript](https://github.com/neuml/txtai.js), [Java](https://github.com/neuml/txtai.java), [Rust](https://github.com/neuml/txtai.rs) and [Go](https://github.com/neuml/txtai.go).
 - 🔋 Batteries included with defaults to get up and running fast
 - ☁️ Run local or scale out with container orchestration
 
-txtai is built with Python 3.9+, [Hugging Face Transformers](https://github.com/huggingface/transformers), [Sentence Transformers](https://github.com/UKPLab/sentence-transformers) and [FastAPI](https://github.com/tiangolo/fastapi). txtai is open-source under an Apache 2.0 license.
+txtai is built with Python 3.11+, [Hugging Face Transformers](https://github.com/huggingface/transformers), [Sentence Transformers](https://github.com/UKPLab/sentence-transformers) and [FastAPI](https://github.com/tiangolo/fastapi). txtai is open-source under an Apache 2.0 license.
 
-*Interested in an easy and secure way to run hosted txtai applications? Then join the [txtai.cloud](https://txtai.cloud) preview to learn more.*
+!!! note
+
+    [NeuML](https://neuml.com) is the company behind txtai and we provide AI consulting services around our stack. [Schedule a meeting](https://cal.com/neuml/intro) or [send a message](mailto:[email protected]) to learn more.
+
+    We're also building an easy and secure way to run hosted txtai applications with [txtai.cloud](https://txtai.cloud).

docs/install.md

@@ -9,7 +9,7 @@ The easiest way to install is via pip and PyPI
 pip install txtai
 ```
 
-Python 3.9+ is supported. Using a Python [virtual environment](https://docs.python.org/3/library/venv.html) is recommended.
+Python 3.11+ is supported. Using a Python [virtual environment](https://docs.python.org/3/library/venv.html) is recommended.
 
 ## Optional dependencies
 
@@ -130,7 +130,11 @@ Additional environment specific prerequisites are below.
 
 ### Linux
 
-The AudioStream and Microphone pipelines require the [PortAudio](https://python-sounddevice.readthedocs.io/en/0.5.0/installation.html) system library. The Transcription pipeline requires the [SoundFile](https://github.com/bastibe/python-soundfile#installation) system library.
+The AudioStream and Microphone pipelines require the [PortAudio](https://python-sounddevice.readthedocs.io/en/0.5.0/installation.html) system library.
+
+The Transcription pipeline requires the [SoundFile](https://github.com/bastibe/python-soundfile#installation) system library.
+
+The LiteRT LLM pipeline requires libegl1, libgles2 and libvulkan1.
 
 ### macOS
 
@@ -146,7 +150,7 @@ The [txtai build workflow](https://github.com/neuml/txtai/blob/master/.github/wo
 
 ## CPU-only
 
-The default install adds PyTorch with GPU support. There are a number of dependencies that come with that. When running in a CPU-only environment or using Embeddings/LLM models without PyTorch (i.e. llama.cpp or API services), the CPU-only PyTorch package can be installed with txtai as follows.
+The default install adds PyTorch with GPU support. There are a number of dependencies that come with that. When running in a CPU-only environment, the CPU-only PyTorch package can be installed with txtai as follows.
 
 ```
 pip install txtai torch==[version]+cpu \
@@ -173,6 +177,22 @@ A [community-supported txtai package](https://anaconda.org/conda-forge/txtai) is
 conda install -c conda-forge txtai
 ```
 
+## Minimal install
+
+A lightweight zero dependency minimal install package is available. This is helpful if the default libraries are too heavy or there are no plans to use them.
+
+All the same extras mentioned above are available in this version of the package.
+
+```
+# Lightweight minimal install
+pip install txtai_minimal
+
+# Same as standard `txtai` install
+pip install txtai_minimal[default]
+```
+
+Note that in order to use the `Embeddings` or `LLM` interface without `torch` requires either `llama.cpp`, `litellm` or `litert`.
+
 ## Run with containers
 
 Docker images are available for txtai. [See this section](../cloud) for more information on container-based installs.

docs/models.md

@@ -8,9 +8,9 @@ See the table below for the current recommended models. These models all allow c
 | ---------------------------------------------------- | ------------------------------------------------------------------------ |
 | [Embeddings](../embeddings)                          | [all-MiniLM-L6-v2](https://hf.co/sentence-transformers/all-MiniLM-L6-v2) | 
 | [Image Captions](./pipeline/image/caption.md)        | [BLIP](https://hf.co/Salesforce/blip-image-captioning-base)              |
-| [Labels - Zero Shot](./pipeline/text/labels.md)      | [BART-Large-MNLI](https://hf.co/facebook/bart-large)                     |
+| [Labels - Zero Shot](./pipeline/text/labels.md)      | [DeBERTa v3 Zeroshot](https://hf.co/MoritzLaurer/deberta-v3-base-zeroshot-v2.0-c)                     |
 | [Labels - Fixed](./pipeline/text/labels.md)          | Fine-tune with [training pipeline](./pipeline/train/trainer.md)          |
-| [Large Language Model (LLM)](./pipeline/text/llm.md) | [Llama 3.1 Instruct](https://hf.co/meta-llama/Llama-3.1-8B-Instruct)     |
+| [Large Language Model (LLM)](./pipeline/text/llm.md) | [Gemma 4 31B](https://hf.co/google/gemma-4-31B)                          |
 | [Summarization](./pipeline/text/summary.md)          | [DistilBART](https://hf.co/sshleifer/distilbart-cnn-12-6)                |
 | [Text-to-Speech](./pipeline/audio/texttospeech.md)   | [ESPnet JETS](https://hf.co/NeuML/ljspeech-jets-onnx)                    |
 | [Transcription](./pipeline/audio/transcription.md)   | [Whisper](https://hf.co/openai/whisper-base)                             | 
@@ -22,6 +22,5 @@ See the following links to learn more.
 
 - [Hugging Face Tasks](https://hf.co/tasks)
 - [Hugging Face Model Hub](https://hf.co/models)
-- [MTEB Leaderboard](https://hf.co/spaces/mteb/leaderboard)
-- [LMSYS LLM Leaderboard](https://chat.lmsys.org/?leaderboard)
-- [Open LLM Leaderboard](https://hf.co/spaces/HuggingFaceH4/open_llm_leaderboard)
+- [Embeddings Leaderboard](https://hf.co/spaces/mteb/leaderboard)
+- [LLM Leaderboard](https://hf.co/spaces/lmarena-ai/arena-leaderboard)

docs/observability.md

@@ -2,7 +2,7 @@
 
 ![agent](https://raw.githubusercontent.com/neuml/mlflow-txtai/master/images/agent.png)
 
-Observability helps track the inner workings of a system without having to change the system. This makes it much easier to debug and evaluate overall performance.
+Observability enables tracking the inner workings of a system without having to change the system. This makes it much easier to debug and evaluate overall performance.
 
 `txtai` has an integration with [MLflow](https://mlflow.org) and it's [tracing module](https://mlflow.org/docs/latest/llms/tracing/index.html) to provide insights into each of the components in `txtai`.
 
@@ -88,7 +88,7 @@ with mlflow.start_run():
     # Create RAG pipeline
     rag = RAG(
         wiki,
-        "hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
+        "openai/gpt-oss-20b",
         system="You are a friendly assistant. You answer questions from users.",
         template=template,
         context=10
@@ -155,8 +155,8 @@ embeddings.load(provider="huggingface-hub", container="neuml/txtai-astronomy")
 
 agent = Agent(
     tools=[search],
-    llm="hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4",
-    max_iterations=10,
+    llm="Qwen/Qwen3-4B-Instruct-2507",
+    max_steps=10,
 )
 
 researcher = """

docs/pipeline/data/filetohtml.md

@@ -17,6 +17,12 @@ Apache Tika requires [Java](https://en.wikipedia.org/wiki/Java_(programming_lang
 
 See [this link](https://github.com/DS4SD/docling?tab=readme-ov-file#features) for a list of supported document formats.
 
+## LiteParse
+
+[LiteParse](https://github.com/run-llama/liteparse) is a fast, helpful, and open-source document parser. It supports PDF out of the box with additional format support via third-party packages.
+
+See [this link](https://github.com/run-llama/liteparse#multi-format-input-support) for more.
+
 ## Example
 
 The following shows a simple example using this pipeline.

docs/pipeline/data/segmentation.md

@@ -16,14 +16,12 @@ from txtai.pipeline import Segmentation
 segment = Segmentation(sentences=True)
 segment("This is a test. And another test.")
 
-# Segment text with Chonkie chunker (word, sentence, semantic, late etc)
+# Load third-party chunkers
 segment = Segmentation(chunker="semantic")
 segment("This is a test. And another test.")
 ```
 
-The Segmentation pipeline supports segmenting `sentences`, `lines`, `paragraphs` and `sections` using a rules-based approach. Each of these modes can be set when creating the pipeline.
-
-More advanced functionality is supported via a [Chonkie chunker](https://docs.chonkie.ai/chunkers/overview). The `chunker` keyword dynamically creates a Chonkie chunker. For example, `chunker='token'` creates a [TokenChunker](https://docs.chonkie.ai/chunkers/token-chunker), `chunker='semantic'` creates a [SemanticChunker](https://docs.chonkie.ai/chunkers/semantic-chunker) and so forth. Additional keyword arguments are passed to the chunker.
+The Segmentation pipeline supports segmenting `sentences`, `lines`, `paragraphs` and `sections` using a rules-based approach. Each of these modes can be set when creating the pipeline. Third-party chunkers are also supported via the `chunker` parameter.
 
 ## Configuration-driven example
 

docs/pipeline/data/textractor.md

@@ -16,6 +16,8 @@ The [backend](../filetohtml/#txtai.pipeline.FileToHTML.__init__) parameter sets
 
 See the [FiletoHTML](../filetohtml) and [HTMLToMarkdown](../htmltomd) pipelines to learn more on the dependencies necessary for each of those pipelines.
 
+Note that the default parameters enable access to all local files and all URLs. The `safeopen` parameter limits this to only files in a specified directory (defaults to temp directory) and public URLs. When running the textractor pipeline through an Application or via the API, `safeopen` defaults to True. It can be disabled if desired via configuration.
+
 ## Example
 
 The following shows a simple example using this pipeline.

docs/pipeline/data/tokenizer.md

@@ -0,0 +1,82 @@
+# Tokenizer
+
+![pipeline](../../images/pipeline.png#only-light)
+![pipeline](../../images/pipeline-dark.png#only-dark)
+
+The Tokenizer pipeline splits text into tokens. This is primarily used for keyword / term indexing.
+
+_Note: Transformers-based models have their own tokenizers and this pipeline isn't designed for working with Transformers models._
+
+## Example
+
+The following shows a simple example using this pipeline.
+
+```python
+from txtai.pipeline import Tokenizer
+
+# Create and run pipeline
+tokenizer = Tokenizer()
+tokenizer("text to tokenize")
+
+# Whitespace tokenization
+tokenizer = Tokenizer(whitespace=True)
+tokenizer("text to tokenize")
+
+# Tokenize using a regular expression
+tokenizer = Tokenizer(regexp=r"\w{5,}")
+tokenizer("text to tokenize")
+
+# Tokenize into trigrams like pg_trgm
+tokenizer = Tokenizer(ngrams={
+  "ngrams": 3, "lpad": "  ", "rpad": " ", "unique": True
+})
+tokenize("text to tokenize")
+
+# Tokenize into edge ngrams
+tokenizer = Tokenizer(ngrams={"nmin": 2, "nmax": 5, "edge": True})
+tokenizer("text to tokenize")
+```
+
+## Configuration-driven example
+
+Pipelines are run with Python or configuration. Pipelines can be instantiated in [configuration](../../../api/configuration/#pipeline) using the lower case name of the pipeline. Configuration-driven pipelines are run with [workflows](../../../workflow/#configuration-driven-example) or the [API](../../../api#local-instance).
+
+### config.yml
+```yaml
+# Create pipeline using lower case class name
+tokenizer:
+
+# Run pipeline with workflow
+workflow:
+  tokenizer:
+    tasks:
+      - action: tokenizer
+```
+
+### Run with Workflows
+
+```python
+from txtai import Application
+
+# Create and run pipeline with workflow
+app = Application("config.yml")
+list(app.workflow("tokenizer", ["text to tokenize"]))
+```
+
+### Run with API
+
+```bash
+CONFIG=config.yml uvicorn "txtai.api:app" &
+
+curl \
+  -X POST "http://localhost:8000/workflow" \
+  -H "Content-Type: application/json" \
+  -d '{"name":"tokenizer", "elements":["text"]}'
+```
+
+## Methods
+
+Python documentation for the pipeline.
+
+### ::: txtai.pipeline.Tokenizer.__init__
+### ::: txtai.pipeline.Tokenizer.__call__

docs/pipeline/data/urlretrieve.md

@@ -0,0 +1,62 @@
+# URL Retrieve
+
+![pipeline](../../images/pipeline.png#only-light)
+![pipeline](../../images/pipeline-dark.png#only-dark)
+
+The URL Retrieve pipeline retrieves content from a HTTP(s) URL.
+
+## Example
+
+The following shows a simple example using this pipeline.
+
+```python
+from txtai.pipeline import URLRetrieve
+
+# Create and run pipeline
+urlretrieve = URLRetrieve()
+urlretrieve("https://github.com/neuml/txtai")
+```
+
+## Configuration-driven example
+
+Pipelines are run with Python or configuration. Pipelines can be instantiated in [configuration](../../../api/configuration/#pipeline) using the lower case name of the pipeline. Configuration-driven pipelines are run with [workflows](../../../workflow/#configuration-driven-example) or the [API](../../../api#local-instance).
+
+### config.yml
+```yaml
+# Create pipeline using lower case class name
+urlretrieve:
+
+# Run pipeline with workflow
+workflow:
+  retrieve:
+    tasks:
+      - action: urlretrieve
+```
+
+### Run with Workflows
+
+```python
+from txtai import Application
+
+# Create and run pipeline with workflow
+app = Application("config.yml")
+list(app.workflow("urlretrieve", ["https://github.com/neuml/txtai"]))
+```
+
+### Run with API
+
+```bash
+CONFIG=config.yml uvicorn "txtai.api:app" &
+
+curl \
+  -X POST "http://localhost:8000/workflow" \
+  -H "Content-Type: application/json" \
+  -d '{"name":"retrieve", "elements":["http://github.com/neuml/txtai"]}'
+```
+
+## Methods
+
+Python documentation for the pipeline.
+
+### ::: txtai.pipeline.URLRetrieve.__init__
+### ::: txtai.pipeline.URLRetrieve.__call__

docs/pipeline/index.md

@@ -26,6 +26,8 @@ The `LLM` and `RAG` pipelines also have integrations for [llama.cpp](https://git
     - [Segmentation](data/segmentation)
     - [Tabular](data/tabular)
     - [Text extraction](data/textractor)
+    - [Tokenizer](data/tokenizer)
+    - [URLRetrieve](data/urlretrieve)
 - Image
     - [Caption](image/caption)
     - [Image Hash](image/imagehash)
@@ -35,6 +37,7 @@ The `LLM` and `RAG` pipelines also have integrations for [llama.cpp](https://git
     - [Labeling](text/labels)
     - [LLM](text/llm)
     - [RAG](text/rag)
+    - [Reranker](text/reranker)
     - [Similarity](text/similarity)
     - [Summary](text/summary)
     - [Translation](text/translation)

docs/pipeline/llm/llm.md

@@ -29,8 +29,8 @@ llm(
   """
 )
 
-# Instruction tuned models typically require string prompts to
-# follow a specific chat template set by the model
+# Prompts with chat templating can be directly passed
+# The template format varies by model
 llm(
   """
   <|im_start|>system
@@ -47,41 +47,70 @@ llm([
   {"role": "user", "content": "Answer the following question..."}
 ])
 
-# Set the default role to user and string inputs are converted to chat messages
+# When there is no system prompt passed to instruction tuned models
+# the default role is inferred `defaultrole="auto"`
+llm("Answer the following question...")
+
+# To always generate chat messages for string inputs
 llm("Answer the following question...", defaultrole="user")
+
+# To never generate chat messages for string inputs
+llm("Answer the following question...", defaultrole="prompt")
 ```
 
-The LLM pipeline automatically detects the underlying LLM framework. This can also be manually set.
+The LLM pipeline automatically detects the underlying LLM framework. This can also be manually set. The following methods are supported.
 
-[Hugging Face Transformers](https://github.com/huggingface/transformers), [llama.cpp](https://github.com/abetlen/llama-cpp-python) and [hosted API models via LiteLLM](https://github.com/BerriAI/litellm) are all supported by this pipeline.
+- [Hugging Face Transformers](https://github.com/huggingface/transformers)
+- [llama.cpp](https://github.com/abetlen/llama-cpp-python)
+- [LLM APIs via LiteLLM](https://github.com/BerriAI/litellm)
+- [LiteRT-LM](https://ai.google.dev/edge/litert-lm)
+- [OpenCode server](https://github.com/anomalyco/opencode)
 
-See the [LiteLLM documentation](https://litellm.vercel.app/docs/providers) for the options available with LiteLLM models. llama.cpp models support both local and remote GGUF paths on the HF Hub.
+`llama.cpp` and `LiteRT-LM` support both local and remote file paths on the HF Hub.
+
+See the [LiteLLM documentation](https://litellm.vercel.app/docs/providers) for the options available with LiteLLM models.
+
+See the [OpenCode documentation](https://opencode.ai/docs/server/) for more on how to integrate the LLM pipeline with a running OpenCode instance.
 
 ```python
 from txtai import LLM
 
 # Transformers
-llm = LLM("meta-llama/Meta-Llama-3.1-8B-Instruct")
-llm = LLM("meta-llama/Meta-Llama-3.1-8B-Instruct", method="transformers")
+llm = LLM("openai/gpt-oss-20b")
+llm = LLM("openai/gpt-oss-20b", method="transformers")
 
 # llama.cpp
-llm = LLM("microsoft/Phi-3-mini-4k-instruct-gguf/Phi-3-mini-4k-instruct-q4.gguf")
-llm = LLM("microsoft/Phi-3-mini-4k-instruct-gguf/Phi-3-mini-4k-instruct-q4.gguf",
+llm = LLM("unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf")
+llm = LLM("unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf",
            method="llama.cpp")
 
 # LiteLLM
-llm = LLM("ollama/llama3.1")
-llm = LLM("ollama/llama3.1", method="litellm")
+llm = LLM("ollama/gpt-oss")
+llm = LLM("ollama/gpt-oss", method="litellm")
+
+# LiteRT-LM
+llm = LLM(
+  "litert-community/gemma-4-E2B-it-litert-lm/gemma-4-E2B-it.litertlm"
+)
+llm = LLM(
+  "litert-community/gemma-4-E2B-it-litert-lm/gemma-4-E2B-it.litertlm",
+  method="litert"
+)
 
 # Custom Ollama endpoint
-llm = LLM("ollama/llama3.1", api_base="http://localhost:11434")
+llm = LLM("ollama/gpt-oss", api_base="http://localhost:11434")
 
 # Custom OpenAI-compatible endpoint
-llm = LLM("openai/llama3.1", api_base="http://localhost:4000")
+llm = LLM("openai/gpt-oss", api_base="http://localhost:4000")
 
 # LLM APIs - must also set API key via environment variable
-llm = LLM("gpt-4o")
-llm = LLM("claude-3-5-sonnet-20240620")
+llm = LLM("gpt-5.2")
+llm = LLM("claude-opus-4-5-20251101")
+llm = LLM("gemini/gemini-3-pro-preview")
+
+# Local OpenCode server started via `opencode serve`
+llm = LLM("opencode")
+llm = LLM("opencode/big-pickle", url="http://localhost:4000")
 ```
 
 Models can be externally loaded and passed to pipelines. This is useful for models that are not yet supported by Transformers and/or need special initialization.
@@ -92,11 +121,11 @@ import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer
 from txtai import LLM
 
-# Load Phi 3.5-mini
-path = "microsoft/Phi-3.5-mini-instruct"
+# Load Qwen3 0.6B
+path = "Qwen/Qwen3-0.6B"
 model = AutoModelForCausalLM.from_pretrained(
   path,
-  torch_dtype=torch.bfloat16,
+  dtype=torch.bfloat16,
 )
 tokenizer = AutoTokenizer.from_pretrained(path)
 
@@ -109,7 +138,7 @@ See the links below for more detailed examples.
 |:----------|:-------------|------:|
 | [Prompt-driven search with LLMs](https://github.com/neuml/txtai/blob/master/examples/42_Prompt_driven_search_with_LLMs.ipynb) | Embeddings-guided and Prompt-driven search with Large Language Models (LLMs) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/42_Prompt_driven_search_with_LLMs.ipynb) |
 | [Prompt templates and task chains](https://github.com/neuml/txtai/blob/master/examples/44_Prompt_templates_and_task_chains.ipynb) | Build model prompts and connect tasks together with workflows | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/44_Prompt_templates_and_task_chains.ipynb) |
-| [Build RAG pipelines with txtai](https://github.com/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) | Guide on retrieval augmented generation including how to create citations | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) |
+| [Build RAG pipelines with txtai](https://github.com/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) [▶️](https://www.youtube.com/watch?v=t_OeAc8NVfQ) | Guide on retrieval augmented generation including how to create citations | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) |
 | [Integrate LLM frameworks](https://github.com/neuml/txtai/blob/master/examples/53_Integrate_LLM_Frameworks.ipynb) | Integrate llama.cpp, LiteLLM and custom generation frameworks | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/53_Integrate_LLM_Frameworks.ipynb) |
 | [Generate knowledge with Semantic Graphs and RAG](https://github.com/neuml/txtai/blob/master/examples/55_Generate_knowledge_with_Semantic_Graphs_and_RAG.ipynb) | Knowledge exploration and discovery with Semantic Graphs and RAG | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/55_Generate_knowledge_with_Semantic_Graphs_and_RAG.ipynb) |
 | [Build knowledge graphs with LLMs](https://github.com/neuml/txtai/blob/master/examples/57_Build_knowledge_graphs_with_LLM_driven_entity_extraction.ipynb) | Build knowledge graphs with LLM-driven entity extraction | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/57_Build_knowledge_graphs_with_LLM_driven_entity_extraction.ipynb) |
@@ -118,13 +147,19 @@ See the links below for more detailed examples.
 | [RAG with llama.cpp and external API services](https://github.com/neuml/txtai/blob/master/examples/62_RAG_with_llama_cpp_and_external_API_services.ipynb) | RAG with additional vector and LLM frameworks | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/62_RAG_with_llama_cpp_and_external_API_services.ipynb) |
 | [How RAG with txtai works](https://github.com/neuml/txtai/blob/master/examples/63_How_RAG_with_txtai_works.ipynb) | Create RAG processes, API services and Docker instances | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/63_How_RAG_with_txtai_works.ipynb) |
 | [Speech to Speech RAG](https://github.com/neuml/txtai/blob/master/examples/65_Speech_to_Speech_RAG.ipynb) [▶️](https://www.youtube.com/watch?v=tH8QWwkVMKA) | Full cycle speech to speech workflow with RAG | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/65_Speech_to_Speech_RAG.ipynb) |
-| [Generative Audio](https://github.com/neuml/txtai/blob/master/examples/66_Generative_Audio.ipynb) | Storytelling with generative audio workflows | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/66_Generative_Audio.ipynb) |
 | [Analyzing Hugging Face Posts with Graphs and Agents](https://github.com/neuml/txtai/blob/master/examples/68_Analyzing_Hugging_Face_Posts_with_Graphs_and_Agents.ipynb) | Explore a rich dataset with Graph Analysis and Agents | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/68_Analyzing_Hugging_Face_Posts_with_Graphs_and_Agents.ipynb) |
 | [Granting autonomy to agents](https://github.com/neuml/txtai/blob/master/examples/69_Granting_autonomy_to_agents.ipynb) | Agents that iteratively solve problems as they see fit | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/69_Granting_autonomy_to_agents.ipynb) |
 | [Getting started with LLM APIs](https://github.com/neuml/txtai/blob/master/examples/70_Getting_started_with_LLM_APIs.ipynb) | Generate embeddings and run LLMs with OpenAI, Claude, Gemini, Bedrock and more | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/70_Getting_started_with_LLM_APIs.ipynb) |
 | [Analyzing LinkedIn Company Posts with Graphs and Agents](https://github.com/neuml/txtai/blob/master/examples/71_Analyzing_LinkedIn_Company_Posts_with_Graphs_and_Agents.ipynb) | Exploring how to improve social media engagement with AI | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/71_Analyzing_LinkedIn_Company_Posts_with_Graphs_and_Agents.ipynb) |
 | [Parsing the stars with txtai](https://github.com/neuml/txtai/blob/master/examples/72_Parsing_the_stars_with_txtai.ipynb) | Explore an astronomical knowledge graph of known stars, planets, galaxies | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/72_Parsing_the_stars_with_txtai.ipynb) |
 | [Chunking your data for RAG](https://github.com/neuml/txtai/blob/master/examples/73_Chunking_your_data_for_RAG.ipynb) | Extract, chunk and index content for effective retrieval | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/73_Chunking_your_data_for_RAG.ipynb) |
+| [Medical RAG Research with txtai](https://github.com/neuml/txtai/blob/master/examples/75_Medical_RAG_Research_with_txtai.ipynb) | Analyze PubMed article metadata with RAG | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/75_Medical_RAG_Research_with_txtai.ipynb) |
+| [GraphRAG with Wikipedia and GPT OSS](https://github.com/neuml/txtai/blob/master/examples/77_GraphRAG_with_Wikipedia_and_GPT_OSS.ipynb) | Deep graph search powered RAG | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/77_GraphRAG_with_Wikipedia_and_GPT_OSS.ipynb) |
+| [RAG is more than Vector Search](https://github.com/neuml/txtai/blob/master/examples/79_RAG_is_more_than_Vector_Search.ipynb) | Context retrieval via Web, SQL and other sources | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/79_RAG_is_more_than_Vector_Search.ipynb) |
+| [OpenCode as a txtai LLM](https://github.com/neuml/txtai/blob/master/examples/81_OpenCode_as_a_txtai_LLM.ipynb) | Integrate OpenCode with the txtai ecosystem | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/81_OpenCode_as_a_txtai_LLM.ipynb) |
+| [Agentic College Search](https://github.com/neuml/txtai/blob/master/examples/82_Agentic_College_Search.ipynb) | Identify a list of strong engineering colleges | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/82_Agentic_College_Search.ipynb) |
+| [TxtAI got skills](https://github.com/neuml/txtai/blob/master/examples/83_TxtAI_got_skills.ipynb) | Integrate skill.md files with your agent | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/83_TxtAI_got_skills.ipynb) |
+| [Agent Tools](https://github.com/neuml/txtai/blob/master/examples/84_Agent_Tools.ipynb) [▶️](https://www.youtube.com/watch?v=RDNaFXQy3GQ) | Learn about the txtai agent toolkit | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/84_Agent_Tools.ipynb) |
 
 ## Configuration-driven example
 
@@ -146,8 +181,8 @@ Similar to the Python example above, the underlying [Hugging Face pipeline param
 
 ```yaml
 llm:
-  path: microsoft/Phi-3.5-mini-instruct
-  torch_dtype: torch.bfloat16
+  path: Qwen/Qwen3-0.6B
+  dtype: torch.bfloat16
 ```
 
 ### Run with Workflows

docs/pipeline/llm/rag.md

@@ -3,9 +3,9 @@
 ![pipeline](../../images/pipeline.png#only-light)
 ![pipeline](../../images/pipeline-dark.png#only-dark)
 
-The RAG pipeline (aka Extractor) joins a prompt, context data store and generative model together to extract knowledge.
+The Retrieval Augmented Generation (RAG) pipeline joins a prompt, context data store and generative model together to extract knowledge.
 
-The data store can be an embeddings database or a similarity instance with associated input text. The generative model can be a prompt-driven large language model (LLM), an extractive question-answering model or a custom pipeline. This is known as retrieval augmented generation (RAG).
+The data store can be an embeddings database or a similarity instance with associated input text. The generative model can be a prompt-driven large language model (LLM), an extractive question-answering model or a custom pipeline.
 
 ## Example
 
@@ -30,8 +30,8 @@ data = [
 embeddings = Embeddings(content=True)
 embeddings.index(data)
 
-# Create and run pipeline
-rag = RAG(embeddings, "google/flan-t5-base", template="""
+# Create the RAG pipeline
+rag = RAG(embeddings, "Qwen/Qwen3-0.6B", template="""
   Answer the following question using the provided context.
 
   Question:
@@ -41,11 +41,12 @@ rag = RAG(embeddings, "google/flan-t5-base", template="""
   {context}
 """)
 
+# Run RAG pipeline
 rag("What was won?")
 
-# Instruction tuned models typically require string prompts to
-# follow a specific chat template set by the model
-rag = RAG(embeddings, "meta-llama/Meta-Llama-3.1-8B-Instruct", template="""
+# Prompts with chat templating can be directly passed
+# The template format varies by model
+rag = RAG(embeddings, "Qwen/Qwen3-0.6B", template="""
   <|im_start|>system
   You are a friendly assistant.<|im_end|>
   <|im_start|>user
@@ -61,8 +62,13 @@ rag = RAG(embeddings, "meta-llama/Meta-Llama-3.1-8B-Instruct", template="""
 )
 rag("What was won?")
 
-# LLM options can be passed as additional arguments
-rag = RAG(embeddings, "meta-llama/Meta-Llama-3.1-8B-Instruct", template="""
+# Inputs are automatically converted to chat messages when a
+# system prompt is provided
+rag = RAG(
+  embeddings,
+  "openai/gpt-oss-20b",
+  system="You are a friendly assistant",
+  template="""
   Answer the following question using the provided context.
 
   Question:
@@ -71,38 +77,35 @@ rag = RAG(embeddings, "meta-llama/Meta-Llama-3.1-8B-Instruct", template="""
   Context:
   {context}
 """)
+rag("What was won?")
 
-# Set the default role to user and string inputs are converted to chat messages
-rag("What was won?", defaultrole="user")
+# LLM options can be passed as additional arguments
+#  - Streaming RAG response with `stream=True`
+#  - String inputs are always converted to user messages with `defaultrole="user"`
+#  - Thinking text is removed with `stripthink=True`
+rag("What was won?", stream=True, defaultrole="user", stripThink=True)
 ```
 
 See the [Embeddings](../../../embeddings) and [LLM](../llm) pages for additional configuration options.
 
-See the links below for more detailed examples.
+Check out this [RAG Quickstart Example](https://github.com/neuml/txtai/blob/master/examples/rag_quickstart.py). Additional examples are listed below.
 
 | Notebook  | Description  |       |
 |:----------|:-------------|------:|
 | [Prompt-driven search with LLMs](https://github.com/neuml/txtai/blob/master/examples/42_Prompt_driven_search_with_LLMs.ipynb) | Embeddings-guided and Prompt-driven search with Large Language Models (LLMs) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/42_Prompt_driven_search_with_LLMs.ipynb) |
-| [Prompt templates and task chains](https://github.com/neuml/txtai/blob/master/examples/44_Prompt_templates_and_task_chains.ipynb) | Build model prompts and connect tasks together with workflows | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/44_Prompt_templates_and_task_chains.ipynb) |
-| [Build RAG pipelines with txtai](https://github.com/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) | Guide on retrieval augmented generation including how to create citations | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) |
+| [Build RAG pipelines with txtai](https://github.com/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) [▶️](https://www.youtube.com/watch?v=t_OeAc8NVfQ) | Guide on retrieval augmented generation including how to create citations | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) |
 | [Integrate LLM frameworks](https://github.com/neuml/txtai/blob/master/examples/53_Integrate_LLM_Frameworks.ipynb) | Integrate llama.cpp, LiteLLM and custom generation frameworks | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/53_Integrate_LLM_Frameworks.ipynb) |
 | [Generate knowledge with Semantic Graphs and RAG](https://github.com/neuml/txtai/blob/master/examples/55_Generate_knowledge_with_Semantic_Graphs_and_RAG.ipynb) | Knowledge exploration and discovery with Semantic Graphs and RAG | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/55_Generate_knowledge_with_Semantic_Graphs_and_RAG.ipynb) |
-| [Build knowledge graphs with LLMs](https://github.com/neuml/txtai/blob/master/examples/57_Build_knowledge_graphs_with_LLM_driven_entity_extraction.ipynb) | Build knowledge graphs with LLM-driven entity extraction | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/57_Build_knowledge_graphs_with_LLM_driven_entity_extraction.ipynb) |
 | [Advanced RAG with graph path traversal](https://github.com/neuml/txtai/blob/master/examples/58_Advanced_RAG_with_graph_path_traversal.ipynb) | Graph path traversal to collect complex sets of data for advanced RAG | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/58_Advanced_RAG_with_graph_path_traversal.ipynb) |
 | [Advanced RAG with guided generation](https://github.com/neuml/txtai/blob/master/examples/60_Advanced_RAG_with_guided_generation.ipynb) | Retrieval Augmented and Guided Generation | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/60_Advanced_RAG_with_guided_generation.ipynb) |
 | [RAG with llama.cpp and external API services](https://github.com/neuml/txtai/blob/master/examples/62_RAG_with_llama_cpp_and_external_API_services.ipynb) | RAG with additional vector and LLM frameworks | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/62_RAG_with_llama_cpp_and_external_API_services.ipynb) |
 | [How RAG with txtai works](https://github.com/neuml/txtai/blob/master/examples/63_How_RAG_with_txtai_works.ipynb) | Create RAG processes, API services and Docker instances | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/63_How_RAG_with_txtai_works.ipynb) |
 | [Speech to Speech RAG](https://github.com/neuml/txtai/blob/master/examples/65_Speech_to_Speech_RAG.ipynb) [▶️](https://www.youtube.com/watch?v=tH8QWwkVMKA) | Full cycle speech to speech workflow with RAG | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/65_Speech_to_Speech_RAG.ipynb) |
-| [Generative Audio](https://github.com/neuml/txtai/blob/master/examples/66_Generative_Audio.ipynb) | Storytelling with generative audio workflows | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/66_Generative_Audio.ipynb) |
-| [Analyzing Hugging Face Posts with Graphs and Agents](https://github.com/neuml/txtai/blob/master/examples/68_Analyzing_Hugging_Face_Posts_with_Graphs_and_Agents.ipynb) | Explore a rich dataset with Graph Analysis and Agents | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/68_Analyzing_Hugging_Face_Posts_with_Graphs_and_Agents.ipynb) |
-| [Granting autonomy to agents](https://github.com/neuml/txtai/blob/master/examples/69_Granting_autonomy_to_agents.ipynb) | Agents that iteratively solve problems as they see fit | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/69_Granting_autonomy_to_agents.ipynb) |
-| [Getting started with LLM APIs](https://github.com/neuml/txtai/blob/master/examples/70_Getting_started_with_LLM_APIs.ipynb) | Generate embeddings and run LLMs with OpenAI, Claude, Gemini, Bedrock and more | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/70_Getting_started_with_LLM_APIs.ipynb) |
-| [Analyzing LinkedIn Company Posts with Graphs and Agents](https://github.com/neuml/txtai/blob/master/examples/71_Analyzing_LinkedIn_Company_Posts_with_Graphs_and_Agents.ipynb) | Exploring how to improve social media engagement with AI | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/71_Analyzing_LinkedIn_Company_Posts_with_Graphs_and_Agents.ipynb) |
-| [Extractive QA with txtai](https://github.com/neuml/txtai/blob/master/examples/05_Extractive_QA_with_txtai.ipynb) | Introduction to extractive question-answering with txtai | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/05_Extractive_QA_with_txtai.ipynb) |
-| [Extractive QA with Elasticsearch](https://github.com/neuml/txtai/blob/master/examples/06_Extractive_QA_with_Elasticsearch.ipynb) | Run extractive question-answering queries with Elasticsearch | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/06_Extractive_QA_with_Elasticsearch.ipynb) |
-| [Extractive QA to build structured data](https://github.com/neuml/txtai/blob/master/examples/20_Extractive_QA_to_build_structured_data.ipynb) | Build structured datasets using extractive question-answering | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/20_Extractive_QA_to_build_structured_data.ipynb) |
 | [Parsing the stars with txtai](https://github.com/neuml/txtai/blob/master/examples/72_Parsing_the_stars_with_txtai.ipynb) | Explore an astronomical knowledge graph of known stars, planets, galaxies | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/72_Parsing_the_stars_with_txtai.ipynb) |
 | [Chunking your data for RAG](https://github.com/neuml/txtai/blob/master/examples/73_Chunking_your_data_for_RAG.ipynb) | Extract, chunk and index content for effective retrieval | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/73_Chunking_your_data_for_RAG.ipynb) |
+| [Medical RAG Research with txtai](https://github.com/neuml/txtai/blob/master/examples/75_Medical_RAG_Research_with_txtai.ipynb) | Analyze PubMed article metadata with RAG | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/75_Medical_RAG_Research_with_txtai.ipynb) |
+| [GraphRAG with Wikipedia and GPT OSS](https://github.com/neuml/txtai/blob/master/examples/77_GraphRAG_with_Wikipedia_and_GPT_OSS.ipynb) | Deep graph search powered RAG | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/77_GraphRAG_with_Wikipedia_and_GPT_OSS.ipynb) |
+| [RAG is more than Vector Search](https://github.com/neuml/txtai/blob/master/examples/79_RAG_is_more_than_Vector_Search.ipynb) | Context retrieval via Web, SQL and other sources | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/79_RAG_is_more_than_Vector_Search.ipynb) |
 
 ## Configuration-driven example
 
@@ -118,7 +121,7 @@ embeddings:
   content: True
 
 rag:
-  path: google/flan-t5-base
+  path: Qwen/Qwen3-0.6B
   template: |
     Answer the following question using the provided context.
 

docs/pipeline/text/reranker.md

@@ -0,0 +1,82 @@
+# Reranker
+
+![pipeline](../../images/pipeline.png#only-light)
+![pipeline](../../images/pipeline-dark.png#only-dark)
+
+The Reranker pipeline runs embeddings queries and re-ranks them using a similarity pipeline. 
+
+## Example
+
+The following shows a simple example using this pipeline.
+
+```python
+from txtai import Embeddings
+from txtai.pipeline import Reranker, Similarity
+
+# Embeddings instance
+embeddings = Embeddings()
+embeddings.load(provider="huggingface-hub", container="neuml/txtai-wikipedia")
+
+# Similarity instance
+similarity = Similarity(path="colbert-ir/colbertv2.0", lateencode=True)
+
+# Reranking pipeline
+reranker = Reranker(embeddings, similarity)
+reranker("Tell me about AI")
+```
+
+_Note: Content must be enabled with the embeddings instance for this to work properly._
+
+See the link below for a more detailed example.
+
+| Notebook  | Description  |       |
+|:----------|:-------------|------:|
+| [What's new in txtai 9.0](https://github.com/neuml/txtai/blob/master/examples/76_Whats_new_in_txtai_9_0.ipynb) | Learned sparse vectors, late interaction models and rerankers | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/76_Whats_new_in_txtai_9_0.ipynb) |
+
+## Configuration-driven example
+
+Pipelines are run with Python or configuration. Pipelines can be instantiated in [configuration](../../../api/configuration/#pipeline) using the lower case name of the pipeline. Configuration-driven pipelines are run with [workflows](../../../workflow/#configuration-driven-example) or the [API](../../../api#local-instance).
+
+### config.yml
+```yaml
+embeddings:
+
+similarity:
+
+# Create pipeline using lower case class name
+reranker:
+
+# Run pipeline with workflow
+workflow:
+  translate:
+    tasks:
+      - reranker
+```
+
+### Run with Workflows
+
+```python
+from txtai import Application
+
+# Create and run pipeline with workflow
+app = Application("config.yml")
+list(app.workflow("reranker", ["Tell me about AI"]))
+```
+
+### Run with API
+
+```bash
+CONFIG=config.yml uvicorn "txtai.api:app" &
+
+curl \
+  -X POST "http://localhost:8000/workflow" \
+  -H "Content-Type: application/json" \
+  -d '{"name":"rerank", "elements":["Tell me about AI"]}'
+```
+
+## Methods 
+
+Python documentation for the pipeline.
+
+### ::: txtai.pipeline.Reranker.__init__
+### ::: txtai.pipeline.Reranker.__call__

docs/pipeline/text/similarity.md

@@ -7,7 +7,7 @@ The Similarity pipeline computes similarity between queries and list of text usi
 
 This pipeline supports both standard text classification models and zero-shot classification models. The pipeline uses the queries as labels for the input text. The results are transposed to get scores per query/label vs scores per input text. 
 
-Cross-encoder models are supported via the `crossencode=True` constructor parameter. These models are loaded with a CrossEncoder pipeline that can also be instantiated directly. The CrossEncoder pipeline has the same methods and functionality as described below.
+Cross-encoder models are supported via the `crossencode=True` constructor parameter. Late interaction (i.e. ColBERT) models are also supported via the `lateencode=True` constructor parameter. CrossEncoder and LateEncoder pipelines back each of these models and can be instantiated directly as well.
 
 ## Example
 

docs/pipeline/train/lemur.md

@@ -0,0 +1,98 @@
+# LemurTrainer
+
+![pipeline](../../images/pipeline.png#only-light)
+![pipeline](../../images/pipeline-dark.png#only-dark)
+
+Trains a LEMUR fixed dimensional encoder for a late interaction model and corpus. Training is a separate pipeline step; the saved
+artifact is then loaded through the embeddings `vectors.lemur` configuration.
+
+`LemurTrainer` owns training-data construction, feature-model setup, epoch training and validation-based selection. The loaded LEMUR
+pooler remains responsible for query-feature and document-weight encoding.
+
+## Example
+
+```python
+from txtai.pipeline import LemurTrainer
+
+corpus = [
+    "First document",
+    "Second document",
+    "Third document",
+]
+
+trainer = LemurTrainer()
+trainer(
+    "colbert-ir/colbertv2.0",
+    corpus,
+    "lemur-model",
+    epochs=100,
+    validationsplit=0.1,
+)
+```
+
+`epochs` is required so the training mode is explicit. Use `epochs=100` for the quality-oriented MLP setting. This can take hours on
+a CPU. Use `epochs=0` to select deterministic random ELM features as a lower-cost fallback.
+
+The learn distribution defaults to `learncategory="query"`. Target documents are always encoded with the data encoder, so the
+default encodes selected corpus texts once as data and once as queries. Set `learncategory="data"` to reuse the data encodings, or
+pass a separate iterable of texts with `learn`.
+
+When `vectors.center` is omitted, the trainer encodes the corpus without centering, computes one mean over all corpus token rows and
+uses that mean to center both data and learn vectors before fitting. The collection mean is stored in the LEMUR artifact. An explicit
+`vectors.center` setting keeps its configured behavior and does not store an automatic collection mean.
+
+Set `corpussubsetsize` to a positive integer to sample that many raw corpus texts under `seed` before either encoding pass. The
+default is `None`, which uses every corpus text. `trainsubsetsize` and `learnsubsetsize` apply later, after token vectors have already
+been created. `corpussubsetsize` applies to `data`; a separate `learn` iterable remains caller-sized.
+
+For trained MLP features, set `validationsplit` to a fraction greater than zero and less than one to retain the epoch with the lowest
+held-out loss. The default is `0.0`, which preserves training-loss selection. The selected one-based epoch, loss and metric are
+available as `selectedepoch`, `selectedloss` and `selectionmetric` on the fitted or reloaded encoder.
+
+MLP training displays a `tqdm` progress bar on an interactive terminal, including percent complete and the current validation loss
+when `validationsplit` is enabled (or training loss otherwise). Progress output is disabled for non-interactive runs.
+
+The artifact contains `config.json` and `model.safetensors`. It stores the inference feature model, output-normalization statistics,
+token sample needed to encode documents added after training and, by default, the collection token mean. The training-only output
+readout is not saved.
+
+Load the artifact in an embeddings configuration.
+
+```yaml
+embeddings:
+    path: colbert-ir/colbertv2.0
+    vectors:
+        lemur:
+            path: lemur-model
+```
+
+## Search behavior
+
+Loading a LEMUR artifact with a stored collection mean automatically centers both query and document token vectors with that mean.
+This keeps training and search on the same representation and makes fixed vectors independent of indexing or query batch size. An
+explicit `vectors.center` setting takes precedence. Older artifacts without a stored mean use the standard late-pooling fallback:
+batch centering for models with multiple linear layers and no centering for models with zero or one linear layer.
+
+LEMUR approximates standardized MaxSim targets, so useful ranking scores can be negative. txtai's dense vector path L2-normalizes the
+fixed vectors and removes results with scores less than or equal to zero. This can change ordering and return fewer than the requested
+number of candidates compared with raw maximum inner-product search.
+
+The Faiss backend uses exact `IDMap,Flat` search through 5,000 rows and switches to an IVF index above that threshold. In the measured
+scifact run, default IVF reduced LEMUR NDCG@10 by 43% relative to exact search, compared with 25% for MUVERA. For LEMUR corpora above
+5,000 rows, either pin `faiss.components` to an exact index or tune the IVF settings for the corpus. The exact configuration is:
+
+```yaml
+embeddings:
+    path: colbert-ir/colbertv2.0
+    vectors:
+        lemur:
+            path: lemur-model
+    faiss:
+        components: IDMap,Flat
+```
+
+## Methods
+
+Python documentation for the pipeline.
+
+### ::: txtai.pipeline.LemurTrainer.__call__

docs/pipeline/train/trainer.md

@@ -98,6 +98,20 @@ See the following PEFT documentation links for more information.
 - [Quantization](https://huggingface.co/docs/peft/developer_guides/quantization)
 - [LoRA](https://huggingface.co/docs/peft/developer_guides/lora)
 
+## Merge
+
+An important parameter for `language-generation` and `language-modeling` tasks is `merge` or the packing of data into chunks.
+
+It supports the following options.
+
+- `concat` (default) - text is split into chunks up to maxlength, data can be split across multiple chunks
+- `pack` - text is split into chunks up to maxlength, data guaranteed to be in same chunk, chunks can be smaller than maxlength
+- `None` - disables merging
+
+Merging helps reduce training time as data can be processed efficiently without padding. `concat` maximizes this as it guarantees each chunk will be up to maxlength size. `pack` is a middle ground where data is combined but records are preserved.
+
+For general language modeling tasks like masked language modeling, `concat` is the best choice. For instruction/prompt fine-tuning, `pack` or None are the better choices as it guarantees complex logic is not split across chunks.
+
 ## Methods 
 
 Python documentation for the pipeline.

docs/poweredby.md

@@ -7,8 +7,8 @@ The following applications are powered by txtai.
 | Application  | Description  |
 |:------------ |:-------------|
 | [rag](https://github.com/neuml/rag) | Retrieval Augmented Generation (RAG) application |
-| [ragdata](https://github.com/neuml/ragdata) | Build knowledge bases for RAG |
-| [paperai](https://github.com/neuml/paperai) | Semantic search and workflows for medical/scientific papers |
+| [ncoder](https://github.com/neuml/ncoder) | Open-Source AI coding agent |
+| [paperai](https://github.com/neuml/paperai) | AI for medical and scientific papers |
 | [annotateai](https://github.com/neuml/annotateai) | Automatically annotate papers with LLMs |
 
 In addition to this list, there are also many other [open-source projects](https://github.com/neuml/txtai/network/dependents), [published research](https://scholar.google.com/scholar?q=txtai&hl=en&as_ylo=2022) and closed proprietary/commercial projects that have built on txtai in production.

docs/usecases.md

@@ -1,6 +1,6 @@
 # Use Cases
 
-The following sections introduce common txtai use cases. A comprehensive set of over 60 [example notebooks and applications](../examples) are also available.
+The following sections introduce common txtai use cases. A comprehensive set of over 70 [example notebooks and applications](../examples) are also available.
 
 ## Semantic Search
 
@@ -42,14 +42,15 @@ Agents connect embeddings, pipelines, workflows and other agents together to aut
 
 ![agent](images/agent.png)
 
-txtai agents are built on top of the Transformers Agent framework. This supports all LLMs txtai supports (Hugging Face, llama.cpp, OpenAI / Claude / AWS Bedrock via LiteLLM).
+txtai agents are built on top of the [smolagents](https://github.com/huggingface/smolagents) framework. This supports all LLMs txtai supports (Hugging Face, llama.cpp, OpenAI / Claude / AWS Bedrock via LiteLLM). Agent prompting with [`agents.md`](https://github.com/agentsmd/agents.md) and [`skill.md`](https://agentskills.io/specification) are also supported.
 
-See the link below to learn more.
+Check out this [Agent Quickstart Example](https://github.com/neuml/txtai/blob/master/examples/agent_quickstart.py). Additional examples are listed below.
 
 | Notebook  | Description  |       |
 |:----------|:-------------|------:|
-| [Analyzing Hugging Face Posts with Graphs and Agents](https://github.com/neuml/txtai/blob/master/examples/68_Analyzing_Hugging_Face_Posts_with_Graphs_and_Agents.ipynb) | Explore a rich dataset with Graph Analysis and Agents | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/68_Analyzing_Hugging_Face_Posts_with_Graphs_and_Agents.ipynb) |
 | [Granting autonomy to agents](https://github.com/neuml/txtai/blob/master/examples/69_Granting_autonomy_to_agents.ipynb) | Agents that iteratively solve problems as they see fit | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/69_Granting_autonomy_to_agents.ipynb) |
+| [TxtAI got skills](https://github.com/neuml/txtai/blob/master/examples/83_TxtAI_got_skills.ipynb) | Integrate skill.md files with your agent | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/83_TxtAI_got_skills.ipynb) |
+| [Agent Tools](https://github.com/neuml/txtai/blob/master/examples/84_Agent_Tools.ipynb) [▶️](https://www.youtube.com/watch?v=RDNaFXQy3GQ) | Learn about the txtai agent toolkit | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/84_Agent_Tools.ipynb) |
 | [Analyzing LinkedIn Company Posts with Graphs and Agents](https://github.com/neuml/txtai/blob/master/examples/71_Analyzing_LinkedIn_Company_Posts_with_Graphs_and_Agents.ipynb) | Exploring how to improve social media engagement with AI | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/71_Analyzing_LinkedIn_Company_Posts_with_Graphs_and_Agents.ipynb) |
 
 ### Retrieval augmented generation
@@ -59,13 +60,13 @@ Retrieval augmented generation (RAG) reduces the risk of LLM hallucinations by c
 ![rag](images/rag.png#gh-light-mode-only)
 ![rag](images/rag-dark.png#gh-dark-mode-only)
 
-A novel feature of txtai is that it can provide both an answer and source citation.
+Check out this [RAG Quickstart Example](https://github.com/neuml/txtai/blob/master/examples/rag_quickstart.py). Additional examples are listed below.
 
 | Notebook  | Description  |       |
 |:----------|:-------------|------:|
-| [Build RAG pipelines with txtai](https://github.com/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) | Guide on retrieval augmented generation including how to create citations | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) |
-| [How RAG with txtai works](https://github.com/neuml/txtai/blob/master/examples/63_How_RAG_with_txtai_works.ipynb) | Create RAG processes, API services and Docker instances | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/63_How_RAG_with_txtai_works.ipynb) |
-| [Advanced RAG with graph path traversal](https://github.com/neuml/txtai/blob/master/examples/58_Advanced_RAG_with_graph_path_traversal.ipynb) | Graph path traversal to collect complex sets of data for advanced RAG | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/58_Advanced_RAG_with_graph_path_traversal.ipynb) |
+| [Build RAG pipelines with txtai](https://github.com/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) [▶️](https://www.youtube.com/watch?v=t_OeAc8NVfQ) | Guide on retrieval augmented generation including how to create citations | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) |
+| [RAG is more than Vector Search](https://github.com/neuml/txtai/blob/master/examples/79_RAG_is_more_than_Vector_Search.ipynb) | Context retrieval via Web, SQL and other sources | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/79_RAG_is_more_than_Vector_Search.ipynb) |
+| [GraphRAG with Wikipedia and GPT OSS](https://github.com/neuml/txtai/blob/master/examples/77_GraphRAG_with_Wikipedia_and_GPT_OSS.ipynb) | Deep graph search powered RAG | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/77_GraphRAG_with_Wikipedia_and_GPT_OSS.ipynb) |
 | [Speech to Speech RAG](https://github.com/neuml/txtai/blob/master/examples/65_Speech_to_Speech_RAG.ipynb) [▶️](https://www.youtube.com/watch?v=tH8QWwkVMKA) | Full cycle speech to speech workflow with RAG | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/65_Speech_to_Speech_RAG.ipynb) |
 
 ## Language Model Workflows
@@ -77,6 +78,8 @@ Language model workflows, also known as semantic workflows, connect language mod
 
 While LLMs are powerful, there are plenty of smaller, more specialized models that work better and faster for specific tasks. This includes models for extractive question-answering, automatic summarization, text-to-speech, transcription and translation.
 
+Check out this [Workflow Quickstart Example](https://github.com/neuml/txtai/blob/master/examples/workflow_quickstart.py). Additional examples are listed below.
+
 | Notebook  | Description  |       |
 |:----------|:-------------|------:|
 | [Run pipeline workflows](https://github.com/neuml/txtai/blob/master/examples/14_Run_pipeline_workflows.ipynb) [▶️](https://www.youtube.com/watch?v=UBMPDCn1gEU) | Simple yet powerful constructs to efficiently process data | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/neuml/txtai/blob/master/examples/14_Run_pipeline_workflows.ipynb) |

docs/workflow/index.md

@@ -133,7 +133,7 @@ Workflows can connect multiple LLM prompting tasks together.
 
 ```yaml
 llm:
-  path: google/flan-t5-xl
+  path: openai/gpt-oss-20b
 
 workflow:
   llm:
@@ -168,7 +168,7 @@ Any txtai pipeline/workflow task can be connected in workflows with LLMs.
 
 ```yaml
 llm:
-  path: google/flan-t5-xl
+  path: openai/gpt-oss-20b
 
 translation:
 
@@ -202,4 +202,4 @@ Workflows are callable objects. Workflows take an input of iterable data element
 
 ## More examples
 
-See [this link](../examples/#workflows) for a full list of workflow examples.
+Check out this [Workflow Quickstart Example](https://github.com/neuml/txtai/blob/master/examples/workflow_quickstart.py). See [this link](../examples/#workflows) for a full list of workflow examples.

docs/workflow/task/console.md

@@ -22,8 +22,9 @@ This task can also be created with workflow configuration.
 
 ```yaml
 workflow:
-  tasks:
-    - task: console
+  name:
+    tasks:
+      - task: console
 ```
 
 ## Methods

docs/workflow/task/export.md

@@ -22,8 +22,9 @@ This task can also be created with workflow configuration.
 
 ```yaml
 workflow:
-  tasks:
-    - task: export
+  name:
+    tasks:
+      - task: export
 ```
 
 ## Methods

docs/workflow/task/file.md

@@ -22,8 +22,9 @@ This task can also be created with workflow configuration.
 
 ```yaml
 workflow:
-  tasks:
-    - task: file
+  name:
+    tasks:
+      - task: file
 ```
 
 ## Methods

docs/workflow/task/image.md

@@ -22,8 +22,9 @@ This task can also be created with workflow configuration.
 
 ```yaml
 workflow:
-  tasks:
-    - task: image
+  name:
+    tasks:
+      - task: image
 ```
 
 ## Methods

docs/workflow/task/index.md

@@ -35,12 +35,26 @@ Tasks can also be created with configuration as part of a workflow.
 
 ```yaml
 workflow:
-  tasks:
-    - action: summary 
+  name:
+    tasks:
+      - action: summary
 ```
 
 ::: txtai.workflow.Task.__init__
 
+## Multi-action task inputs
+
+A multi-action `Task` requires reusable inputs. Passing a single-pass iterator, such as a generator or the output of a `StreamTask`, raises a `ValueError` before consuming inputs or running any actions. Automatically buffering an iterator could exhaust memory, so tasks do not convert these inputs to a list.
+
+Iterators passed to a `Workflow` remain supported: the workflow collects bounded batches before passing them to its tasks. For example:
+
+```python
+workflow = Workflow([Task([list, list])], batch=2)
+list(workflow(iter([1, 2, 3])))  # [(1, 1), (2, 2), (3, 3)]
+```
+
+This batching occurs at the workflow input; it does not batch iterator outputs produced by tasks inside that workflow. Single-action tasks continue to accept iterator inputs.
+
 ## Multi-action task concurrency
 
 The default processing mode is to run actions sequentially. Multiprocessing support is already built in at a number of levels. Any of the GPU models will maximize GPU utilization for example and even in CPU mode, concurrency is utilized. But there are still use cases for task action concurrency. For example, if the system has multiple GPUs, the task runs external sequential code, or the task has a large number of I/O tasks.
@@ -72,7 +86,7 @@ workflow = Workflow([Task(lambda x: [y * 3 for y in x], unpack=False, column=0)]
 list(workflow([(2, 8)]))
 ```
 
-For the example input tuple of (2, 2), the workflow will only select the first element (2) and run the task against that element. 
+For the example input tuple of (2, 8), the workflow will only select the first element (2) and run the task against that element.
 
 ```python
 workflow = Workflow([Task([lambda x: [y * 3 for y in x], 

docs/workflow/task/retrieve.md

@@ -22,9 +22,10 @@ This task can also be created with workflow configuration.
 
 ```yaml
 workflow:
-  tasks:
-    - task: retrieve
-      directory: /tmp
+  name:
+    tasks:
+      - task: retrieve
+        directory: /tmp
 ```
 
 ## Methods

docs/workflow/task/service.md

@@ -22,9 +22,10 @@ This task can also be created with workflow configuration.
 
 ```yaml
 workflow:
-  tasks:
-    - task: service
-      url: https://service.url/action
+  name:
+    tasks:
+      - task: service
+        url: https://service.url/action
 ```
 
 ## Methods

docs/workflow/task/storage.md

@@ -22,8 +22,9 @@ This task can also be created with workflow configuration.
 
 ```yaml
 workflow:
-  tasks:
-    - task: storage
+  name:
+    tasks:
+      - task: storage
 ```
 
 ## Methods

docs/workflow/task/template.md

@@ -23,9 +23,10 @@ This task can also be created with workflow configuration.
 
 ```yaml
 workflow:
-  tasks:
-    - task: template
-      template: This is a {text} task
+  name:
+    tasks:
+      - task: template
+        template: This is a {text} task
 ```
 
 ## Methods

docs/workflow/task/url.md

@@ -22,8 +22,9 @@ This task can also be created with workflow configuration.
 
 ```yaml
 workflow:
-  tasks:
-    - task: url
+  name:
+    tasks:
+      - task: url
 ```
 
 ## Methods

examples/33_Query_translation.ipynb

@@ -1,28 +1,4 @@
 {
-  "nbformat": 4,
-  "nbformat_minor": 0,
-  "metadata": {
-    "kernelspec": {
-      "name": "python3",
-      "display_name": "Python 3",
-      "language": "python"
-    },
-    "language_info": {
-      "name": "python",
-      "version": "3.7.6",
-      "mimetype": "text/x-python",
-      "codemirror_mode": {
-        "name": "ipython",
-        "version": 3
-      },
-      "pygments_lexer": "ipython3",
-      "nbconvert_exporter": "python",
-      "file_extension": ".py"
-    },
-    "colab": {
-      "provenance": []
-    }
-  },
   "cells": [
     {
       "cell_type": "markdown",
@@ -61,44 +37,59 @@
     },
     {
       "cell_type": "code",
+      "execution_count": null,
       "metadata": {
-        "_uuid": "8f2839f25d086af736a60e9eeb907d3b93b6e0e5",
         "_cell_guid": "b1076dfc-b9ad-4769-8c92-a6c4dae69d19",
-        "trusted": true,
         "_kg_hide-output": true,
-        "id": "24q-1n5i6XzQ"
+        "_uuid": "8f2839f25d086af736a60e9eeb907d3b93b6e0e5",
+        "id": "24q-1n5i6XzQ",
+        "trusted": true
       },
+      "outputs": [],
       "source": [
         "%%capture\n",
         "!pip install git+https://github.com/neuml/txtai#egg=txtai[pipeline]"
-      ],
-      "execution_count": null,
-      "outputs": []
+      ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "0p3WCDniUths"
+      },
       "source": [
         "# Create index\n",
         "Let's first recap how to create an index. We'll use the classic txtai example.\n"
-      ],
-      "metadata": {
-        "id": "0p3WCDniUths"
-      }
+      ]
     },
     {
       "cell_type": "code",
+      "execution_count": 1,
       "metadata": {
-        "_uuid": "d629ff2d2480ee46fbb7e2d37f6b5fab8052498a",
         "_cell_guid": "79c7e3d0-c299-4dcb-8224-4455121ee9b0",
-        "trusted": true,
-        "id": "2j_CFGDR6Xzp",
+        "_uuid": "d629ff2d2480ee46fbb7e2d37f6b5fab8052498a",
         "colab": {
           "base_uri": "https://localhost:8080/"
         },
-        "outputId": "e65238c5-d67f-4fe4-9cbf-c88b0ff3a8bb"
+        "id": "2j_CFGDR6Xzp",
+        "outputId": "e65238c5-d67f-4fe4-9cbf-c88b0ff3a8bb",
+        "trusted": true
       },
+      "outputs": [
+        {
+          "data": {
+            "text/plain": [
+              "[{'id': '4',\n",
+              "  'text': 'Maine man wins $1M from $25 lottery ticket',\n",
+              "  'score': 0.08329025655984879}]"
+            ]
+          },
+          "execution_count": 1,
+          "metadata": {},
+          "output_type": "execute_result"
+        }
+      ],
       "source": [
-        "from txtai.embeddings import Embeddings\n",
+        "from txtai import Embeddings\n",
         "\n",
         "data = [\"US tops 5 million confirmed virus cases\",\n",
         "        \"Canada's last fully intact ice shelf has suddenly collapsed, forming a Manhattan-sized iceberg\",\n",
@@ -108,65 +99,29 @@
         "        \"Make huge profits without work, earn up to $100,000 a day\"]\n",
         "\n",
         "# Create embeddings index with content enabled. The default behavior is to only store indexed vectors.\n",
-        "embeddings = Embeddings({\"path\": \"sentence-transformers/nli-mpnet-base-v2\", \"content\": True})\n",
+        "embeddings = Embeddings(path=\"sentence-transformers/nli-mpnet-base-v2\", content=True)\n",
         "\n",
         "# Create an index for the list of text\n",
-        "embeddings.index([(uid, text, None) for uid, text in enumerate(data)])\n",
+        "embeddings.index(data)\n",
         "\n",
         "# Run a search\n",
         "embeddings.search(\"feel good story\", 1)"
-      ],
-      "execution_count": null,
-      "outputs": [
-        {
-          "output_type": "execute_result",
-          "data": {
-            "text/plain": [
-              "[{'id': '4',\n",
-              "  'score': 0.08329011499881744,\n",
-              "  'text': 'Maine man wins $1M from $25 lottery ticket'}]"
-            ]
-          },
-          "metadata": {},
-          "execution_count": 30
-        }
       ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "QTee7YMNDD4R"
+      },
       "source": [
         "# Query translation models\n",
         "\n",
         "Next we'll explore how query translation models work with examples. "
-      ],
-      "metadata": {
-        "id": "QTee7YMNDD4R"
-      }
+      ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "from txtai.pipeline import Sequences\n",
-        "\n",
-        "sequences = Sequences(\"NeuML/t5-small-txtsql\")\n",
-        "\n",
-        "queries = [\n",
-        "  \"feel good story\",\n",
-        "  \"feel good story since yesterday\",\n",
-        "  \"feel good story with lottery in text\",\n",
-        "  \"how many feel good story\",\n",
-        "  \"feel good story translated to fr\",\n",
-        "  \"feel good story summarized\"\n",
-        "]\n",
-        "\n",
-        "# Prefix to pass to T5 model\n",
-        "prefix = \"translate English to SQL: \"\n",
-        "\n",
-        "for query in queries:\n",
-        "  print(f\"Input: {query}\")\n",
-        "  print(f\"SQL: {sequences(query, prefix)}\")\n",
-        "  print()\n"
-      ],
+      "execution_count": 2,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/"
@@ -174,11 +129,10 @@
         "id": "04iOgP_ojSfK",
         "outputId": "41e73130-75e6-4ae9-b690-685972a13565"
       },
-      "execution_count": null,
       "outputs": [
         {
-          "output_type": "stream",
           "name": "stdout",
+          "output_type": "stream",
           "text": [
             "Input: feel good story\n",
             "SQL: select id, text, score from txtai where similar('feel good story')\n",
@@ -186,8 +140,8 @@
             "Input: feel good story since yesterday\n",
             "SQL: select id, text, score from txtai where similar('feel good story') and entry >= date('now', '-1 day')\n",
             "\n",
-            "Input: feel good story with lottery in text\n",
-            "SQL: select id, text, score from txtai where similar('feel good story') and text like '% lottery%'\n",
+            "Input: feel good story with 'lottery' in text\n",
+            "SQL: select id, text, score from txtai where similar('feel good story') and text like '%lottery%'\n",
             "\n",
             "Input: how many feel good story\n",
             "SQL: select count(*) from txtai where similar('feel good story')\n",
@@ -200,32 +154,73 @@
             "\n"
           ]
         }
+      ],
+      "source": [
+        "from txtai import LLM\n",
+        "\n",
+        "llm = LLM(\"NeuML/t5-small-txtsql\", template=\"translate English to SQL: {text}\")\n",
+        "\n",
+        "queries = [\n",
+        "  \"feel good story\",\n",
+        "  \"feel good story since yesterday\",\n",
+        "  \"feel good story with 'lottery' in text\",\n",
+        "  \"how many feel good story\",\n",
+        "  \"feel good story translated to fr\",\n",
+        "  \"feel good story summarized\"\n",
+        "]\n",
+        "\n",
+        "for query in queries:\n",
+        "  print(f\"Input: {query}\")\n",
+        "  print(f\"SQL: {llm(query)}\")\n",
+        "  print()\n"
       ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "rAnEMaiWlOXm"
+      },
       "source": [
         "Looking at the query translations above gives an idea on how this model works.\n",
         "\n",
         "[t5-small-txtsql](https://huggingface.co/NeuML/t5-small-txtsql) is the default model. Custom domain query syntax languages can be created using this same methodology, including for other languages. Natural language can be translated to functions, query clauses, column selection and more!"
-      ],
-      "metadata": {
-        "id": "rAnEMaiWlOXm"
-      }
+      ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "P9hOcgNfjSyL"
+      },
       "source": [
         "# Natural language filtering\n",
         "\n",
         "Now it's time for this in action! Let's first initialize the embeddings index with the appropriate settings."
-      ],
-      "metadata": {
-        "id": "P9hOcgNfjSyL"
-      }
+      ]
     },
     {
       "cell_type": "code",
+      "execution_count": 2,
+      "metadata": {
+        "colab": {
+          "base_uri": "https://localhost:8080/"
+        },
+        "id": "L2DDJrd0RAaN",
+        "outputId": "c60bfc29-187b-4926-8656-04063b2ca85c"
+      },
+      "outputs": [
+        {
+          "data": {
+            "text/plain": [
+              "{'id': '4',\n",
+              " 'score': 0.08329025655984879,\n",
+              " 'text': 'Maine Mann gewinnt $1M von $25 Lotterie-Ticket'}"
+            ]
+          },
+          "execution_count": 2,
+          "metadata": {},
+          "output_type": "execute_result"
+        }
+      ],
       "source": [
         "from txtai.pipeline import Translation\n",
         "\n",
@@ -235,56 +230,34 @@
         "translation = Translation()\n",
         "\n",
         "# Create embeddings index with content enabled. The default behavior is to only store indexed vectors.\n",
-        "embeddings = Embeddings({\"path\": \"sentence-transformers/nli-mpnet-base-v2\",\n",
-        "                         \"content\": True,\n",
-        "                         \"query\": {\"path\": \"NeuML/t5-small-txtsql\"},\n",
-        "                         \"functions\": [translate]})\n",
+        "embeddings = Embeddings(\n",
+        "  path=\"sentence-transformers/nli-mpnet-base-v2\",\n",
+        "  content=True,\n",
+        "  query={\"path\": \"NeuML/t5-small-txtsql\"},\n",
+        "  functions=[translate]\n",
+        ")\n",
         "\n",
         "# Create an index for the list of text\n",
-        "embeddings.index([(uid, text, None) for uid, text in enumerate(data)])\n",
+        "embeddings.index(data)\n",
         "\n",
         "query = \"select id, score, translate(text, 'de') 'text' from txtai where similar('feel good story')\"\n",
         "\n",
         "# Run a search using a custom SQL function\n",
         "embeddings.search(query)[0]"
-      ],
-      "metadata": {
-        "colab": {
-          "base_uri": "https://localhost:8080/"
-        },
-        "id": "L2DDJrd0RAaN",
-        "outputId": "c60bfc29-187b-4926-8656-04063b2ca85c"
-      },
-      "execution_count": null,
-      "outputs": [
-        {
-          "output_type": "execute_result",
-          "data": {
-            "text/plain": [
-              "{'id': '4',\n",
-              " 'score': 0.08329011499881744,\n",
-              " 'text': 'Maine Mann gewinnt $1M von $25 Lotterie-Ticket'}"
-            ]
-          },
-          "metadata": {},
-          "execution_count": 32
-        }
       ]
     },
     {
       "cell_type": "markdown",
-      "source": [
-        "Note how the query model was provided as a embeddings index configuration parameter. Custom SQL functions were also added in. Let's now see if the same SQL statement can be run with a natural language query."
-      ],
       "metadata": {
         "id": "MNd7QmFmnh-f"
-      }
+      },
+      "source": [
+        "Note how the query model was provided as a embeddings index configuration parameter. Custom SQL functions were also added in. Let's now see if the same SQL statement can be run with a natural language query."
+      ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "embeddings.search(\"feel good story translated to de\")[0]"
-      ],
+      "execution_count": 4,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/"
@@ -292,36 +265,36 @@
         "id": "mnGSBNMonur2",
         "outputId": "dc14d898-b46f-42e6-88e0-434863cc15d6"
       },
-      "execution_count": null,
       "outputs": [
         {
-          "output_type": "execute_result",
           "data": {
             "text/plain": [
               "{'id': '4',\n",
-              " 'score': 0.08329011499881744,\n",
-              " 'text': 'Maine Mann gewinnt $1M von $25 Lotterie-Ticket'}"
+              " 'text': 'Maine Mann gewinnt $1M von $25 Lotterie-Ticket',\n",
+              " 'score': 0.08329025655984879}"
             ]
           },
+          "execution_count": 4,
           "metadata": {},
-          "execution_count": 33
+          "output_type": "execute_result"
         }
+      ],
+      "source": [
+        "embeddings.search(\"feel good story translated to de\")[0]"
       ]
     },
     {
       "cell_type": "markdown",
-      "source": [
-        "Same result. Let's try a few more."
-      ],
       "metadata": {
         "id": "OnmVUK2DoZkh"
-      }
+      },
+      "source": [
+        "Same result. Let's try a few more."
+      ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "embeddings.search(\"feel good story since yesterday\")[0]"
-      ],
+      "execution_count": 5,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/"
@@ -329,27 +302,27 @@
         "id": "Js2b_M61oitg",
         "outputId": "fab080cb-d082-4cf3-8314-12d23acc3c02"
       },
-      "execution_count": null,
       "outputs": [
         {
-          "output_type": "execute_result",
           "data": {
             "text/plain": [
               "{'id': '4',\n",
-              " 'score': 0.08329011499881744,\n",
-              " 'text': 'Maine man wins $1M from $25 lottery ticket'}"
+              " 'text': 'Maine man wins $1M from $25 lottery ticket',\n",
+              " 'score': 0.08329025655984879}"
             ]
           },
+          "execution_count": 5,
           "metadata": {},
-          "execution_count": 34
+          "output_type": "execute_result"
         }
+      ],
+      "source": [
+        "embeddings.search(\"feel good story since yesterday\")[0]"
       ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "embeddings.search(\"feel good story with lottery in text\")[0]"
-      ],
+      "execution_count": 6,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/"
@@ -357,36 +330,36 @@
         "id": "pcxPgriwomWv",
         "outputId": "7e71dd09-7c4f-4147-ba9e-47a45c1cf90f"
       },
-      "execution_count": null,
       "outputs": [
         {
-          "output_type": "execute_result",
           "data": {
             "text/plain": [
-              "{'id': '4',\n",
-              " 'score': 0.08329011499881744,\n",
-              " 'text': 'Maine man wins $1M from $25 lottery ticket'}"
+              "[{'id': '4',\n",
+              "  'text': 'Maine man wins $1M from $25 lottery ticket',\n",
+              "  'score': 0.08329025655984879}]"
             ]
           },
+          "execution_count": 6,
           "metadata": {},
-          "execution_count": 35
+          "output_type": "execute_result"
         }
+      ],
+      "source": [
+        "embeddings.search(\"feel good story with 'lottery' in text\")"
       ]
     },
     {
       "cell_type": "markdown",
-      "source": [
-        "For good measure, a couple queries with filters that return no results."
-      ],
       "metadata": {
         "id": "1ob4Q_CLpGBm"
-      }
+      },
+      "source": [
+        "For good measure, a couple queries with filters that return no results."
+      ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "embeddings.search(\"feel good story with missing in text\")"
-      ],
+      "execution_count": 7,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/"
@@ -394,25 +367,25 @@
         "id": "LMaf7JJzonAC",
         "outputId": "71080bd1-2d9e-41ba-9501-2cdedaf26d6d"
       },
-      "execution_count": null,
       "outputs": [
         {
-          "output_type": "execute_result",
           "data": {
             "text/plain": [
               "[]"
             ]
           },
+          "execution_count": 7,
           "metadata": {},
-          "execution_count": 36
+          "output_type": "execute_result"
         }
+      ],
+      "source": [
+        "embeddings.search(\"feel good story with missing in text\")"
       ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "embeddings.search(\"feel good story with field equal 14\")"
-      ],
+      "execution_count": 8,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/"
@@ -420,33 +393,57 @@
         "id": "ADFNLhvto_0J",
         "outputId": "12e08ea8-7203-4fb0-cb84-1133b5db9dc0"
       },
-      "execution_count": null,
       "outputs": [
         {
-          "output_type": "execute_result",
           "data": {
             "text/plain": [
               "[]"
             ]
           },
+          "execution_count": 8,
           "metadata": {},
-          "execution_count": 37
+          "output_type": "execute_result"
         }
+      ],
+      "source": [
+        "embeddings.search(\"feel good story with field equal 14\")"
       ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "mTT8nopiRdVH"
+      },
       "source": [
         "# Query translation with applications\n",
         "\n",
         "Of course this is all available with YAML-configured applications."
-      ],
-      "metadata": {
-        "id": "mTT8nopiRdVH"
-      }
+      ]
     },
     {
       "cell_type": "code",
+      "execution_count": 9,
+      "metadata": {
+        "colab": {
+          "base_uri": "https://localhost:8080/"
+        },
+        "id": "FZ_7G6M4RUbz",
+        "outputId": "b00ce1c2-8df2-4289-8d03-991174a0e74f"
+      },
+      "outputs": [
+        {
+          "data": {
+            "text/plain": [
+              "{'id': '4',\n",
+              " 'text': 'Maine Mann gewinnt $1M von $25 Lotterie-Ticket',\n",
+              " 'score': 0.08329025655984879}"
+            ]
+          },
+          "execution_count": 9,
+          "metadata": {},
+          "output_type": "execute_result"
+        }
+      ],
       "source": [
         "config = \"\"\"\n",
         "translation:\n",
@@ -470,28 +467,6 @@
         "\n",
         "# Run search query\n",
         "app.search(\"feel good story translated to de\")[0]"
-      ],
-      "metadata": {
-        "colab": {
-          "base_uri": "https://localhost:8080/"
-        },
-        "id": "FZ_7G6M4RUbz",
-        "outputId": "b00ce1c2-8df2-4289-8d03-991174a0e74f"
-      },
-      "execution_count": null,
-      "outputs": [
-        {
-          "output_type": "execute_result",
-          "data": {
-            "text/plain": [
-              "{'id': '4',\n",
-              " 'score': 0.08329011499881744,\n",
-              " 'text': 'Maine Mann gewinnt $1M von $25 Lotterie-Ticket'}"
-            ]
-          },
-          "metadata": {},
-          "execution_count": 38
-        }
       ]
     },
     {
@@ -505,5 +480,29 @@
         "This notebook introduced natural language filtering with query translation models. This powerful feature adds filtering and pipelines to natural language statements. Custom domain-specific query languages can be created to enable rich queries natively in txtai."
       ]
     }
-  ]
-}
+  ],
+  "metadata": {
+    "colab": {
+      "provenance": []
+    },
+    "kernelspec": {
+      "display_name": "local",
+      "language": "python",
+      "name": "python3"
+    },
+    "language_info": {
+      "codemirror_mode": {
+        "name": "ipython",
+        "version": 3
+      },
+      "file_extension": ".py",
+      "mimetype": "text/x-python",
+      "name": "python",
+      "nbconvert_exporter": "python",
+      "pygments_lexer": "ipython3",
+      "version": "3.10.19"
+    }
+  },
+  "nbformat": 4,
+  "nbformat_minor": 0
+}

examples/41_Train_a_language_model_from_scratch.ipynb

@@ -30,6 +30,7 @@
         "- Changing the vocabulary size\n",
         "- Changing the number of hidden dimensions\n",
         "- Changing the number of attention heads or layers\n",
+        "- Create a custom model architecture\n",
         "\n",
         "This notebook will show how to build a new tokenizer and train a small language model (known as a micromodel) from scratch.\n"
       ],

examples/42_Prompt_driven_search_with_LLMs.ipynb

@@ -1,17 +1,4 @@
 {
-  "nbformat": 4,
-  "nbformat_minor": 0,
-  "metadata": {
-    "colab": {
-      "provenance": []
-    },
-    "kernelspec": {
-      "name": "python3",
-      "display_name": "Python 3"
-    },
-    "accelerator": "GPU",
-    "gpuClass": "standard"
-  },
   "cells": [
     {
       "cell_type": "markdown",
@@ -44,15 +31,15 @@
     },
     {
       "cell_type": "code",
+      "execution_count": null,
       "metadata": {
         "id": "LPQTb25tASIG"
       },
+      "outputs": [],
       "source": [
         "%%capture\n",
         "!pip install git+https://github.com/neuml/txtai datasets"
-      ],
-      "execution_count": null,
-      "outputs": []
+      ]
     },
     {
       "cell_type": "markdown",
@@ -71,32 +58,65 @@
     },
     {
       "cell_type": "code",
+      "execution_count": 1,
       "metadata": {
         "id": "OUc9gqTyAYnm"
       },
+      "outputs": [],
       "source": [
         "%%capture\n",
         "\n",
         "from txtai import Embeddings, RAG\n",
         "\n",
         "# Create embeddings model with content support\n",
-        "embeddings = Embeddings({\"path\": \"sentence-transformers/all-MiniLM-L6-v2\", \"content\": True})\n",
+        "embeddings = Embeddings(path=\"sentence-transformers/all-MiniLM-L6-v2\", content=True)\n",
         "\n",
-        "# Create RAG instance\n",
-        "rag = RAG(embeddings, \"google/flan-t5-base\")"
-      ],
-      "execution_count": null,
-      "outputs": []
+        "# Create the RAG pipeline\n",
+        "rag = RAG(embeddings, \"Qwen/Qwen3-4B-Instruct-2507\", template=\"\"\"\n",
+        "  Answer the following question using the provided context.\n",
+        "\n",
+        "  Question:\n",
+        "  {question}\n",
+        "\n",
+        "  Context:\n",
+        "  {context}\n",
+        "\"\"\")"
+      ]
     },
     {
       "cell_type": "code",
+      "execution_count": 2,
       "metadata": {
-        "id": "4X5z3UjnAGe7",
         "colab": {
           "base_uri": "https://localhost:8080/"
         },
+        "id": "4X5z3UjnAGe7",
         "outputId": "cacb7e9d-471a-437d-c68d-a8b51f876413"
       },
+      "outputs": [
+        {
+          "name": "stdout",
+          "output_type": "stream",
+          "text": [
+            "---- Red Sox - Blue Jays ----\n",
+            "{'answer': 'The Blue Jays won the game.'}\n",
+            "{'answer': 'The score was 2-1 in favor of the Blue Jays.'}\n",
+            "\n",
+            "---- Phillies - Braves ----\n",
+            "{'answer': 'The Phillies won the game.'}\n",
+            "{'answer': 'The score was 5-0 in favor of the Phillies.'}\n",
+            "\n",
+            "---- Dodgers - Giants ----\n",
+            "{'answer': 'The Giants won the game.'}\n",
+            "{'answer': 'The score was Giants 5, Dodgers 4.'}\n",
+            "\n",
+            "---- Flyers - Lightning ----\n",
+            "{'answer': 'The Flyers won the game.'}\n",
+            "{'answer': 'The score was Flyers 4, Lightning 1.'}\n",
+            "\n"
+          ]
+        }
+      ],
       "source": [
         "data = [\"Giants hit 3 HRs to down Dodgers\",\n",
         "        \"Giants 5 Dodgers 4 final\",\n",
@@ -111,88 +131,40 @@
         "        \"Flyers 4 Lightning 1 final\",\n",
         "        \"Flyers win 4-1\"]\n",
         "\n",
-        "def prompt(question):\n",
-        "  return f\"\"\"\n",
-        "    Answer the following question using the context below.\n",
-        "    Question: {question}\n",
-        "    Context:\n",
-        "  \"\"\"\n",
-        "\n",
         "questions = [\"What team won the game?\", \"What was score?\"]\n",
         "\n",
-        "execute = lambda query: rag([(question, query, prompt(question), False) for question in questions], data)\n",
-        "\n",
         "for query in [\"Red Sox - Blue Jays\", \"Phillies - Braves\", \"Dodgers - Giants\", \"Flyers - Lightning\"]:\n",
         "    print(\"----\", query, \"----\")\n",
-        "    for answer in execute(query):\n",
+        "    for answer in rag([f\"{query} {x}\" for x in questions], data):\n",
         "        print(answer)\n",
         "    print()"
-      ],
-      "execution_count": null,
-      "outputs": [
-        {
-          "output_type": "stream",
-          "name": "stdout",
-          "text": [
-            "---- Red Sox - Blue Jays ----\n",
-            "('What team won the game?', 'Blue Jays')\n",
-            "('What was score?', '2-1')\n",
-            "\n",
-            "---- Phillies - Braves ----\n",
-            "('What team won the game?', 'Phillies')\n",
-            "('What was score?', '5-0')\n",
-            "\n",
-            "---- Dodgers - Giants ----\n",
-            "('What team won the game?', 'Giants')\n",
-            "('What was score?', '5-4')\n",
-            "\n",
-            "---- Flyers - Lightning ----\n",
-            "('What team won the game?', 'Flyers')\n",
-            "('What was score?', '4-1')\n",
-            "\n"
-          ]
-        }
       ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "7AnPvSeM3N1Z"
+      },
       "source": [
         "This code runs a series of questions. First it runs an embeddings filtering query to find the most relevant text. For example, `Red Sox - Blue Jays` finds text related to those teams. Then `What team won the game?` and `What was the score?` are asked.\n",
         "\n",
         "This logic is the same logic found in Notebook 5 - Extractive QA with txtai but uses prompt-based QA vs extractive QA. "
-      ],
-      "metadata": {
-        "id": "7AnPvSeM3N1Z"
-      }
+      ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "Aj8GoDk331cS"
+      },
       "source": [
         "# Embeddings-guided and Prompt-driven Search\n",
         "\n",
         "Now for the fun stuff. Let's build an embeddings index for the `ag_news` dataset (a set of news stories from the mid 2000s). Then we'll use prompts to ask questions with embeddings results as the context."
-      ],
-      "metadata": {
-        "id": "Aj8GoDk331cS"
-      }
+      ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "from datasets import load_dataset\n",
-        "\n",
-        "dataset = load_dataset(\"ag_news\", split=\"train\")\n",
-        "\n",
-        "# List of all text elements\n",
-        "texts = dataset[\"text\"]\n",
-        "\n",
-        "# Create an embeddings index over the dataset\n",
-        "embeddings = Embeddings({\"path\": \"sentence-transformers/all-MiniLM-L6-v2\", \"content\": True})\n",
-        "embeddings.index((x, text, None) for x, text in enumerate(texts))\n",
-        "\n",
-        "# Create RAG instance\n",
-        "rag = RAG(embeddings, \"google/flan-t5-large\")"
-      ],
+      "execution_count": null,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/"
@@ -200,40 +172,41 @@
         "id": "yL716oEZ43t-",
         "outputId": "23f4b0e7-a60a-4e89-fb57-06966e6612f8"
       },
-      "execution_count": null,
-      "outputs": []
+      "outputs": [
+      ],
+      "source": [
+        "from datasets import load_dataset\n",
+        "\n",
+        "dataset = load_dataset(\"ag_news\", split=\"train\")\n",
+        "\n",
+        "# Create an embeddings index over the dataset\n",
+        "embeddings = Embeddings(path=\"sentence-transformers/all-MiniLM-L6-v2\", content=True)\n",
+        "embeddings.index(dataset[\"text\"])\n",
+        "\n",
+        "# Create RAG instance\n",
+        "rag = RAG(embeddings, \"Qwen/Qwen3-4B-Instruct-2507\", template=\"\"\"\n",
+        "  Answer the following question using the provided context.\n",
+        "\n",
+        "  Question:\n",
+        "  {question}\n",
+        "\n",
+        "  Context:\n",
+        "  {context}\n",
+        "\"\"\", output=\"flatten\")"
+      ]
     },
     {
       "cell_type": "markdown",
-      "source": [
-        "Now let's run a prompt-driven search!"
-      ],
       "metadata": {
         "id": "Ifl8JwLDBL7k"
-      }
+      },
+      "source": [
+        "Now let's run a prompt-driven search!"
+      ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "def prompt(question):\n",
-        "  return f\"\"\"Answer the following question using only the context below. Say 'no answer' when the question can't be answered.\n",
-        "Question: {question}\n",
-        "Context: \"\"\"\n",
-        "\n",
-        "def search(query, question=None):\n",
-        "  # Default question to query if empty\n",
-        "  if not question:\n",
-        "    question = query\n",
-        "\n",
-        "  return rag([(\"answer\", query, prompt(question), False)])[0][1]\n",
-        "\n",
-        "question = \"Who won the 2004 presidential election?\"\n",
-        "answer = search(question)\n",
-        "print(question, answer)\n",
-        "\n",
-        "nquestion = \"Who did the candidate beat?\"\n",
-        "print(nquestion, search(f\"{question} {answer}. {nquestion}\"))"
-      ],
+      "execution_count": 22,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/"
@@ -241,91 +214,52 @@
         "id": "5O1WBJ8153Mo",
         "outputId": "ddf09da2-7d4c-4fd3-b0da-df5631bacd13"
       },
-      "execution_count": null,
       "outputs": [
         {
-          "output_type": "stream",
           "name": "stdout",
+          "output_type": "stream",
           "text": [
-            "Who won the 2004 presidential election? George W. Bush\n",
-            "Who did the candidate beat? John F. Kerry\n"
+            "Who won the 2004 presidential election? George W. Bush won the 2004 presidential election.\n",
+            "Who did the candidate beat? George W. Bush beat John F. Kerry in the 2004 presidential election.\n"
           ]
         }
+      ],
+      "source": [
+        "question = \"Who won the 2004 presidential election?\"\n",
+        "answer = rag(question)\n",
+        "print(question, answer)\n",
+        "\n",
+        "nquestion = \"Who did the candidate beat?\"\n",
+        "print(nquestion, rag(f\"{question} {answer}. {nquestion}\"))"
       ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "AhViFXH_BZSo"
+      },
       "source": [
         "And there are the answers. Let's unpack how this works.\n",
         "\n",
         "The first thing the RAG pipeline does is run an embeddings search to find the most relevant text within the index. A context string is then built using those search results.\n",
         "\n",
-        "After that, a prompt is generated, run and the answer printed. Let's see what a full prompt looks like."
-      ],
-      "metadata": {
-        "id": "AhViFXH_BZSo"
-      }
-    },
-    {
-      "cell_type": "code",
-      "source": [
-        "text = prompt(question)\n",
-        "text += \"\\n\" + \"\\n\".join(x[\"text\"]for x in embeddings.search(question))\n",
-        "print(text)"
-      ],
-      "metadata": {
-        "colab": {
-          "base_uri": "https://localhost:8080/"
-        },
-        "id": "h7BsoJC3CFOq",
-        "outputId": "44ae0dec-4f79-4d44-88f5-585e5eae9cb2"
-      },
-      "execution_count": null,
-      "outputs": [
-        {
-          "output_type": "stream",
-          "name": "stdout",
-          "text": [
-            "Answer the following question using only the context below. Say 'no answer' when the question can't be answered.\n",
-            "Question: Who won the 2004 presidential election?\n",
-            "Context: \n",
-            "Right- and left-click politics The 2004 presidential race ended last week in a stunning defeat for Massachusetts Senator John F. Kerry, as incumbent President George W. Bush cruised to an easy victory.\n",
-            "2004 Presidential Endorsements (AP) AP - Newspaper endorsements in the 2004 presidential campaign between President Bush, a Republican, and Sen. John Kerry, a Democrat.\n",
-            "Presidential Campaign to Nov. 2, 2004 (Reuters) Reuters - The following diary of events\\leading up to the presidential election on Nov. 2.\n"
-          ]
-        }
+        "After that, a prompt is generated, run and the answer printed."
       ]
     },
     {
       "cell_type": "markdown",
-      "source": [
-        "The prompt has the information needed to determine the answers to the questions."
-      ],
       "metadata": {
-        "id": "4m9XketCty3m"
-      }
-    },
-    {
-      "cell_type": "markdown",
+        "id": "JtDcVPdOB0Rv"
+      },
       "source": [
         "# Additional examples\n",
         "\n",
         "Before moving on, a couple more example questions."
-      ],
-      "metadata": {
-        "id": "JtDcVPdOB0Rv"
-      }
+      ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "question = \"Who won the World Series in 2004?\"\n",
-        "answer = search(question)\n",
-        "print(question, answer)\n",
-        "\n",
-        "nquestion = \"Who did they beat?\"\n",
-        "print(nquestion, search(f\"{question} {answer}. {nquestion}\"))"
-      ],
+      "execution_count": 23,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/"
@@ -333,23 +267,28 @@
         "id": "0NNLBwC-83MM",
         "outputId": "4b9fabd7-baf9-4f7d-bb8b-c9c60c5101e4"
       },
-      "execution_count": null,
       "outputs": [
         {
-          "output_type": "stream",
           "name": "stdout",
+          "output_type": "stream",
           "text": [
-            "Who won the World Series in 2004? Boston\n",
-            "Who did they beat? St Louis\n"
+            "Who won the World Series in 2004? The Boston Red Sox won the World Series in 2004.\n",
+            "What team did the Red Sox beat in the World Series? The Boston Red Sox beat the St. Louis Cardinals in the World Series.\n"
           ]
         }
+      ],
+      "source": [
+        "question = \"Who won the World Series in 2004?\"\n",
+        "answer = rag(question)\n",
+        "print(question, answer)\n",
+        "\n",
+        "nquestion = \"What team did the Red Sox beat in the World Series?\"\n",
+        "print(nquestion, rag(f\"{question} {answer}. {nquestion}\"))"
       ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "search(\"Tell me something interesting?\")"
-      ],
+      "execution_count": 29,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/",
@@ -358,43 +297,36 @@
         "id": "1P0zqkTW9cZW",
         "outputId": "4f2232db-761b-464f-b47d-6695c84ffb80"
       },
-      "execution_count": null,
       "outputs": [
         {
-          "output_type": "execute_result",
           "data": {
             "text/plain": [
-              "'herrings communicate by farting'"
-            ],
-            "application/vnd.google.colaboratory.intrinsic+json": {
-              "type": "string"
-            }
+              "'An interesting fact is that herrings communicate by farting—a quirky and unusual discovery that was honored with an Ig Nobel Prize for its oddball research.'"
+            ]
           },
+          "execution_count": 29,
           "metadata": {},
-          "execution_count": 8
+          "output_type": "execute_result"
         }
+      ],
+      "source": [
+        "rag(\"Tell me something interesting\")"
       ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "ygFFcwWPGI9p"
+      },
       "source": [
         "Whhaaaattt???  Is this a model hallucination?\n",
         "\n",
         "Let's run an embeddings query and see if that text is in the results."
-      ],
-      "metadata": {
-        "id": "ygFFcwWPGI9p"
-      }
+      ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "answer = \"herrings communicate by farting\"\n",
-        "for x in embeddings.search(\"Tell me something interesting?\"):\n",
-        "  if answer in x[\"text\"]:\n",
-        "    start = x[\"text\"].find(answer)\n",
-        "    print(x[\"text\"][start:start + len(answer)])"
-      ],
+      "execution_count": 28,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/"
@@ -402,143 +334,68 @@
         "id": "qZPhLqSxGMbK",
         "outputId": "e3b2909a-1ee8-480f-afa3-201a8e27cb08"
       },
-      "execution_count": null,
       "outputs": [
         {
-          "output_type": "stream",
           "name": "stdout",
+          "output_type": "stream",
           "text": [
             "herrings communicate by farting\n"
           ]
         }
+      ],
+      "source": [
+        "answer = \"herrings communicate by farting\"\n",
+        "for x in embeddings.search(\"Tell me something interesting\"):\n",
+        "  if answer in x[\"text\"]:\n",
+        "    start = x[\"text\"].find(answer)\n",
+        "    print(x[\"text\"][start:start + len(answer)])"
       ]
     },
     {
       "cell_type": "markdown",
-      "source": [
-        "Sure enough it is. It appears that the FLAN-T5 model has a bit of an immature sense of humor 😃"
-      ],
       "metadata": {
         "id": "IpgxMc1DZcds"
-      }
-    },
-    {
-      "cell_type": "markdown",
-      "source": [
-        "# External API Integration\n",
-        "\n",
-        "In addition to support for Hugging Face models, the RAG pipeline also supports custom question-answer models. This could be a call to the OpenAI API (GPT-3), Cohere API, Hugging Face API or using langchain to manage that. All that is needed is a Callable object or a function!\n",
-        "\n",
-        "Let's see an example that uses the Hugging Face API to answer questions. We'll use the original sports dataset to demonstrate."
-      ],
-      "metadata": {
-        "id": "7bitjzeyYOqh"
-      }
-    },
-    {
-      "cell_type": "code",
-      "source": [
-        "import requests\n",
-        "\n",
-        "data = [\"Giants hit 3 HRs to down Dodgers\",\n",
-        "        \"Giants 5 Dodgers 4 final\",\n",
-        "        \"Dodgers drop Game 2 against the Giants, 5-4\",\n",
-        "        \"Blue Jays beat Red Sox final score 2-1\",\n",
-        "        \"Red Sox lost to the Blue Jays, 2-1\",\n",
-        "        \"Blue Jays at Red Sox is over. Score: 2-1\",\n",
-        "        \"Phillies win over the Braves, 5-0\",\n",
-        "        \"Phillies 5 Braves 0 final\",\n",
-        "        \"Final: Braves lose to the Phillies in the series opener, 5-0\",\n",
-        "        \"Lightning goaltender pulled, lose to Flyers 4-1\",\n",
-        "        \"Flyers 4 Lightning 1 final\",\n",
-        "        \"Flyers win 4-1\"]\n",
-        "\n",
-        "def prompt(question):\n",
-        "  return f\"\"\"\n",
-        "    Answer the following question using the context below.\n",
-        "    Question: {question}\n",
-        "    Context:\n",
-        "  \"\"\"\n",
-        "\n",
-        "# Submits a series of prompts to the Hugging Face API.\n",
-        "# This call can easily be switched to use the OpenAI API (GPT-3), Cohere API or a library like langchain.\n",
-        "def api(prompts):\n",
-        "  response = requests.post(\"https://api-inference.huggingface.co/models/google/flan-t5-base\",\n",
-        "                           json={\"inputs\": prompts})\n",
-        "\n",
-        "  return [x[\"generated_text\"] for x in response.json()]\n",
-        "\n",
-        "# Create embeddings model with content support\n",
-        "embeddings = Embeddings({\"path\": \"sentence-transformers/all-MiniLM-L6-v2\", \"content\": True})\n",
-        "\n",
-        "# Create RAG instance, submit prompts to the Hugging Face inference API\n",
-        "rag = RAG(embeddings, api)\n",
-        "\n",
-        "questions = [\"What team won the game?\", \"What was score?\"]\n",
-        "\n",
-        "execute = lambda query: rag([(question, query, prompt(question), False) for question in questions], data)\n",
-        "\n",
-        "for query in [\"Red Sox - Blue Jays\", \"Phillies - Braves\", \"Dodgers - Giants\", \"Flyers - Lightning\"]:\n",
-        "    print(\"----\", query, \"----\")\n",
-        "    for answer in execute(query):\n",
-        "        print(answer)\n",
-        "    print()\n"
-      ],
-      "metadata": {
-        "colab": {
-          "base_uri": "https://localhost:8080/"
-        },
-        "id": "yYAa9N7EYm7E",
-        "outputId": "72208327-bfb7-466f-c664-3630f207b6dd"
       },
-      "execution_count": null,
-      "outputs": [
-        {
-          "output_type": "stream",
-          "name": "stdout",
-          "text": [
-            "---- Red Sox - Blue Jays ----\n",
-            "('What team won the game?', 'Blue Jays')\n",
-            "('What was score?', '2-1')\n",
-            "\n",
-            "---- Phillies - Braves ----\n",
-            "('What team won the game?', 'Phillies')\n",
-            "('What was score?', '5-0')\n",
-            "\n",
-            "---- Dodgers - Giants ----\n",
-            "('What team won the game?', 'Giants')\n",
-            "('What was score?', '5-4')\n",
-            "\n",
-            "---- Flyers - Lightning ----\n",
-            "('What team won the game?', 'Flyers')\n",
-            "('What was score?', '4-1')\n",
-            "\n"
-          ]
-        }
+      "source": [
+        "Sure enough it is 😃"
       ]
     },
     {
       "cell_type": "markdown",
-      "source": [
-        "Everything matches with first example above in [Create Embeddings and RAG instances](#Create-Embeddings-and-RAG-instances) except the prompts are run as an external API call.\n",
-        "\n",
-        "The Embeddings instance can also swap out the vectorization, database and vector store components with external API services. Check out the [txtai documentation](https://neuml.github.io/txtai/embeddings/configuration/) documentation for more information.\n",
-        "\n"
-      ],
       "metadata": {
-        "id": "Xh8UhC3NdsGf"
-      }
-    },
-    {
-      "cell_type": "markdown",
+        "id": "KqfvCXp2B3li"
+      },
       "source": [
         "# Wrapping up\n",
         "\n",
         "This notebook covered how to run embeddings-guided and prompt-driven search with LLMs. This functionality is a major step forward towards `Generative Semantic Search` for txtai. More to come, stay tuned!"
-      ],
-      "metadata": {
-        "id": "KqfvCXp2B3li"
-      }
+      ]
     }
-  ]
-}
+  ],
+  "metadata": {
+    "accelerator": "GPU",
+    "colab": {
+      "provenance": []
+    },
+    "gpuClass": "standard",
+    "kernelspec": {
+      "display_name": "local",
+      "language": "python",
+      "name": "python3"
+    },
+    "language_info": {
+      "codemirror_mode": {
+        "name": "ipython",
+        "version": 3
+      },
+      "file_extension": ".py",
+      "mimetype": "text/x-python",
+      "name": "python",
+      "nbconvert_exporter": "python",
+      "pygments_lexer": "ipython3",
+      "version": "3.10.19"
+    }
+  },
+  "nbformat": 4,
+  "nbformat_minor": 0
+}

examples/44_Prompt_templates_and_task_chains.ipynb

@@ -1,17 +1,4 @@
 {
-  "nbformat": 4,
-  "nbformat_minor": 0,
-  "metadata": {
-    "colab": {
-      "provenance": []
-    },
-    "kernelspec": {
-      "name": "python3",
-      "display_name": "Python 3"
-    },
-    "accelerator": "GPU",
-    "gpuClass": "standard"
-  },
   "cells": [
     {
       "cell_type": "markdown",
@@ -41,15 +28,15 @@
     },
     {
       "cell_type": "code",
+      "execution_count": null,
       "metadata": {
         "id": "LPQTb25tASIG"
       },
+      "outputs": [],
       "source": [
         "%%capture\n",
         "!pip install git+https://github.com/neuml/txtai#egg=txtai[api]"
-      ],
-      "execution_count": null,
-      "outputs": []
+      ]
     },
     {
       "cell_type": "markdown",
@@ -61,36 +48,46 @@
         "\n",
         "First, we'll look at building a workflow with a series of model prompts. This workflow creates a conditional translation using a statement and target language. Another task reads that output text and detects the language.\n",
         "\n",
-        "This workflow uses a sequences pipeline. The sequences pipeline loads a Hugging Face sequence to sequence model for inference, in this case [FLAN-T5](https://huggingface.co/google/flan-t5-base). The sequences pipeline takes a prompt as input and outputs the model inference result.\n",
+        "This workflow uses a LLM pipeline. The LLM pipeline loads a local model for inference, in this case [Qwen3-4B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507). The [LLM pipeline](https://neuml.github.io/txtai/pipeline/llm/llm) supports local transformers models, llama.cpp models and LLM APIs such as Ollama, vLLM, OpenAI, Claude etc. \n",
         "\n",
         "It's important to note that a pipeline is simply a callable function. It can easily be replaced with a call to an external API."
       ]
     },
     {
       "cell_type": "code",
+      "execution_count": null,
       "metadata": {
-        "id": "OUc9gqTyAYnm",
         "colab": {
           "base_uri": "https://localhost:8080/"
         },
+        "id": "OUc9gqTyAYnm",
         "outputId": "83300311-736c-47c8-bc16-ec0303274054"
       },
+      "outputs": [
+        {
+          "name": "stdout",
+          "output_type": "stream",
+          "text": [
+            "['French', 'German']\n"
+          ]
+        }
+      ],
       "source": [
-        "from txtai.pipeline import Sequences\n",
-        "from txtai.workflow import Workflow, TemplateTask\n",
+        "from txtai import LLM, Workflow\n",
+        "from txtai.workflow import TemplateTask\n",
         "\n",
-        "# Create sequences pipeline\n",
-        "sequences = Sequences(\"google/flan-t5-large\")\n",
+        "# Create LLM\n",
+        "llm = LLM(\"Qwen/Qwen3-4B-Instruct-2507\")\n",
         "\n",
         "# Define workflow or chaining of tasks together.\n",
         "workflow = Workflow([\n",
         "    TemplateTask(\n",
-        "        template=\"Translate '{statement}' to {language} if it's English\",\n",
-        "        action=sequences\n",
+        "        template=\"Translate text '{statement}' to {language} if the text is English, otherwise keep the original text\",\n",
+        "        action=llm\n",
         "    ),\n",
         "    TemplateTask(\n",
-        "        template=\"What language is the following text? {text}\",\n",
-        "        action=sequences\n",
+        "        template=\"What language is the following text. Only print the answer? {text}\",\n",
+        "        action=llm\n",
         "    )\n",
         "])\n",
         "\n",
@@ -100,20 +97,13 @@
         "]\n",
         "\n",
         "print(list(workflow(inputs)))"
-      ],
-      "execution_count": null,
-      "outputs": [
-        {
-          "output_type": "stream",
-          "name": "stdout",
-          "text": [
-            "['French', 'German']\n"
-          ]
-        }
       ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "_zz4Do8BV-Lk"
+      },
       "source": [
         "Let's recap what happened here. The first workflow task conditionally translates text to a language if it's English.\n",
         "\n",
@@ -122,40 +112,22 @@
         "The second statement is German, so it's not converted to French.\n",
         "\n",
         "The next step asks the model what the language is and it correctly prints `French` and `German`."
-      ],
-      "metadata": {
-        "id": "_zz4Do8BV-Lk"
-      }
+      ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "iXDAKP4CX0W9"
+      },
       "source": [
         "# Prompt Workflow as YAML\n",
         "\n",
         "The same workflow above can be created with YAML configuration."
-      ],
-      "metadata": {
-        "id": "iXDAKP4CX0W9"
-      }
+      ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "%%writefile workflow.yml\n",
-        "\n",
-        "sequences:\n",
-        "  path: google/flan-t5-large\n",
-        "\n",
-        "workflow:\n",
-        "  chain:\n",
-        "    tasks:\n",
-        "      - task: template\n",
-        "        template: Translate '{statement}' to {language} if it's English\n",
-        "        action: sequences\n",
-        "      - task: template\n",
-        "        template: What language is the following text? {text}\n",
-        "        action: sequences"
-      ],
+      "execution_count": 2,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/"
@@ -163,25 +135,35 @@
         "id": "GwV5A9xRYtYs",
         "outputId": "ffe6ee65-95a7-46c6-e6b9-5324eab26ca8"
       },
-      "execution_count": null,
       "outputs": [
         {
-          "output_type": "stream",
           "name": "stdout",
+          "output_type": "stream",
           "text": [
             "Writing workflow.yml\n"
           ]
         }
+      ],
+      "source": [
+        "%%writefile workflow.yml\n",
+        "\n",
+        "llm:\n",
+        "  path: Qwen/Qwen3-4B-Instruct-2507\n",
+        "\n",
+        "workflow:\n",
+        "  chain:\n",
+        "    tasks:\n",
+        "      - task: template\n",
+        "        template: Translate text '{statement}' to {language} if the text is English, otherwise keep the original text\n",
+        "        action: llm\n",
+        "      - task: template\n",
+        "        template: What language is the following text. Only print the answer? {text}\n",
+        "        action: llm"
       ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "from txtai import Application\n",
-        "\n",
-        "app = Application(\"workflow.yml\")\n",
-        "print(list(app.workflow(\"chain\", inputs)))"
-      ],
+      "execution_count": 2,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/"
@@ -189,58 +171,58 @@
         "id": "dr7Lv5S5X98e",
         "outputId": "d6ac0427-671d-4525-aa21-664430109af3"
       },
-      "execution_count": null,
       "outputs": [
         {
-          "output_type": "stream",
           "name": "stdout",
+          "output_type": "stream",
           "text": [
             "['French', 'German']\n"
           ]
         }
+      ],
+      "source": [
+        "from txtai import Application\n",
+        "\n",
+        "app = Application(\"workflow.yml\")\n",
+        "print(list(app.workflow(\"chain\", inputs)))"
       ]
     },
     {
       "cell_type": "markdown",
-      "source": [
-        "As expected, the same result! This is a matter of preference on how you want to create a workflow. One advantage of YAML workflows is that an API can easily be created from the workflow file."
-      ],
       "metadata": {
         "id": "EGqiV45fYVse"
-      }
+      },
+      "source": [
+        "As expected, the same result! This is a matter of preference on how you want to create a workflow. One advantage of YAML workflows is that an API can easily be created from the workflow file."
+      ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "9PqMU0bNYinf"
+      },
       "source": [
         "# Prompt Workflow via an API call\n",
         "\n",
         "Let's say you want the workflow to be available via an API call. Well good news, txtai has a built in API mechanism using FastAPI. "
-      ],
-      "metadata": {
-        "id": "9PqMU0bNYinf"
-      }
+      ]
     },
     {
       "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "id": "vDxQj1ZIYsz3"
+      },
+      "outputs": [],
       "source": [
         "# Start an API service\n",
         "!CONFIG=workflow.yml nohup uvicorn \"txtai.api:app\" &> api.log &\n",
         "!sleep 60"
-      ],
-      "metadata": {
-        "id": "vDxQj1ZIYsz3"
-      },
-      "execution_count": null,
-      "outputs": []
+      ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "import requests\n",
-        "\n",
-        "# Run API request\n",
-        "requests.post(\"http://localhost:8000/workflow\", json={\"name\": \"chain\", \"elements\": inputs}).json()"
-      ],
+      "execution_count": 3,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/"
@@ -248,31 +230,53 @@
         "id": "R1o08SVtZW7h",
         "outputId": "99875acd-18a8-4c2c-ead3-cb6975a4b2d2"
       },
-      "execution_count": null,
       "outputs": [
         {
-          "output_type": "execute_result",
           "data": {
             "text/plain": [
               "['French', 'German']"
             ]
           },
+          "execution_count": 3,
           "metadata": {},
-          "execution_count": 5
+          "output_type": "execute_result"
         }
+      ],
+      "source": [
+        "import requests\n",
+        "\n",
+        "# Run API request\n",
+        "requests.post(\"http://localhost:8000/workflow\", json={\"name\": \"chain\", \"elements\": inputs}).json()"
       ]
     },
     {
       "cell_type": "markdown",
-      "source": [
-        "Just like the previous steps, except through an API call. Let's run via cURL for good measure."
-      ],
       "metadata": {
         "id": "B88mCrGFl5W-"
-      }
+      },
+      "source": [
+        "Just like the previous steps, except through an API call. Let's run via cURL for good measure."
+      ]
     },
     {
       "cell_type": "code",
+      "execution_count": 4,
+      "metadata": {
+        "colab": {
+          "base_uri": "https://localhost:8080/"
+        },
+        "id": "hRUyh0cQl_P2",
+        "outputId": "9db8481d-0b6e-4a31-bdf6-5443df5f768a"
+      },
+      "outputs": [
+        {
+          "name": "stdout",
+          "output_type": "stream",
+          "text": [
+            "[\"French\",\"German\"]"
+          ]
+        }
+      ],
       "source": [
         "%%bash\n",
         "\n",
@@ -287,51 +291,51 @@
         "  ]\n",
         "}\n",
         "EOF"
-      ],
-      "metadata": {
-        "colab": {
-          "base_uri": "https://localhost:8080/"
-        },
-        "id": "hRUyh0cQl_P2",
-        "outputId": "9db8481d-0b6e-4a31-bdf6-5443df5f768a"
-      },
-      "execution_count": null,
-      "outputs": [
-        {
-          "output_type": "stream",
-          "name": "stdout",
-          "text": [
-            "[\"French\",\"German\"]"
-          ]
-        }
       ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "W0zL93WPoaCo"
+      },
       "source": [
         "One last time, the same output is shown.\n",
         "\n",
         "If your primary development environment isn't Python, txtai does have API bindings for [JavaScript](https://github.com/neuml/txtai.js), [Rust](https://github.com/neuml/txtai.rs), [Go](https://github.com/neuml/txtai.go) and [Java](https://github.com/neuml/txtai.java).\n",
         "\n",
         "More information on the API is available [here](https://neuml.github.io/txtai/api/)."
-      ],
-      "metadata": {
-        "id": "W0zL93WPoaCo"
-      }
+      ]
     },
     {
       "cell_type": "markdown",
-      "source": [
-        "# Conversational chain\n",
-        "\n",
-        "Conversational search is another big area of focus in 2023. [txtchat](https://github.com/neuml/txtchat) is a framework for building conversational search applications. It relies heavily on txtai. Let's see a conversational example."
-      ],
       "metadata": {
         "id": "q9WiFG6fpzw5"
-      }
+      },
+      "source": [
+        "# Chat with your data\n",
+        "\n",
+        "\"Chat with your data\" is a popular entry point into the AI space. Let's run an example."
+      ]
     },
     {
       "cell_type": "code",
+      "execution_count": 10,
+      "metadata": {
+        "colab": {
+          "base_uri": "https://localhost:8080/"
+        },
+        "id": "rM3Y551LqF-J",
+        "outputId": "85623785-c15f-4996-9460-0644f69cf5bf"
+      },
+      "outputs": [
+        {
+          "name": "stdout",
+          "output_type": "stream",
+          "text": [
+            "Writing search.yml\n"
+          ]
+        }
+      ],
       "source": [
         "%%writefile search.yml\n",
         "\n",
@@ -341,48 +345,27 @@
         "  container: neuml/txtai-intro\n",
         "\n",
         "rag:\n",
-        "  path: google/flan-t5-large\n",
+        "  path: Qwen/Qwen3-4B-Instruct-2507\n",
         "  output: reference\n",
+        "  template: |\n",
+        "    Answer the following question using only the context below.\n",
+        "\n",
+        "    Question: {question}\n",
+        "    Context: {context}\n",
         "\n",
         "workflow:\n",
         "  search:\n",
         "    tasks:\n",
-        "      - task: rag\n",
-        "        template: |\n",
-        "          Answer the following question using only the context below. Give a detailed answer.\n",
-        "          Say 'I don't have data on that' when the question can't be answered.\n",
-        "          Question: {text}\n",
-        "          Context: \n",
-        "        action: rag\n",
+        "      - action: rag\n",
         "      - task: template\n",
         "        template: \"{answer}\\n\\nReference: {reference}\"\n",
         "        rules:\n",
         "          answer: I don't have data on that"
-      ],
-      "metadata": {
-        "colab": {
-          "base_uri": "https://localhost:8080/"
-        },
-        "id": "rM3Y551LqF-J",
-        "outputId": "85623785-c15f-4996-9460-0644f69cf5bf"
-      },
-      "execution_count": null,
-      "outputs": [
-        {
-          "output_type": "stream",
-          "name": "stdout",
-          "text": [
-            "Overwriting search.yml\n"
-          ]
-        }
       ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "app = Application(\"search.yml\")\n",
-        "print(list(app.workflow(\"search\", [\"Tell me something about North America\"])))"
-      ],
+      "execution_count": null,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/"
@@ -390,40 +373,67 @@
         "id": "1Elb8JANqpwX",
         "outputId": "b1f1ffa1-6c47-4d90-b6f1-8098d4dc45f8"
       },
-      "execution_count": null,
       "outputs": [
         {
-          "output_type": "stream",
           "name": "stdout",
+          "output_type": "stream",
           "text": [
-            "[\"Canada's last fully intact ice shelf has suddenly collapsed, forming a Manhattan-sized iceberg\\n\\nReference: 1\"]\n"
+            "[\"Canada's last fully intact ice shelf has suddenly collapsed, forming a Manhattan-sized iceberg.\\n\\nReference: 1\"]\n"
           ]
         }
+      ],
+      "source": [
+        "app = Application(\"search.yml\")\n",
+        "print(list(app.workflow(\"search\", [\"Find something about North America\"])))"
       ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "4r49V4c9s5nf"
+      },
       "source": [
         "The first thing the code above does is run an embeddings search to build a conversational context. That context is then used to build a prompt and inference is run against the LLM. \n",
         "\n",
-        "The next task formats the outputs with a reference to the best matching record. In this case, it's only an id of 1. But this can be much more useful if the id is a URL or there is logic to format the id back to a unique reference string.\n",
-        "\n",
-        "The [txtchat](https://github.com/neuml/txtchat) project has much more on this, check it out!"
-      ],
-      "metadata": {
-        "id": "4r49V4c9s5nf"
-      }
+        "The next task formats the outputs with a reference to the best matching record. In this case, it's only an id of 1. But this can be much more useful if the id is a URL or there is logic to format the id back to a unique reference string."
+      ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "KqfvCXp2B3li"
+      },
       "source": [
         "# Wrapping up\n",
         "\n",
         "This notebook covered how to build prompt templates and task chains through a series of results. txtai has long had a robust and efficient workflow framework for connecting models together. This can be small and simple models and/or prompting with large models. Go ahead and give it a try!"
-      ],
-      "metadata": {
-        "id": "KqfvCXp2B3li"
-      }
+      ]
     }
-  ]
-}
+  ],
+  "metadata": {
+    "accelerator": "GPU",
+    "colab": {
+      "provenance": []
+    },
+    "gpuClass": "standard",
+    "kernelspec": {
+      "display_name": "local",
+      "language": "python",
+      "name": "python3"
+    },
+    "language_info": {
+      "codemirror_mode": {
+        "name": "ipython",
+        "version": 3
+      },
+      "file_extension": ".py",
+      "mimetype": "text/x-python",
+      "name": "python",
+      "nbconvert_exporter": "python",
+      "pygments_lexer": "ipython3",
+      "version": "3.10.19"
+    }
+  },
+  "nbformat": 4,
+  "nbformat_minor": 0
+}

examples/51_Custom_API_Endpoints.ipynb

@@ -78,8 +78,7 @@
         "\n",
         "# LLM pipeline\n",
         "llm:\n",
-        "  path: google/flan-t5-large\n",
-        "  torch_dtype: torch.bfloat16"
+        "  path: Qwen/Qwen3-4B-Instruct-2507"
       ],
       "metadata": {
         "colab": {

examples/52_Build_RAG_pipelines_with_txtai.ipynb

@@ -1,61 +1,45 @@
 {
-  "nbformat": 4,
-  "nbformat_minor": 0,
-  "metadata": {
-    "colab": {
-      "provenance": [],
-      "gpuType": "T4"
-    },
-    "kernelspec": {
-      "name": "python3",
-      "display_name": "Python 3"
-    },
-    "language_info": {
-      "name": "python"
-    },
-    "accelerator": "GPU"
-  },
   "cells": [
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "VGeVB8M41jqW"
+      },
       "source": [
         "# Build RAG pipelines with txtai\n",
         "\n",
-        "Large Language Models (LLMs) have completely dominated the AI and machine learning space in 2023. The results have been amazing and the public imagination is almost endless.\n",
+        "Large Language Models (LLMs) have completely dominated the tech space in recent years. The results have been amazing and the public imagination is almost endless.\n",
         "\n",
         "While LLMs have been impressive, they are not problem free. The biggest challenge is with hallucinations. Hallucinations is the term for when a LLM generates output that is factually incorrect. The alarming part of this is that on a cursory glance, it actually sounds like good content. The default behavior of LLMs is to produce plausible answers even when no plausible answer exists. LLMs are not great at saying I don't know.\n",
         "\n",
         "Retrieval augmented generation (RAG) helps reduce the risk of hallucinations by limiting the context in which a LLM can generate answers. This is typically done with a vector search query that hydrates a prompt with a relevant context. RAG is one of the most practical and production-ready use cases for *Generative AI*. It's so popular now, that some are creating their entire companies around it.\n",
         "\n",
-        "[txtai](https://github.com/neuml/txtai) has long had question-answering pipelines, which employ the same process of retrieving a relevant context. LLMs are now the preferred approach for analyzing that context and RAG pipelines are one of the main features of txtai. One of the other main features of txtai is that it's a vector database! You can build your prompts and limit your context all with one library. Hence the phrase *all-in-one embeddings database*.\n",
+        "[txtai](https://github.com/neuml/txtai) has long had question-answering pipelines, which employ the same process of retrieving a relevant context. LLMs are now the preferred approach for analyzing that context and RAG pipelines are one of the main features of txtai. One of the other main features of txtai is that it's a vector database! You can build your prompts and limit your context all with one library. Hence the phrase *all-in-one AI framework*.\n",
         "\n",
         "This notebook shows how to build RAG pipelines with txtai."
-      ],
-      "metadata": {
-        "id": "VGeVB8M41jqW"
-      }
+      ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "ZQrHIw351lwE"
+      },
       "source": [
         "# Install dependencies\n",
         "\n",
         "Install `txtai` and all dependencies. Since this notebook is using optional pipelines, we need to install the pipeline extras package."
-      ],
-      "metadata": {
-        "id": "ZQrHIw351lwE"
-      }
+      ]
     },
     {
       "cell_type": "code",
-      "execution_count": 1,
+      "execution_count": null,
       "metadata": {
         "id": "R0AqRP7v1hdr"
       },
       "outputs": [],
       "source": [
         "%%capture\n",
-        "!pip install git+https://github.com/neuml/txtai#egg=txtai[pipeline] autoawq\n",
+        "!pip install git+https://github.com/neuml/txtai#egg=txtai[pipeline]\n",
         "\n",
         "# Get test data\n",
         "!wget -N https://github.com/neuml/txtai/releases/download/v6.2.0/tests.tar.gz\n",
@@ -68,50 +52,43 @@
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "xmPN8RDF1pXd"
+      },
       "source": [
         "# Start with the basics\n",
         "\n",
         "Let's jump right in and start with a simple LLM pipeline. The [LLM pipeline](https://neuml.github.io/txtai/pipeline/text/llm/) supports local LLM models via [Hugging Face Transformers](https://github.com/huggingface/transformers) and [llama.cpp](https://github.com/abetlen/llama-cpp-python).\n",
         "\n",
         "The LLM pipeline also supports [API services (i.e. OpenAI, Claude, Bedrock etc) via LiteLLM](https://github.com/BerriAI/litellm). The LLM pipeline automatically detects the underlying LLM framework from the `path` parameter.\n"
-      ],
-      "metadata": {
-        "id": "xmPN8RDF1pXd"
-      }
+      ]
     },
     {
       "cell_type": "code",
+      "execution_count": 1,
+      "metadata": {
+        "id": "XZ7vPBIs1rGZ"
+      },
+      "outputs": [],
       "source": [
         "from txtai import LLM\n",
         "\n",
         "# Create LLM\n",
-        "llm = LLM(\"TheBloke/Mistral-7B-OpenOrca-AWQ\")"
-      ],
-      "metadata": {
-        "id": "XZ7vPBIs1rGZ"
-      },
-      "execution_count": 21,
-      "outputs": []
+        "llm = LLM(\"Qwen/Qwen3-4B-Instruct-2507\")"
+      ]
     },
     {
       "cell_type": "markdown",
-      "source": [
-        "Next, we'll load a document to query. The [Textractor pipeline](https://neuml.github.io/txtai/pipeline/data/textractor/) has support for extracting text from common document formats (docx, pdf, xlsx)."
-      ],
       "metadata": {
         "id": "9rmTWMxAH3Vx"
-      }
+      },
+      "source": [
+        "Next, we'll load a document to query. The [Textractor pipeline](https://neuml.github.io/txtai/pipeline/data/textractor/) has support for extracting text from common document formats (docx, pdf, xlsx, web)."
+      ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "from txtai.pipeline import Textractor\n",
-        "\n",
-        "# Create Textractor\n",
-        "textractor = Textractor()\n",
-        "text = textractor(\"txtai/document.docx\")\n",
-        "print(text)"
-      ],
+      "execution_count": null,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/"
@@ -119,24 +96,24 @@
         "id": "nifStGtOHuyc",
         "outputId": "5a4010e0-75f9-4095-a24c-cd4c859847d0"
       },
-      "execution_count": 22,
       "outputs": [
         {
-          "output_type": "stream",
           "name": "stdout",
+          "output_type": "stream",
           "text": [
-            "txtai – the all-in-one embeddings database\n",
+            "# txtai – the all-in-one embeddings database\n",
             "txtai is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\n",
             "\n",
             "Summary of txtai features:\n",
-            "· Vector search with SQL, object storage, topic modeling\n",
-            "· Create embeddings for text, documents, audio, images and video\n",
-            "· Pipelines powered by language models that run LLM prompts\n",
-            "· Workflows to join pipelines together and aggregate business logic\n",
-            "· Build with Python or YAML. API bindings available for JavaScript, Java, Rust and Go.\n",
-            "· Run local or scale out with container orchestration\n",
+            "· *Vector search* with SQL, object storage, topic modeling\n",
+            "· Create *embeddings* for text, documents, audio, images and video\n",
+            "· *Pipelines* powered by language models that run LLM prompts\n",
+            "· *Workflows* to join pipelines together and aggregate business logic\n",
+            "· Build with *Python* or *YAML* . API bindings available for JavaScript, Java, Rust and Go.\n",
+            "· *Run local or scale out with container orchestration* \n",
             "\n",
-            "Examples\n",
+            "\n",
+            "## Examples\n",
             "List of example notebooks.\n",
             "|Notebook|Description|\n",
             "|---|---|\n",
@@ -145,16 +122,17 @@
             "|Build a QA database|Question matching with semantic search|\n",
             "|Semantic Graphs|Explore topics, data connectivity and run network analysis|\n",
             "\n",
-            "Install\n",
+            "## Install\n",
             "The easiest way to install is via pip and PyPI\n",
             "pip install txtai\n",
-            "Python 3.9+ is supported. Using a Python virtual environment is recommended.\n",
+            "Python 3.11+ is supported. Using a Python virtual environment is **recommended** .\n",
             "See the detailed install instructions for more information covering optional dependencies, environment specific prerequisites, installing from source, conda support and how to run with containers.\n",
             "\n",
             "\n",
             "\n",
             "\n",
-            "Model guide\n",
+            "\n",
+            "## Model guide\n",
             "The following shows a list of suggested models.\n",
             "|Component|Model(s)|\n",
             "|---|---|\n",
@@ -171,35 +149,28 @@
             "|Translation|OPUS Model Series|\n"
           ]
         }
+      ],
+      "source": [
+        "from txtai import Textractor\n",
+        "\n",
+        "# Create Textractor\n",
+        "textractor = Textractor()\n",
+        "text = textractor(\"txtai/document.docx\")\n",
+        "print(text)"
       ]
     },
     {
       "cell_type": "markdown",
-      "source": [
-        "Now we'll define a simple LLM pipeline. It takes a question and context (which in this case is the whole file), creates a prompt and runs it with the LLM."
-      ],
       "metadata": {
         "id": "2jkamwgdIgEp"
-      }
+      },
+      "source": [
+        "Now we'll define a simple LLM pipeline. It takes a question and context (which in this case is the whole file), creates a prompt and runs it with the LLM."
+      ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "def execute(question, text):\n",
-        "  prompt = f\"\"\"<|im_start|>system\n",
-        "  You are a friendly assistant. You answer questions from users.<|im_end|>\n",
-        "  <|im_start|>user\n",
-        "  Answer the following question using only the context below. Only include information specifically discussed.\n",
-        "\n",
-        "  question: {question}\n",
-        "  context: {text} <|im_end|>\n",
-        "  <|im_start|>assistant\n",
-        "  \"\"\"\n",
-        "\n",
-        "  return llm(prompt, maxlength=4096, pad_token_id=32000)\n",
-        "\n",
-        "execute(\"Tell me about txtai in one sentence\", text)"
-      ],
+      "execution_count": 18,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/",
@@ -208,28 +179,36 @@
         "id": "9HU6C0OIIAKn",
         "outputId": "f9d556c4-cd7a-4774-ef62-1f3fff90aa47"
       },
-      "execution_count": 23,
       "outputs": [
         {
-          "output_type": "execute_result",
           "data": {
             "text/plain": [
-              "'Txtai is an all-in-one embeddings database for semantic search, LLM orchestration, and language model workflows, offering features such as vector search, pipeline creation, workflow management, and API bindings for various programming languages.'"
-            ],
-            "application/vnd.google.colaboratory.intrinsic+json": {
-              "type": "string"
-            }
+              "'txtai is an all-in-one embeddings database that supports semantic search, LLM orchestration, and language model workflows with features like vector search, embeddings for text, audio, images, and video, pipelines powered by language models, and scalable workflows available via Python or YAML with API bindings for multiple languages.'"
+            ]
           },
+          "execution_count": 18,
           "metadata": {},
-          "execution_count": 23
+          "output_type": "execute_result"
         }
+      ],
+      "source": [
+        "def execute(question, text):\n",
+        "  return llm([\n",
+        "    {\"role\": \"system\", \"content\": \"You are a friendly assistant. You answer questions from users.\"},\n",
+        "    {\"role\": \"user\", \"content\": f\"\"\"\n",
+        "        Answer the following question using only the context below. Only include information specifically discussed.\n",
+        "\n",
+        "        question: {question}\n",
+        "        context: {text} \n",
+        "    \"\"\"}\n",
+        "  ], maxlength=4096)\n",
+        "\n",
+        "execute(\"Tell me about txtai in one sentence\", text)"
       ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "execute(\"What model does txtai recommend for transcription?\", text)"
-      ],
+      "execution_count": 19,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/",
@@ -238,28 +217,25 @@
         "id": "xxF7ajCPJP5_",
         "outputId": "0e6f6dbb-c784-4841-fe3f-82754ef478eb"
       },
-      "execution_count": 24,
       "outputs": [
         {
-          "output_type": "execute_result",
           "data": {
             "text/plain": [
-              "'The model that txtai recommends for transcription is Whisper.'"
-            ],
-            "application/vnd.google.colaboratory.intrinsic+json": {
-              "type": "string"
-            }
+              "'txtai recommends using Whisper for transcription.'"
+            ]
           },
+          "execution_count": 19,
           "metadata": {},
-          "execution_count": 24
+          "output_type": "execute_result"
         }
+      ],
+      "source": [
+        "execute(\"What model does txtai recommend for transcription?\", text)"
       ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "execute(\"I don't know anything about txtai, what would be the best thing to read?\", text)"
-      ],
+      "execution_count": 20,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/",
@@ -268,57 +244,38 @@
         "id": "AKmmTqsnJa5X",
         "outputId": "834bf3ee-b7ed-4e38-e2ef-d5950f99ed9a"
       },
-      "execution_count": 25,
       "outputs": [
         {
-          "output_type": "execute_result",
           "data": {
             "text/plain": [
-              "'The best thing to read to learn about txtai is the \"Introducing txtai\" notebook, which provides an overview of the functionality provided by txtai. This notebook covers various features such as vector search with SQL, object storage, topic modeling, creating embeddings for text, documents, audio, images, and video, and running language model workflows. Additionally, you can explore other example notebooks like \"Similarity search with images,\" \"Build a QA database,\" and \"Semantic Graphs\" to learn more about specific use cases and features. To install txtai, use pip and PyPI with Python 3.9+, and follow the detailed install instructions for more information on optional dependencies and environment-specific prerequisites.'"
-            ],
-            "application/vnd.google.colaboratory.intrinsic+json": {
-              "type": "string"
-            }
+              "'The best thing to read if you don\\'t know anything about txtai would be the \"Introducing txtai\" notebook, as it provides an overview of the functionality offered by txtai.'"
+            ]
           },
+          "execution_count": 20,
           "metadata": {},
-          "execution_count": 25
+          "output_type": "execute_result"
         }
+      ],
+      "source": [
+        "execute(\"I don't know anything about txtai, what would be the best thing to read?\", text)"
       ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "WaVeEHrIMpFr"
+      },
       "source": [
         "If this is the first time you've seen *Generative AI*, then these statements are 🤯. Even if you've been in the space a while, it's still amazing how much a language model can understand and the high level of quality in it's answers.\n",
         "\n",
         "While this use case is fun, lets try to scale it to a larger set of documents."
-      ],
-      "metadata": {
-        "id": "WaVeEHrIMpFr"
-      }
-    },
-    {
-      "cell_type": "markdown",
-      "metadata": {},
-      "source": [
-        "_Before continuing, it's important to note that txtai has multiple ways to run LLM inference. In the past, prior to \"Chat Templates\", it was expected that the submitted text had all the required chat tokens embedded. The same prompt above can also be written with chat messages. This is especially important when working with LLM APIs (i.e. OpenAI, Claude, Bedrock etc)._\n",
-        "\n",
-        "```python\n",
-        "llm([\n",
-        "    {\"role\": \"system\": \"You are a friendly assistant. You answer questions from users.\"}\n",
-        "    {\"role\": \"user\", \"content\": f\"\"\"\n",
-        "        Answer the following question using only the context below. Only include information specifically discussed.\n",
-        "\n",
-        "        question: {question}\n",
-        "        context: {text} \n",
-        "    \"\"\"}\n",
-        "])\n",
-        "```\n",
-        "\n",
-        "_See the [LLM pipeline documentation](https://neuml.github.io/txtai/pipeline/text/llm/) for more information._"
       ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "viVVft59NbKv"
+      },
       "source": [
         "# Build a RAG pipeline with vector search\n",
         "\n",
@@ -329,13 +286,30 @@
         "Typically, this candidate query uses vector search but it can be anything that runs a search and returns results. In fact, many complex production systems have customized retrieval pipelines that feed a context into LLM prompts.\n",
         "\n",
         "The first step in building our RAG pipeline is creating the knowledge store. In this case, it's a vector database of file content. The files will be split into paragraphs with each paragraph stored as a separate row."
-      ],
-      "metadata": {
-        "id": "viVVft59NbKv"
-      }
+      ]
     },
     {
       "cell_type": "code",
+      "execution_count": null,
+      "metadata": {
+        "colab": {
+          "base_uri": "https://localhost:8080/"
+        },
+        "id": "ipmsmtN1NahT",
+        "outputId": "64733e1f-fb7b-4a2d-bf02-8930478a8ee8"
+      },
+      "outputs": [
+        {
+          "name": "stdout",
+          "output_type": "stream",
+          "text": [
+            "Indexing txtai/article.pdf\n",
+            "Indexing txtai/document.docx\n",
+            "Indexing txtai/document.pdf\n",
+            "Indexing txtai/spreadsheet.xlsx\n"
+          ]
+        }
+      ],
       "source": [
         "import os\n",
         "\n",
@@ -357,39 +331,40 @@
         "# Vector Database\n",
         "embeddings = Embeddings(content=True)\n",
         "embeddings.index(stream(\"txtai\"))"
-      ],
-      "metadata": {
-        "colab": {
-          "base_uri": "https://localhost:8080/"
-        },
-        "id": "ipmsmtN1NahT",
-        "outputId": "64733e1f-fb7b-4a2d-bf02-8930478a8ee8"
-      },
-      "execution_count": 26,
-      "outputs": [
-        {
-          "output_type": "stream",
-          "name": "stdout",
-          "text": [
-            "Indexing txtai/article.pdf\n",
-            "Indexing txtai/document.docx\n",
-            "Indexing txtai/document.pdf\n",
-            "Indexing txtai/spreadsheet.xlsx\n"
-          ]
-        }
       ]
     },
     {
       "cell_type": "markdown",
-      "source": [
-        "The next step is defining the RAG pipeline. This pipeline takes the input question, runs a vector search and builds a context using the search results. The context is then inserted into a prompt template and run with the LLM."
-      ],
       "metadata": {
         "id": "ASlmAaR3nBPN"
-      }
+      },
+      "source": [
+        "The next step is defining the RAG pipeline. This pipeline takes the input question, runs a vector search and builds a context using the search results. The context is then inserted into a prompt template and run with the LLM."
+      ]
     },
     {
       "cell_type": "code",
+      "execution_count": 23,
+      "metadata": {
+        "colab": {
+          "base_uri": "https://localhost:8080/",
+          "height": 39
+        },
+        "id": "_-9SW6r4P5ha",
+        "outputId": "6a7bcd69-bcd0-4f6e-81c1-e32c323b3ffb"
+      },
+      "outputs": [
+        {
+          "data": {
+            "text/plain": [
+              "'txtai recommends using BLIP for image captioning.'"
+            ]
+          },
+          "execution_count": 23,
+          "metadata": {},
+          "output_type": "execute_result"
+        }
+      ],
       "source": [
         "def context(question):\n",
         "  context =  \"\\n\".join(x[\"text\"] for x in embeddings.search(question))\n",
@@ -399,38 +374,11 @@
         "  return execute(question, context(question))\n",
         "\n",
         "rag(\"What model does txtai recommend for image captioning?\")"
-      ],
-      "metadata": {
-        "colab": {
-          "base_uri": "https://localhost:8080/",
-          "height": 39
-        },
-        "id": "_-9SW6r4P5ha",
-        "outputId": "6a7bcd69-bcd0-4f6e-81c1-e32c323b3ffb"
-      },
-      "execution_count": 27,
-      "outputs": [
-        {
-          "output_type": "execute_result",
-          "data": {
-            "text/plain": [
-              "'Based on the provided context, txtai recommends the model \"BLIP\" for image captioning.'"
-            ],
-            "application/vnd.google.colaboratory.intrinsic+json": {
-              "type": "string"
-            }
-          },
-          "metadata": {},
-          "execution_count": 27
-        }
       ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "result = rag(\"When was the BLIP model added for image captioning?\")\n",
-        "print(result)"
-      ],
+      "execution_count": 24,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/"
@@ -438,47 +386,47 @@
         "id": "NbQhSunPQtB0",
         "outputId": "de0caf04-4cdb-48e8-aadf-37283be9909a"
       },
-      "execution_count": 28,
       "outputs": [
         {
-          "output_type": "stream",
           "name": "stdout",
+          "output_type": "stream",
           "text": [
             "The BLIP model was added for image captioning on 2022-03-17.\n"
           ]
         }
+      ],
+      "source": [
+        "result = rag(\"When was the BLIP model added for image captioning?\")\n",
+        "print(result)"
       ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "D6HW-3GtnTFl"
+      },
       "source": [
         "As we can see, the result is similar to what we had before without vector search. The difference is that we only used a relevant portion of the documents to generate the answer.\n",
         "\n",
         "As we discussed before, this is important when dealing with large volumes of data. Not all of the data can be added to a LLM prompt. Additionally, having only the most relevant context helps the LLM generate higher quality answers."
-      ],
-      "metadata": {
-        "id": "D6HW-3GtnTFl"
-      }
+      ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "FZ-yPC-xiUqa"
+      },
       "source": [
         "# Citations for LLMs\n",
         "\n",
         "A healthy level of skepticism should be applied to answers generated by AI. We're far from the day where we can blindly trust answers from an AI model.\n",
         "\n",
         "txtai has a couple approaches for generating citations. The basic approach is to take the answer and search the vector database for the closest match."
-      ],
-      "metadata": {
-        "id": "FZ-yPC-xiUqa"
-      }
+      ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "for x in embeddings.search(result):\n",
-        "  print(x[\"text\"])"
-      ],
+      "execution_count": 25,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/"
@@ -486,16 +434,15 @@
         "id": "obttSg_dSFT5",
         "outputId": "c7ae7675-6959-4bcb-ad06-065ea8609c31"
       },
-      "execution_count": 29,
       "outputs": [
         {
-          "output_type": "stream",
           "name": "stdout",
+          "output_type": "stream",
           "text": [
             "E5-base-v2\n",
             "Image Captions BLIP\n",
             "Labels - Zero Shot BART-Large-MNLI\n",
-            "Model Guide\n",
+            "# Model Guide\n",
             "|Component |Model(s)|Date Added|\n",
             "|---|---|---|\n",
             "|Embeddings |all-MiniLM-L6-v2|2022-04-15|\n",
@@ -507,66 +454,65 @@
             "|Transcription |Whisper|2022-08-01|\n",
             "|Translation |OPUS Model Series|2021-04-06|\n",
             "&\"Times New Roman,Regular\"&12&A\n",
-            "Notebook Description\n",
-            "Introducing txtai Overview of the functionality provided by txtai\n",
-            "Similarity search with \n",
-            "images Embed images and text into the same space for search\n"
+            "## Model guide\n",
+            "The following shows a list of suggested models.\n",
+            "|Component|Model(s)|\n",
+            "|---|---|\n",
+            "|Embeddings|all-MiniLM-L6-v2|\n",
+            "||E5-base-v2|\n",
+            "|Image Captions|BLIP|\n",
+            "|Labels - Zero Shot|BART-Large-MNLI|\n",
+            "|Labels - Fixed|Fine-tune with training pipeline|\n",
+            "|Large Language Model (LLM)|Flan T5 XL|\n",
+            "||Mistral 7B OpenOrca|\n",
+            "|Summarization|DistilBART|\n",
+            "|Text-to-Speech|ESPnet JETS|\n",
+            "|Transcription|Whisper|\n",
+            "|Translation|OPUS Model Series|\n"
           ]
         }
+      ],
+      "source": [
+        "for x in embeddings.search(result):\n",
+        "  print(x[\"text\"])"
       ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "MDcB7GCWY6TO"
+      },
       "source": [
         "While the basic approach above works in this case, txtai has a more robust pipeline to handle citations and references.\n",
         "\n",
         "The RAG pipeline is defined below. A RAG pipeline works in the same way as a LLM + Vector Search pipeline, except it has special logic for generating citations. This pipeline takes the answers and compares it to the context passed to the LLM to determine the most likely reference."
-      ],
-      "metadata": {
-        "id": "MDcB7GCWY6TO"
-      }
+      ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "from txtai import RAG\n",
-        "\n",
-        "# RAG prompt\n",
-        "def prompt(question):\n",
-        "  return [{\n",
-        "    \"query\": question,\n",
-        "    \"question\": f\"\"\"\n",
-        "Answer the following question using only the context below. Only include information specifically discussed.\n",
-        "\n",
-        "question: {question}\n",
-        "context:\n",
-        "\"\"\"\n",
-        "}]\n",
-        "\n",
-        "# Create LLM with system prompt template\n",
-        "llm = LLM(\"TheBloke/Mistral-7B-OpenOrca-AWQ\", template=\"\"\"<|im_start|>system\n",
-        "You are a friendly assistant. You answer questions from users.<|im_end|>\n",
-        "<|im_start|>user\n",
-        "{text} <|im_end|>\n",
-        "<|im_start|>assistant\n",
-        "\"\"\")\n",
-        "\n",
-        "# Create RAG instance\n",
-        "rag = RAG(embeddings, llm, output=\"reference\")"
-      ],
+      "execution_count": 26,
       "metadata": {
         "id": "Lm6gg85_Y7ot"
       },
-      "execution_count": null,
-      "outputs": []
+      "outputs": [],
+      "source": [
+        "from txtai import RAG\n",
+        "\n",
+        "# Create the RAG pipeline\n",
+        "rag = RAG(embeddings, \"Qwen/Qwen3-4B-Instruct-2507\", template=\"\"\"\n",
+        "  Answer the following question using the provided context.\n",
+        "\n",
+        "  Question:\n",
+        "  {question}\n",
+        "\n",
+        "  Context:\n",
+        "  {context}\n",
+        "\"\"\", output=\"reference\")"
+      ]
     },
     {
       "cell_type": "code",
-      "source": [
-        "result = rag(prompt(\"What version of Python is supported?\"), maxlength=4096, pad_token_id=32000)[0]\n",
-        "print(\"ANSWER:\", result[\"answer\"])\n",
-        "print(\"CITATION:\", embeddings.search(\"select id, text from txtai where id = :id\", limit=1, parameters={\"id\": result[\"reference\"]}))"
-      ],
+      "execution_count": 27,
       "metadata": {
         "colab": {
           "base_uri": "https://localhost:8080/"
@@ -574,36 +520,36 @@
         "id": "4pOfE5paZatH",
         "outputId": "2bed2de5-22ff-4f7b-dba5-41b8e4cc6c75"
       },
-      "execution_count": 31,
       "outputs": [
         {
-          "output_type": "stream",
           "name": "stdout",
+          "output_type": "stream",
           "text": [
-            "ANSWER: Python 3.9+ is supported. Using a Python virtual environment is recommended. The easiest way to install is via pip and PyPI.\n",
-            "CITATION: [{'id': '24', 'text': 'Python 3.9+ is supported. Using a Python virtual environment is recommended.'}]\n"
+            "ANSWER: Python 3.11 and later versions (Python 3.11+) are supported.\n",
+            "CITATION: [{'id': '24', 'text': 'Python 3.11+ is supported. Using a Python virtual environment is recommended.'}]\n"
           ]
         }
+      ],
+      "source": [
+        "result = rag(\"What version of Python is supported?\", maxlength=4096)\n",
+        "print(\"ANSWER:\", result[\"answer\"])\n",
+        "print(\"CITATION:\", embeddings.search(\"select id, text from txtai where id = :id\", limit=1, parameters={\"id\": result[\"reference\"]}))"
       ]
     },
     {
       "cell_type": "markdown",
-      "source": [
-        "And as we can see, not only is the answer to the statement shown, the RAG pipeline also provides a citation. This step is crucial in any line of work where answers must be verified (which is most lines of work)."
-      ],
       "metadata": {
         "id": "vHdE2Q59jNnF"
-      }
-    },
-    {
-      "cell_type": "markdown",
-      "metadata": {},
+      },
       "source": [
-        "_As with the LLM pipeline, the RAG pipeline also supports chat messages. See the [RAG pipeline documentation](https://neuml.github.io/txtai/pipeline/text/rag/) for more._"
+        "And as we can see, not only is the answer to the statement shown, the RAG pipeline also provides a citation. This step is crucial in any line of work where answers must be verified (which is most lines of work)."
       ]
     },
     {
       "cell_type": "markdown",
+      "metadata": {
+        "id": "oPwgCgBc2Er2"
+      },
       "source": [
         "# Wrapping up\n",
         "\n",
@@ -611,13 +557,36 @@
         "\n",
         "The advantages of building RAG pipelines with txtai are:\n",
         "\n",
-        "- **All-in-one database** - one library can handle LLM inference and vector search retrieval\n",
+        "- **All-in-one AI framework** - one library can handle LLM inference and vector search retrieval\n",
         "- **Generating citations** - generating answers is useful but referencing where those answers came from is crucial in gaining the trust of users\n",
         "- **Simple yet powerful** - building pipelines can be done in a small amount of Python. Options are available to build pipelines in YAML and/or run through the API"
-      ],
-      "metadata": {
-        "id": "oPwgCgBc2Er2"
-      }
+      ]
     }
-  ]
+  ],
+  "metadata": {
+    "accelerator": "GPU",
+    "colab": {
+      "gpuType": "T4",
+      "provenance": []
+    },
+    "kernelspec": {
+      "display_name": "local",
+      "language": "python",
+      "name": "python3"
+    },
+    "language_info": {
+      "codemirror_mode": {
+        "name": "ipython",
+        "version": 3
+      },
+      "file_extension": ".py",
+      "mimetype": "text/x-python",
+      "name": "python",
+      "nbconvert_exporter": "python",
+      "pygments_lexer": "ipython3",
+      "version": "3.10.19"
+    }
+  },
+  "nbformat": 4,
+  "nbformat_minor": 0
 }

examples/53_Integrate_LLM_Frameworks.ipynb

@@ -90,7 +90,7 @@
         "embeddings.index(data)\n",
         "\n",
         "# Create LLM with llama.cpp - GGUF file is automatically downloaded\n",
-        "llm = LLM(\"TheBloke/Mistral-7B-OpenOrca-GGUF/mistral-7b-openorca.Q4_K_M.gguf\", verbose=True)\n",
+        "llm = LLM(\"unsloth/Qwen3-4B-Instruct-2507-GGUF/Qwen3-4B-Instruct-2507-Q4_K_M.gguf\", verbose=True)\n",
         "\n",
         "template = \"\"\"<|im_start|>system\n",
         "You are a friendly assistant. You answer questions from users.<|im_end|>\n",

examples/54_API_Authorization_and_Authentication.ipynb

@@ -19,7 +19,7 @@
       "source": [
         "# API Authorization and Authentication\n",
         "\n",
-        "[txtai](https://github.com/neuml/txtai) is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows. txtai can run in Python, with YAML configuration and through an API service.\n",
+        "txtai can run in Python, with YAML configuration and through an API service.\n",
         "\n",
         "The default API service implementation runs without any security. This may be OK for a local prototype or if it's run on a small internal network. But in most cases, additional security measures should be taken.\n",
         "\n",

examples/55_Generate_knowledge_with_Semantic_Graphs_and_RAG.ipynb

@@ -8,8 +8,6 @@
       "source": [
         "# Generate knowledge with Semantic Graphs and RAG\n",
         "\n",
-        "[txtai](https://github.com/neuml/txtai) is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\n",
-        "\n",
         "Embeddings databases are a union of vector indexes (sparse and dense), graph networks and relational databases. This enables vector search with SQL, topic modeling, retrieval augmented generation and more.\n",
         "\n",
         "Semantic graphs were introduced in late 2022 with txtai 5.0, right before the start of the LLM era. This notebook picks back up where that work left off, opening a new frontier in semantic graph integration."
@@ -35,7 +33,7 @@
       "outputs": [],
       "source": [
         "%%capture\n",
-        "!pip install git+https://github.com/neuml/txtai#egg=txtai[graph] autoawq"
+        "!pip install git+https://github.com/neuml/txtai#egg=txtai[graph]"
       ]
     },
     {
@@ -296,7 +294,7 @@
       "outputs": [],
       "source": [
         "from txtai import LLM\n",
-        "llm = LLM(\"TheBloke/Mistral-7B-OpenOrca-AWQ\")"
+        "llm = LLM(\"Qwen/Qwen3-4B-Instruct-2507\")"
       ]
     },
     {

examples/56_External_vectorization.ipynb

@@ -21,8 +21,6 @@
       "source": [
         "# External vectorization\n",
         "\n",
-        "[txtai](https://github.com/neuml/txtai) is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\n",
-        "\n",
         "Vectorization is the process of transforming data into numbers using machine learning models. Input data is run through a model and fixed dimension vectors are returned. These vectors can then be loaded into a vector database for similarity search.\n",
         "\n",
         "txtai is an open-source first system. Given it's own open-source roots, like-minded projects such as [sentence-transformers](https://github.com/UKPLab/sentence-transformers) are prioritized during development. But that doesn't mean txtai can't work with Embeddings API services.\n",

examples/57_Build_knowledge_graphs_with_LLM_driven_entity_extraction.ipynb

@@ -8,8 +8,6 @@
       "source": [
         "# Build knowledge graphs with LLM-driven entity extraction\n",
         "\n",
-        "[txtai](https://github.com/neuml/txtai) is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\n",
-        "\n",
         "Embeddings databases are a union of vector indexes (sparse and dense), graph networks and relational databases. This enables vector search with SQL, topic modeling, retrieval augmented generation and more.\n",
         "\n",
         "Up until txtai 7.0, semantic graphs only supported automatic relationship detection. Now relationships can be loaded directly into a txtai database. This notebook will demonstrate how to work with this new feature."
@@ -35,7 +33,7 @@
       "outputs": [],
       "source": [
         "%%capture\n",
-        "!pip install git+https://github.com/neuml/txtai#egg=txtai[graph] autoawq"
+        "!pip install git+https://github.com/neuml/txtai#egg=txtai[graph]"
       ]
     },
     {
@@ -99,7 +97,7 @@
         "from txtai import LLM\n",
         "\n",
         "# Load LLM\n",
-        "llm = LLM(\"TheBloke/Mistral-7B-OpenOrca-AWQ\")\n",
+        "llm = LLM(\"Qwen/Qwen3-4B-Instruct-2507\")\n",
         "\n",
         "data = []\n",
         "for result in results:\n",

examples/58_Advanced_RAG_with_graph_path_traversal.ipynb

@@ -8,8 +8,6 @@
       "source": [
         "# Advanced RAG with graph path traversal\n",
         "\n",
-        "[txtai](https://github.com/neuml/txtai) is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\n",
-        "\n",
         "Embeddings databases are a union of vector indexes (sparse and dense), graph networks and relational databases. This enables vector search with SQL, topic modeling, retrieval augmented generation (RAG) and more.\n",
         "\n",
         "A standard RAG process typically runs a single vector search query and returns the closest matches. Those matches are then passed into a LLM prompt and used to limit the context and help ensure more factually correct answers are generated. This works well with most simple cases. More complex use cases, require a more advanced approach.\n",
@@ -37,7 +35,7 @@
       "outputs": [],
       "source": [
         "%%capture\n",
-        "!pip install git+https://github.com/neuml/txtai#egg=txtai[graph] autoawq"
+        "!pip install git+https://github.com/neuml/txtai#egg=txtai[graph]"
       ]
     },
     {
@@ -215,7 +213,7 @@
       "source": [
         "from txtai import LLM\n",
         "\n",
-        "llm = LLM(\"TheBloke/Mistral-7B-OpenOrca-AWQ\")"
+        "llm = LLM(\"Qwen/Qwen3-4B-Instruct-2507\")"
       ]
     },
     {

examples/59_Whats_new_in_txtai_7_0.ipynb

@@ -54,7 +54,7 @@
       "outputs": [],
       "source": [
         "%%capture\n",
-        "!pip install git+https://github.com/neuml/txtai#egg=txtai[api,graph,pipeline-train] datasets autoawq"
+        "!pip install git+https://github.com/neuml/txtai#egg=txtai[api,graph,pipeline-train] datasets"
       ]
     },
     {
@@ -244,7 +244,7 @@
       "source": [
         "from txtai import LLM\n",
         "\n",
-        "llm = LLM(\"TheBloke/Mistral-7B-OpenOrca-AWQ\")\n",
+        "llm = LLM(\"Qwen/Qwen3-4B-Instruct-2507\")\n",
         "\n",
         "def rag(question, text):\n",
         "    prompt = f\"\"\"<|im_start|>system\n",

examples/60_Advanced_RAG_with_guided_generation.ipynb

@@ -8,8 +8,6 @@
       "source": [
         "# Advanced RAG with guided generation\n",
         "\n",
-        "[txtai](https://github.com/neuml/txtai) is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\n",
-        "\n",
         "A standard RAG process typically runs a single vector search query and returns the closest matches. Those matches are then passed into a LLM prompt and used to limit the context and help ensure more factually correct answers are generated. This works well with most simple cases. More complex use cases, require a more advanced approach.\n",
         "\n",
         "This notebook will demonstrate how constrained or guided generation can be applied to better control LLM output."
@@ -35,7 +33,7 @@
       "outputs": [],
       "source": [
         "%%capture\n",
-        "!pip install git+https://github.com/neuml/txtai#egg=txtai autoawq outlines"
+        "!pip install git+https://github.com/neuml/txtai#egg=txtai outlines"
       ]
     },
     {
@@ -59,7 +57,7 @@
       "source": [
         "from txtai import LLM\n",
         "\n",
-        "llm = LLM(\"TheBloke/Mistral-7B-OpenOrca-AWQ\")\n",
+        "llm = LLM(\"Qwen/Qwen3-4B-Instruct-2507\")\n",
         "\n",
         "def rag(question, text):\n",
         "    prompt = f\"\"\"<|im_start|>system\n",

examples/61_Integrate_txtai_with_Postgres.ipynb

@@ -21,8 +21,6 @@
       "source": [
         "# Integrate txtai with Postgres\n",
         "\n",
-        "[txtai](https://github.com/neuml/txtai) is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\n",
-        "\n",
         "The default persistence methods for txtai are local and file-based. `SQLite` for content, `Faiss` for vectors and `NetworkX` for graph data. The main value add of txtai is getting up and running quickly with a minimal amount of external dependencies.\n",
         "\n",
         "Another key feature of txtai is being able to quickly move from prototyping to production. This notebook will demonstrate how txtai can integrate with [Postgres](https://www.postgresql.org/), a powerful, production-ready and open source object-relational database system. After txtai persists content to Postgres, we'll show it can be directly queried with SQL from any Postgres client 🔥"

examples/62_RAG_with_llama_cpp_and_external_API_services.ipynb

@@ -8,8 +8,6 @@
       "source": [
         "# RAG with llama.cpp and external API services\n",
         "\n",
-        "[txtai](https://github.com/neuml/txtai) is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\n",
-        "\n",
         "txtai has been and always will be a local-first framework. It was originally designed to run models on local hardware using Hugging Face Transformers. As the AI space has evolved over the last year, so has txtai. Additional LLM inference frameworks have been available for a while using llama.cpp and external API services (via LiteLLM). Recent changes have added the ability to use these frameworks for vectorization and made it easier to use for LLM inference.\n",
         "\n",
         "This notebook will demonstrate how to run retrieval-augmented-generation (RAG) processes (vectorization and LLM inference) with llama.cpp and external API services."
@@ -182,7 +180,7 @@
         "from txtai import LLM\n",
         "\n",
         "# LLM instance\n",
-        "llm = LLM(path=\"TheBloke/Mistral-7B-OpenOrca-GGUF/mistral-7b-openorca.Q4_K_M.gguf\")\n",
+        "llm = LLM(path=\"unsloth/Qwen3-4B-Instruct-2507-GGUF/Qwen3-4B-Instruct-2507-Q4_K_M.gguf\")\n",
         "\n",
         "# Question and context\n",
         "question = \"Write a list of invented electric-powered devices\"\n",
@@ -237,11 +235,11 @@
         "\n",
         "# Download models\n",
         "!wget https://huggingface.co/second-state/All-MiniLM-L6-v2-Embedding-GGUF/resolve/main/all-MiniLM-L6-v2-Q4_K_M.gguf\n",
-        "!wget https://huggingface.co/TheBloke/Mistral-7B-OpenOrca-GGUF/resolve/main/mistral-7b-openorca.Q4_K_M.gguf\n",
+        "!wget https://huggingface.co/unsloth/Qwen3-4B-Instruct-2507-GGUF/blob/main/Qwen3-4B-Instruct-2507-Q4_K_M.gguf\n",
         "\n",
         "# Start local API services\n",
         "!nohup python -m llama_cpp.server --n_gpu_layers -1 --model all-MiniLM-L6-v2-Q4_K_M.gguf --host 127.0.0.1 --port 8000 &> vector.log &\n",
-        "!nohup python -m llama_cpp.server --n_gpu_layers -1 --model mistral-7b-openorca.Q4_K_M.gguf --chat_format chatml --host 127.0.0.1 --port 8001 &> llm.log &\n",
+        "!nohup python -m llama_cpp.server --n_gpu_layers -1 --model Qwen3-4B-Instruct-2507-Q4_K_M.gguf --chat_format chatml --host 127.0.0.1 --port 8001 &> llm.log &\n",
         "!sleep 30"
       ]
     },

examples/63_How_RAG_with_txtai_works.ipynb

@@ -8,8 +8,6 @@
       "source": [
         "# How RAG with txtai works\n",
         "\n",
-        "[txtai](https://github.com/neuml/txtai) is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\n",
-        "\n",
         "Large Language Models (LLMs) have captured the public's attention with their impressive capabilities. The Generative AI era has reached a fever pitch with some predicting the coming rise of superintelligence.\n",
         "\n",
         "LLMs are far from perfect though and we're still a ways away from true AI. The biggest challenge is with hallucinations. Hallucinations is the term for when a LLM generates output that is factually incorrect. The alarming part of this is that on a cursory glance, it actually sounds like factual content. The default behavior of LLMs is to produce plausible answers even when no plausible answer exists. LLMs are not great at saying I don't know.\n",
@@ -45,7 +43,7 @@
       "outputs": [],
       "source": [
         "%%capture\n",
-        "!pip install git+https://github.com/neuml/txtai#egg=txtai[api,pipeline] autoawq"
+        "!pip install git+https://github.com/neuml/txtai#egg=txtai[api,pipeline]"
       ]
     },
     {
@@ -74,7 +72,7 @@
         "embeddings.load(provider=\"huggingface-hub\", container=\"neuml/txtai-wikipedia\")\n",
         "\n",
         "# Create LLM\n",
-        "llm = LLM(\"TheBloke/Mistral-7B-OpenOrca-AWQ\")"
+        "llm = LLM(\"Qwen/Qwen3-4B-Instruct-2507\")"
       ]
     },
     {
@@ -310,7 +308,7 @@
         "\n",
         "# RAG pipeline configuration\n",
         "rag:\n",
-        "  path: TheBloke/Mistral-7B-OpenOrca-AWQ\n",
+        "  path: Qwen/Qwen3-4B-Instruct-2507\n",
         "  output: flatten\n",
         "  template: |\n",
         "    <|im_start|>system\n",

examples/64_Embeddings_index_format_for_open_data_access.ipynb

@@ -21,8 +21,6 @@
       "source": [
         "# Embeddings index format for open data access\n",
         "\n",
-        "[txtai](https://github.com/neuml/txtai) is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\n",
-        "\n",
         "The main programming language with txtai is Python. A key tenet is that the underlying data in an embeddings index is accessible without txtai.\n",
         "\n",
         "This notebook will demonstrate this through a series of examples.\n"

examples/65_Speech_to_Speech_RAG.ipynb

@@ -6,8 +6,6 @@
    "source": [
     "# Speech to Speech RAG\n",
     "\n",
-    "[txtai](https://github.com/neuml/txtai) is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\n",
-    "\n",
     "There are many articles, notebooks and examples covering how to perform vector search and/or retrieval augmented generation (RAG) with txtai. A lesser known component of txtai is it's built-in workflow component.\n",
     "\n",
     "Workflows are a simple yet powerful construct that takes a callable and returns elements. Workflows enable efficient processing of pipeline data. Workflows are streaming by nature and work on data in batches. This allows large volumes of data to be processed efficiently.\n",
@@ -33,7 +31,7 @@
    "outputs": [],
    "source": [
     "%%capture\n",
-    "!pip install git+https://github.com/neuml/txtai#egg=txtai[pipeline-audio] autoawq"
+    "!pip install git+https://github.com/neuml/txtai#egg=txtai[pipeline-audio]"
    ]
   },
   {
@@ -86,7 +84,7 @@
     "# Create RAG pipeline\n",
     "rag = RAG(\n",
     "    embeddings,\n",
-    "    \"hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4\",\n",
+    "    \"Qwen/Qwen3-4B-Instruct-2507\",\n",
     "    system=\"You are a friendly assistant. You answer questions from users.\",\n",
     "    template=template,\n",
     "    context=10\n",
@@ -160,7 +158,7 @@
     "\n",
     "# RAG\n",
     "rag:\n",
-    "  path: \"hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4\"\n",
+    "  path: \"Qwen/Qwen3-4B-Instruct-2507\"\n",
     "  system: You are a friendly assistant. You answer questions from users.\n",
     "  template: |\n",
     "    Answer the following question using only the context below. Only include information\n",

examples/66_Generative_Audio.ipynb

@@ -6,8 +6,6 @@
    "source": [
     "# Generative Audio\n",
     "\n",
-    "[txtai](https://github.com/neuml/txtai) is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\n",
-    "\n",
     "txtai works with much more than just text! It has rich multimedia and multimodal capabilities.\n",
     "\n",
     "This notebook will demonstrate how to build generative audio workflows. These workflows will generate a combined audio stream with text and relevant audio for a series of poems."
@@ -29,7 +27,7 @@
    "outputs": [],
    "source": [
     "%%capture\n",
-    "!pip install git+https://github.com/neuml/txtai#egg=txtai[pipeline-audio] autoawq"
+    "!pip install git+https://github.com/neuml/txtai#egg=txtai[pipeline-audio]"
    ]
   },
   {
@@ -41,7 +39,7 @@
     "The next section defines a generative audio workflow. This workflow consists of a set of pipelines as follows:\n",
     "\n",
     "- LLM\n",
-    "  - Llama 3 model used to describe the emotions of a given story or poem\n",
+    "  - Model used to describe the emotions of a given story or poem\n",
     "- Text To Audio\n",
     "  - Builds audio given a text prompt\n",
     "- Text To Speech\n",
@@ -81,7 +79,7 @@
     "  return audio\n",
     "\n",
     "# LLM\n",
-    "llm = LLM(\"hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4\")\n",
+    "llm = LLM(\"Qwen/Qwen3-4B-Instruct-2507\")\n",
     "\n",
     "# Text to Audio\n",
     "# Important: The code for musicgen is licensed as MIT but model weights are CC-BY-NC\n",

examples/67_Whats_new_in_txtai_8_0.ipynb

@@ -8,8 +8,6 @@
    "source": [
     "# 💡 What's new in txtai 8.0\n",
     "\n",
-    "[txtai](https://github.com/neuml/txtai) is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\n",
-    "\n",
     "The 8.0 release brings a major new feature: Agents 🚀\n",
     "\n",
     "Agents automatically create workflows to answer multi-faceted user requests. Agents iteratively prompt and/or interface with tools to\n",
@@ -42,7 +40,7 @@
    "outputs": [],
    "source": [
     "%%capture\n",
-    "!pip install git+https://github.com/neuml/txtai autoawq model2vec"
+    "!pip install git+https://github.com/neuml/txtai#egg=txtai[agent] model2vec"
    ]
   },
   {
@@ -94,7 +92,7 @@
     "\n",
     "agent = Agent(\n",
     "    tools=[today, wikipedia, arxiv, \"websearch\"],\n",
-    "    llm=\"hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4\",\n",
+    "    llm=\"Qwen/Qwen3-4B-Instruct-2507\",\n",
     "    max_iterations=10,\n",
     ")"
    ]
@@ -254,7 +252,7 @@
    "source": [
     "Let's unpack what just happened here. The Agent reviewed multiple sources and aggregated the references into a single report. This is quite powerful 💪\n",
     "\n",
-    "Note that depending on the LLM used, errors can be seen as the Agent tries to get the function parameters right. This example is using Llama 3.1 8B. A more powerful LLM will likely lead to even better results. Remember that txtai supports a number of LLM frameworks (Hugging Face, llama.cpp and LiteLLM APIs)."
+    "Note that depending on the LLM used, errors can be seen as the Agent tries to get the function parameters right. This example is using Qwen3 4B. A more powerful LLM will likely lead to even better results. Remember that txtai supports a number of LLM frameworks (Hugging Face, llama.cpp and LiteLLM APIs)."
    ]
   },
   {
@@ -277,7 +275,7 @@
     "from txtai import Agent, LLM\n",
     "\n",
     "# Share the LLM instance across multiple agents\n",
-    "llm = LLM(\"hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4\")\n",
+    "llm = LLM(\"Qwen/Qwen3-4B-Instruct-2507\")\n",
     "\n",
     "websearcher = Agent(\n",
     "    tools=[\"websearch\"],\n",
@@ -418,7 +416,7 @@
     "              container: neuml/txtai-arxiv\n",
     "\n",
     "llm:\n",
-    "    path: hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4"
+    "    path: Qwen/Qwen3-4B-Instruct-2507"
    ]
   },
   {

examples/68_Analyzing_Hugging_Face_Posts_with_Graphs_and_Agents.ipynb

@@ -6,8 +6,6 @@
    "source": [
     "# Analyzing Hugging Face Posts with Graphs and Agents\n",
     "\n",
-    "[txtai](https://github.com/neuml/txtai) is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\n",
-    "\n",
     "Embeddings databases can optionally build a graph network alongside stored vectors. The graph automatically infers relationships of the content using the computed vector embeddings. This provides an additional way to explore and analyze content. \n",
     "\n",
     "txtai 8.0 was recently released and added the ability to run agents. Agents automatically create workflows to answer multi-faceted user requests.\n",
@@ -36,7 +34,7 @@
    "outputs": [],
    "source": [
     "%%capture\n",
-    "!pip install git+https://github.com/neuml/txtai#egg=txtai[graph] datasets autoawq"
+    "!pip install git+https://github.com/neuml/txtai#egg=txtai[agent,graph] datasets"
    ]
   },
   {
@@ -59,7 +57,7 @@
     "from datasets import load_dataset\n",
     "from txtai import LLM\n",
     "\n",
-    "llm = LLM(\"hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4\")\n",
+    "llm = LLM(\"Qwen/Qwen3-4B-Instruct-2507\")\n",
     "\n",
     "def title(text):\n",
     "    prompt = f\"\"\"\n",
@@ -815,7 +813,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.9.20"
+   "version": "3.10.16"
   }
  },
  "nbformat": 4,

examples/69_Granting_autonomy_to_agents.ipynb

@@ -6,8 +6,6 @@
    "source": [
     "# Granting autonomy to agents\n",
     "\n",
-    "[txtai](https://github.com/neuml/txtai) is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\n",
-    "\n",
     "txtai 8.0 was recently released and added the ability to run agents. Agents automatically create workflows to answer multi-faceted user requests.\n",
     "\n",
     "Agents connect a series of tools with a reasoning engine (i.e. LLM). We're giving the agent a degree of latitude to go through it's own internal logic to address a user's request.\n",
@@ -33,7 +31,7 @@
    "outputs": [],
    "source": [
     "%%capture\n",
-    "!pip install git+https://github.com/neuml/txtai#egg=txtai[graph] autoawq"
+    "!pip install git+https://github.com/neuml/txtai#egg=txtai[agent,graph]"
    ]
   },
   {
@@ -62,7 +60,7 @@
     "        \"provider\": \"huggingface-hub\",\n",
     "        \"container\": \"neuml/txtai-wikipedia\"\n",
     "    }],\n",
-    "    llm=\"hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4\",\n",
+    "    llm=\"Qwen/Qwen3-4B-Instruct-2507\",\n",
     "    max_iterations=10,\n",
     ")"
    ]
@@ -145,9 +143,7 @@
    "source": [
     "💥 Interesting indeed. The fundamental concept of search is we need to know what to look for. In this case, we didn't have that (i.e. we're bored 😀).\n",
     "\n",
-    "Let's go to another example. This time we'll look at the [txtai-hfposts](https://huggingface.co/NeuML/txtai-hfposts) embeddings database. We'll ask the agent to research a specific topic then write a report about what was found.\n",
-    "\n",
-    "It's important to mention that txtai agents support any [LLM supported by txtai](https://neuml.github.io/txtai/pipeline/text/llm/). So while all of the examples have been using `Llama 3.1 8B`, other LLMs local and remote are supported (i.e. OpenAI, AWS Bedrock, Anthropic). To change it up, we'll use one of the [Qwen models](https://huggingface.co/Qwen) which have been receiving a lot of positive reviews lately."
+    "Let's go to another example. This time we'll look at the [txtai-hfposts](https://huggingface.co/NeuML/txtai-hfposts) embeddings database. We'll ask the agent to research a specific topic then write a report about what was found.\n"
    ]
   },
   {
@@ -163,7 +159,7 @@
     "        \"provider\": \"huggingface-hub\",\n",
     "        \"container\": \"neuml/txtai-hfposts\"\n",
     "    }],\n",
-    "    llm=\"Qwen/Qwen2.5-7B-Instruct-AWQ\",\n",
+    "    llm=\"Qwen/Qwen3-4B-Instruct-2507\",\n",
     "    max_iterations=10,\n",
     ")"
    ]
@@ -293,7 +289,7 @@
     "\n",
     "agent = Agent(\n",
     "    tools=[insert, embeddings.search, \"websearch\"],\n",
-    "    llm=\"Qwen/Qwen2.5-7B-Instruct-AWQ\",\n",
+    "    llm=\"Qwen/Qwen3-4B-Instruct-2507\",\n",
     "    max_iterations=10\n",
     ")"
    ]

examples/71_Analyzing_LinkedIn_Company_Posts_with_Graphs_and_Agents.ipynb

@@ -6,8 +6,6 @@
    "source": [
     "# Analyzing LinkedIn Company Posts with Graphs and Agents\n",
     "\n",
-    "[txtai](https://github.com/neuml/txtai) is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\n",
-    "\n",
     "This notebook will analyze [NeuML's LinkedIn company posts](https://hf.co/datasets/neuml/neuml-linkedin-202501) over the last 12 months as of January 2025. It will build an embeddings database with an associated graph and topic modeling. The most popular topics will be explored using graphs and vector search queries.\n",
     "\n",
     "From there we'll use Agents to investigate the dataset and find ways to increase engagement with our future posts using past history."
@@ -24,12 +22,12 @@
   },
   {
    "cell_type": "code",
-   "execution_count": 1,
+   "execution_count": null,
    "metadata": {},
    "outputs": [],
    "source": [
     "%%capture\n",
-    "!pip install git+https://github.com/neuml/txtai#egg=txtai[graph] datasets autoawq autoawq-kernels"
+    "!pip install git+https://github.com/neuml/txtai#egg=txtai[agent,graph] datasets"
    ]
   },
   {
@@ -74,7 +72,7 @@
     "\n",
     "posts = load_dataset(\"neuml/neuml-linkedin-202501\", split=\"train\")\n",
     "\n",
-    "llm = LLM(\"hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4\")\n",
+    "llm = LLM(\"Qwen/Qwen3-4B-Instruct-2507\")\n",
     "\n",
     "embeddings = Embeddings(\n",
     "    autoid=\"uuid5\",\n",

examples/72_Parsing_the_stars_with_txtai.ipynb

@@ -6,8 +6,6 @@
    "source": [
     "# Parsing the ⭐'s with txtai\n",
     "\n",
-    "[txtai](https://github.com/neuml/txtai) is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\n",
-    "\n",
     "In honor of txtai's recent milestone of reaching 10K ⭐'s on GitHub, this notebook will build an astronomical knowledge graph of known stars, planets, galaxies and other astronomical entities. It will show how a compendium of information can be built from a public knowledge source such as Wikipedia.\n",
     "\n",
     "While agents and retrieval augmented generation (RAG) can do great things, they can do even greater things if provided with an assist from the humans in the form of compiled knowledge. It's not only better accuracy, it's also faster as we don't need to wait for an agent to figure out what we already know.\n",
@@ -31,7 +29,7 @@
    "outputs": [],
    "source": [
     "%%capture\n",
-    "!pip install git+https://github.com/neuml/txtai#egg=txtai[graph,pipeline-text] datasets autoawq autoawq-kernels"
+    "!pip install git+https://github.com/neuml/txtai#egg=txtai[agent,graph,pipeline-text] datasets"
    ]
   },
   {
@@ -401,7 +399,7 @@
     "context: {context}\"\"\"\n",
     "\n",
     "# Load LLM\n",
-    "llm = LLM(\"hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4\")\n",
+    "llm = LLM(\"Qwen/Qwen3-4B-Instruct-2507\")\n",
     "\n",
     "# Create RAG pipeline\n",
     "rag = RAG(\n",

examples/73_Chunking_your_data_for_RAG.ipynb

@@ -6,8 +6,6 @@
    "source": [
     "# Chunking your data for RAG\n",
     "\n",
-    "[txtai](https://github.com/neuml/txtai) is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\n",
-    "\n",
     "One of the major workflows in `txtai` is Retrieval Augmented Generation (RAG). Large Language Models (LLM) are built to generate coherent sounding text. While in many cases it is factually accurate, that is not what they're built to do. RAG steps in to help inject smaller pieces of knowledge into a LLM prompt and increase the overall accuracy of responses. The `R` in RAG is very important.\n",
     "\n",
     "This notebook will demonstrate how to extract, chunk and index text to support retrieval operations for RAG."
@@ -40,7 +38,7 @@
     "\n",
     "Let's dive right in and keep this example simple. The next section creates a [Textractor pipeline](https://neuml.github.io/txtai/pipeline/data/textractor/) and an [Embeddings database](https://neuml.github.io/txtai/embeddings/).\n",
     "\n",
-    "The `Textractor` extracts chunks of text from files and the `Embeddings` takes those chunks and builds an index/database. We'll use a [late chunker](https://docs.chonkie.ai/chunkers/late-chunker) backed by [Chonkie](https://github.com/chonkie-ai/chonkie).\n",
+    "The `Textractor` extracts chunks of text from files and the `Embeddings` takes those chunks and builds an index/database. We'll use a [late chunker](https://docs.chonkie.ai/chunkers/late-chunker) backed by [Chonkie](https://github.com/chonkie-inc/chonkie).\n",
     "\n",
     "Then, we'll build an indexing workflow that streams chunks from two files."
    ]
@@ -102,12 +100,12 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "{'indexid': 0, 'id': '0', 'url': 'https://github.com/neuml/txtai', 'text': '**GitHub - neuml/txtai: 💡 All-in-one open-source embeddings database for semantic search, LLM orchestration and language model workflows**\\n\\n*💡 All-in-one open-source embeddings database for semantic search, LLM orchestration and language model workflows - neuml/txtai*\\n\\n\\n\\n**All-in-one embeddings database** \\ntxtai is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\\n\\nEmbeddings databases are a union of vector indexes (sparse and dense), graph networks and relational databases.\\n\\nThis foundation enables vector search and/or serves as a powerful knowledge source for large language model (LLM) applications.\\n\\nBuild autonomous agents, retrieval augmented generation (RAG) processes, multi-model workflows and more.\\n\\nSummary of txtai features:\\n\\n- 🔎 Vector search with SQL, object storage, topic modeling, graph analysis and multimodal indexing\\n- 📄 Create embeddings for text, documents, audio, images and video\\n- 💡 Pipelines powered by language models that run LLM prompts, question-answering, labeling, transcription, translation, summarization and more\\n- ↪️️ Workflows to join pipelines together and aggregate business logic. txtai processes can be simple microservices or multi-model workflows.\\n- 🤖 Agents that intelligently connect embeddings, pipelines, workflows and other agents together to autonomously solve complex problems\\n- ⚙️ Build with Python or YAML. API bindings available for [JavaScript](https://github.com/neuml/txtai.js) , [Java](https://github.com/neuml/txtai.java) , [Rust](https://github.com/neuml/txtai.rs) and [Go](https://github.com/neuml/txtai.go) .\\n- 🔋 Batteries included with defaults to get up and running fast\\n- ☁️ Run local or scale out with container orchestration\\ntxtai is built with Python 3.9+, [Hugging Face Transformers](https://github.com/huggingface/transformers) , [Sentence Transformers](https://github.'}\n",
+      "{'indexid': 0, 'id': '0', 'url': 'https://github.com/neuml/txtai', 'text': '**GitHub - neuml/txtai: 💡 All-in-one open-source embeddings database for semantic search, LLM orchestration and language model workflows**\\n\\n*💡 All-in-one open-source embeddings database for semantic search, LLM orchestration and language model workflows - neuml/txtai*\\n\\n\\n\\n**All-in-one embeddings database** \\ntxtai is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\\n\\nEmbeddings databases are a union of vector indexes (sparse and dense), graph networks and relational databases.\\n\\nThis foundation enables vector search and/or serves as a powerful knowledge source for large language model (LLM) applications.\\n\\nBuild autonomous agents, retrieval augmented generation (RAG) processes, multi-model workflows and more.\\n\\nSummary of txtai features:\\n\\n- 🔎 Vector search with SQL, object storage, topic modeling, graph analysis and multimodal indexing\\n- 📄 Create embeddings for text, documents, audio, images and video\\n- 💡 Pipelines powered by language models that run LLM prompts, question-answering, labeling, transcription, translation, summarization and more\\n- ↪️️ Workflows to join pipelines together and aggregate business logic. txtai processes can be simple microservices or multi-model workflows.\\n- 🤖 Agents that intelligently connect embeddings, pipelines, workflows and other agents together to autonomously solve complex problems\\n- ⚙️ Build with Python or YAML. API bindings available for [JavaScript](https://github.com/neuml/txtai.js) , [Java](https://github.com/neuml/txtai.java) , [Rust](https://github.com/neuml/txtai.rs) and [Go](https://github.com/neuml/txtai.go) .\\n- 🔋 Batteries included with defaults to get up and running fast\\n- ☁️ Run local or scale out with container orchestration\\ntxtai is built with Python 3.11+, [Hugging Face Transformers](https://github.com/huggingface/transformers) , [Sentence Transformers](https://github.'}\n",
       "{'indexid': 1, 'id': '0', 'url': 'https://github.com/neuml/txtai', 'text': 'com/UKPLab/sentence-transformers) and [FastAPI](https://github.com/tiangolo/fastapi) . txtai is open-source under an Apache 2.0 license.\\n\\n*Interested in an easy and secure way to run hosted txtai applications? Then join the* [txtai.cloud](https://txtai.cloud) *preview to learn more.* \\n\\n## Why txtai?\\nNew vector databases, LLM frameworks and everything in between are sprouting up daily. Why build with txtai?\\n\\n- Up and running in minutes with [pip](https://neuml.github.io/txtai/install/) or [Docker](https://neuml.github.io/txtai/cloud/) \\n```\\n# Get started in a couple lines\\nimport txtai\\n\\nembeddings = txtai.Embeddings()\\nembeddings.index([\"Correct\", \"Not what we hoped\"])\\nembeddings.search(\"positive\", 1)\\n#[(0, 0.29862046241760254)]\\n```\\n\\n- Built-in API makes it easy to develop applications using your programming language of choice\\n```\\n# app.yml\\nembeddings:\\n path: sentence-transformers/all-MiniLM-L6-v2\\n```\\n\\n```\\nCONFIG=app.yml uvicorn \"txtai.api:app\"\\ncurl -X GET \"http://localhost:8000/search?query=positive\"\\n```\\n\\n- Run local - no need to ship data off to disparate remote services\\n- Work with micromodels all the way up to large language models (LLMs)\\n- Low footprint - install additional dependencies and scale up when needed\\n- [Learn by example](https://neuml.github.io/txtai/examples) - notebooks cover all available functionality\\n\\n## Use Cases\\nThe following sections introduce common txtai use cases. A comprehensive set of over 60 [example notebooks and applications](https://neuml.github.io/txtai/examples) are also available.\\n\\n\\n### Semantic Search\\nBuild semantic/similarity/vector/neural search applications.'}\n",
       "{'indexid': 2, 'id': '0', 'url': 'https://github.com/neuml/txtai', 'text': 'Traditional search systems use keywords to find data. Semantic search has an understanding of natural language and identifies results that have the same meaning, not necessarily the same keywords.\\n\\nGet started with the following examples.\\n\\n|Notebook|Description||\\n|---|---|---|\\n|[Introducing txtai](https://github.com/neuml/txtai/blob/master/examples/01_Introducing_txtai.ipynb) |Overview of the functionality provided by txtai||\\n|[Similarity search with images](https://github.com/neuml/txtai/blob/master/examples/13_Similarity_search_with_images.ipynb) |Embed images and text into the same space for search||\\n|[Build a QA database](https://github.com/neuml/txtai/blob/master/examples/34_Build_a_QA_database.ipynb) |Question matching with semantic search||\\n|[Semantic Graphs](https://github.com/neuml/txtai/blob/master/examples/38_Introducing_the_Semantic_Graph.ipynb) |Explore topics, data connectivity and run network analysis||\\n\\n### LLM Orchestration\\nAutonomous agents, retrieval augmented generation (RAG), chat with your data, pipelines and workflows that interface with large language models (LLMs).\\n\\nSee below to learn more.\\n\\n|Notebook|Description||\\n|---|---|---|\\n|[Prompt templates and task chains](https://github.com/neuml/txtai/blob/master/examples/44_Prompt_templates_and_task_chains.ipynb) |Build model prompts and connect tasks together with workflows||\\n|[Integrate LLM frameworks](https://github.com/neuml/txtai/blob/master/examples/53_Integrate_LLM_Frameworks.ipynb) |Integrate llama.cpp, LiteLLM and custom generation frameworks||\\n|[Build knowledge graphs with LLMs](https://github.'}\n",
       "{'indexid': 3, 'id': '0', 'url': 'https://github.com/neuml/txtai', 'text': 'com/neuml/txtai/blob/master/examples/57_Build_knowledge_graphs_with_LLM_driven_entity_extraction.ipynb) |Build knowledge graphs with LLM-driven entity extraction||\\n\\n#### Agents\\nAgents connect embeddings, pipelines, workflows and other agents together to autonomously solve complex problems.\\n\\ntxtai agents are built on top of the Transformers Agent framework. This supports all LLMs txtai supports (Hugging Face, llama.cpp, OpenAI / Claude / AWS Bedrock via LiteLLM).\\n\\nSee the link below to learn more.\\n\\n|Notebook|Description||\\n|---|---|---|\\n|[Analyzing Hugging Face Posts with Graphs and Agents](https://github.com/neuml/txtai/blob/master/examples/68_Analyzing_Hugging_Face_Posts_with_Graphs_and_Agents.ipynb) |Explore a rich dataset with Graph Analysis and Agents||\\n|[Granting autonomy to agents](https://github.com/neuml/txtai/blob/master/examples/69_Granting_autonomy_to_agents.ipynb) |Agents that iteratively solve problems as they see fit||\\n|[Analyzing LinkedIn Company Posts with Graphs and Agents](https://github.com/neuml/txtai/blob/master/examples/71_Analyzing_LinkedIn_Company_Posts_with_Graphs_and_Agents.ipynb) |Exploring how to improve social media engagement with AI||\\n\\n#### Retrieval augmented generation\\nRetrieval augmented generation (RAG) reduces the risk of LLM hallucinations by constraining the output with a knowledge base as context. RAG is commonly used to \"chat with your data\".\\n\\nA novel feature of txtai is that it can provide both an answer and source citation.\\n\\n|Notebook|Description||\\n|---|---|---|\\n|[Build RAG pipelines with txtai](https://github.'}\n",
       "{'indexid': 4, 'id': '0', 'url': 'https://github.com/neuml/txtai', 'text': 'com/neuml/txtai/blob/master/examples/52_Build_RAG_pipelines_with_txtai.ipynb) |Guide on retrieval augmented generation including how to create citations||\\n|[How RAG with txtai works](https://github.com/neuml/txtai/blob/master/examples/63_How_RAG_with_txtai_works.ipynb) |Create RAG processes, API services and Docker instances||\\n|[Advanced RAG with graph path traversal](https://github.com/neuml/txtai/blob/master/examples/58_Advanced_RAG_with_graph_path_traversal.ipynb) |Graph path traversal to collect complex sets of data for advanced RAG||\\n|[Speech to Speech RAG](https://github.com/neuml/txtai/blob/master/examples/65_Speech_to_Speech_RAG.ipynb) |Full cycle speech to speech workflow with RAG||\\n\\n### Language Model Workflows\\nLanguage model workflows, also known as semantic workflows, connect language models together to build intelligent applications.\\n\\nWhile LLMs are powerful, there are plenty of smaller, more specialized models that work better and faster for specific tasks. This includes models for extractive question-answering, automatic summarization, text-to-speech, transcription and translation.\\n\\n|Notebook|Description||\\n|---|---|---|\\n|[Run pipeline workflows](https://github.com/neuml/txtai/blob/master/examples/14_Run_pipeline_workflows.ipynb) |Simple yet powerful constructs to efficiently process data||\\n|[Building abstractive text summaries](https://github.com/neuml/txtai/blob/master/examples/09_Building_abstractive_text_summaries.ipynb) |Run abstractive text summarization||\\n|[Transcribe audio to text](https://github.com/neuml/txtai/blob/master/examples/11_Transcribe_audio_to_text.'}\n",
-      "{'indexid': 5, 'id': '0', 'url': 'https://github.com/neuml/txtai', 'text': 'ipynb) |Convert audio files to text||\\n|[Translate text between languages](https://github.com/neuml/txtai/blob/master/examples/12_Translate_text_between_languages.ipynb) |Streamline machine translation and language detection||\\n\\n## Installation\\nThe easiest way to install is via pip and PyPI\\n\\n```\\npip install txtai\\n```\\n\\nPython 3.9+ is supported. Using a Python [virtual environment](https://docs.python.org/3/library/venv.html) is recommended.\\n\\nSee the detailed [install instructions](https://neuml.github.io/txtai/install) for more information covering [optional dependencies](https://neuml.github.io/txtai/install/#optional-dependencies) , [environment specific prerequisites](https://neuml.github.io/txtai/install/#environment-specific-prerequisites) , [installing from source](https://neuml.github.io/txtai/install/#install-from-source) , [conda support](https://neuml.github.io/txtai/install/#conda) and how to [run with containers](https://neuml.github.io/txtai/cloud) .\\n\\n\\n## Model guide\\nSee the table below for the current recommended models. These models all allow commercial use and offer a blend of speed and performance.\\n\\n|Component|Model(s)|\\n|---|---|\\n|[Embeddings](https://neuml.github.io/txtai/embeddings) |[all-MiniLM-L6-v2](https://hf.co/sentence-transformers/all-MiniLM-L6-v2) |\\n|[Image Captions](https://neuml.github.io/txtai/pipeline/image/caption) |[BLIP](https://hf.co/Salesforce/blip-image-captioning-base) |'}\n",
+      "{'indexid': 5, 'id': '0', 'url': 'https://github.com/neuml/txtai', 'text': 'ipynb) |Convert audio files to text||\\n|[Translate text between languages](https://github.com/neuml/txtai/blob/master/examples/12_Translate_text_between_languages.ipynb) |Streamline machine translation and language detection||\\n\\n## Installation\\nThe easiest way to install is via pip and PyPI\\n\\n```\\npip install txtai\\n```\\n\\nPython 3.11+ is supported. Using a Python [virtual environment](https://docs.python.org/3/library/venv.html) is recommended.\\n\\nSee the detailed [install instructions](https://neuml.github.io/txtai/install) for more information covering [optional dependencies](https://neuml.github.io/txtai/install/#optional-dependencies) , [environment specific prerequisites](https://neuml.github.io/txtai/install/#environment-specific-prerequisites) , [installing from source](https://neuml.github.io/txtai/install/#install-from-source) , [conda support](https://neuml.github.io/txtai/install/#conda) and how to [run with containers](https://neuml.github.io/txtai/cloud) .\\n\\n\\n## Model guide\\nSee the table below for the current recommended models. These models all allow commercial use and offer a blend of speed and performance.\\n\\n|Component|Model(s)|\\n|---|---|\\n|[Embeddings](https://neuml.github.io/txtai/embeddings) |[all-MiniLM-L6-v2](https://hf.co/sentence-transformers/all-MiniLM-L6-v2) |\\n|[Image Captions](https://neuml.github.io/txtai/pipeline/image/caption) |[BLIP](https://hf.co/Salesforce/blip-image-captioning-base) |'}\n",
       "{'indexid': 6, 'id': '0', 'url': 'https://github.com/neuml/txtai', 'text': '|[Labels - Zero Shot](https://neuml.github.io/txtai/pipeline/text/labels) |[BART-Large-MNLI](https://hf.co/facebook/bart-large) |\\n|[Labels - Fixed](https://neuml.github.io/txtai/pipeline/text/labels) |Fine-tune with [training pipeline](https://neuml.github.io/txtai/pipeline/train/trainer) |\\n|[Large Language Model (LLM)](https://neuml.github.io/txtai/pipeline/text/llm) |[Llama 3.1 Instruct](https://hf.co/meta-llama/Llama-3.1-8B-Instruct) |\\n|[Summarization](https://neuml.github.io/txtai/pipeline/text/summary) |[DistilBART](https://hf.co/sshleifer/distilbart-cnn-12-6) |\\n|[Text-to-Speech](https://neuml.github.io/txtai/pipeline/audio/texttospeech) |[ESPnet JETS](https://hf.co/NeuML/ljspeech-jets-onnx) |\\n|[Transcription](https://neuml.github.io/txtai/pipeline/audio/transcription) |[Whisper](https://hf.co/openai/whisper-base) |\\n|[Translation](https://neuml.github.io/txtai/pipeline/text/translation) |[OPUS Model Series](https://hf.co/Helsinki-NLP) |\\nModels can be loaded as either a path from the Hugging Face Hub or a local directory. Model paths are optional, defaults are loaded when not specified. For tasks with no recommended model, txtai uses the default models as shown in the Hugging Face Tasks guide.\\n\\nSee the following links to learn more.\\n\\n\\n## Powered by txtai\\nThe following applications are powered by txtai.'}\n",
       "{'indexid': 7, 'id': '0', 'url': 'https://github.com/neuml/txtai', 'text': \"|Application|Description|\\n|---|---|\\n|[rag](https://github.com/neuml/rag) |Retrieval Augmented Generation (RAG) application|\\n|[ragdata](https://github.com/neuml/ragdata) |Build knowledge bases for RAG|\\n|[paperai](https://github.com/neuml/paperai) |Semantic search and workflows for medical/scientific papers|\\n|[annotateai](https://github.com/neuml/annotateai) |Automatically annotate papers with LLMs|\\nIn addition to this list, there are also many other [open-source projects](https://github.com/neuml/txtai/network/dependents) , [published research](https://scholar.google.com/scholar?q=txtai&hl=en&as_ylo=2022) and closed proprietary/commercial projects that have built on txtai in production.\\n\\n\\n## Further Reading\\n- [Tutorial series on Hashnode](https://neuml.hashnode.dev/series/txtai-tutorial) | [dev.to](https://dev.to/neuml/tutorial-series-on-txtai-ibg) \\n- [What's new in txtai 8.0](https://medium.com/neuml/whats-new-in-txtai-8-0-2d7d0ab4506b) | [7.0](https://medium.com/neuml/whats-new-in-txtai-7-0-855ad6a55440) | [6.0](https://medium.com/neuml/whats-new-in-txtai-6-0-7d93eeedf804) | [5.0](https://medium.com/neuml/whats-new-in-txtai-5-0-e5c75a13b101) | [4.0](https://medium.\"}\n",
       "{'indexid': 8, 'id': '0', 'url': 'https://github.com/neuml/txtai', 'text': 'com/neuml/whats-new-in-txtai-4-0-bbc3a65c3d1c) \\n- [Getting started with semantic search](https://medium.com/neuml/getting-started-with-semantic-search-a9fd9d8a48cf) | [workflows](https://medium.com/neuml/getting-started-with-semantic-workflows-2fefda6165d9) | [rag](https://medium.com/neuml/getting-started-with-rag-9a0cca75f748) \\n\\n## Documentation\\n[Full documentation on txtai](https://neuml.github.io/txtai) including configuration settings for embeddings, pipelines, workflows, API and a FAQ with common questions/issues is available.\\n\\n\\n## Contributing\\nFor those who would like to contribute to txtai, please see [this guide](https://github.com/neuml/.github/blob/master/CONTRIBUTING.md) .'}\n",

examples/74_OpenAI_Compatible_API.ipynb

@@ -6,8 +6,6 @@
    "source": [
     "# OpenAI Compatible API\n",
     "\n",
-    "[txtai](https://github.com/neuml/txtai) is an all-in-one embeddings database for semantic search, LLM orchestration and language model workflows.\n",
-    "\n",
     "`txtai` has long been able host a [FastAPI based service](https://neuml.github.io/txtai/api/). There are clients for [Python](https://github.com/neuml/txtai.py), [JavaScript](https://github.com/neuml/txtai.js), [Java](https://github.com/neuml/txtai.java), [Rust](https://github.com/neuml/txtai.rs), [Go](https://github.com/neuml/txtai.go).\n",
     "\n",
     "The API service also supports hosting OpenAI-compatible API endpoints. A standard OpenAI client can then be used to connect to a `txtai` service. This enables quickly trying `txtai` with a familiar-to-use client. It's also a way to do local/offline development testing using the OpenAI client.\n",
@@ -41,7 +39,7 @@
     "\n",
     "# RAG pipeline configuration\n",
     "rag:\n",
-    "  path: hugging-quants/Meta-Llama-3.1-8B-Instruct-AWQ-INT4\n",
+    "  path: Qwen/Qwen3-4B-Instruct-2507\n",
     "  output: flatten\n",
     "  system: You are a friendly assistant. You answer questions from users.\n",
     "  template: |\n",
@@ -70,7 +68,7 @@
     "Alternatively, txtai can be directly installed and run as follows:\n",
     "\n",
     "```\n",
-    "pip install txtai[all] autoawq autoawq-kernels\n",
+    "pip install txtai[all]\n",
     "CONFIG=/tmp/config/config.yml uvicorn \"txtai.api:app\"\n",
     "```\n",
     "\n",

examples/75_Medical_RAG_Research_with_txtai.ipynb

@@ -0,0 +1,454 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "666da745",
+   "metadata": {},
+   "source": [
+    "# Medical RAG Research with txtai\n",
+    "\n",
+    "Large Language Models (LLMs) have captured the public's attention with their impressive capabilities. The Generative AI era has reached a fever pitch with some predicting the coming rise of superintelligence.\n",
+    "\n",
+    "LLMs are far from perfect though and we're still a ways away from true AI. One big challenge is with hallucinations. Hallucinations is the term for when an LLM generates output that is factually incorrect. The alarming part of this is that on a cursory glance, it actually sounds like factual content. The default behavior of LLMs is to produce plausible answers even when no plausible answer exists. LLMs are not great at saying I don't know.\n",
+    "\n",
+    "Retrieval Augmented Generation (RAG) helps reduce the risk of hallucinations by limiting the context in which a LLM can generate answers. This is typically done with a search query that hydrates a prompt with a relevant context. RAG has been one of the most practical use cases of the Generative AI era.\n",
+    "\n",
+    "This notebook will demonstrate how to build a Medical RAG Research process with txtai."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "a4c7cc15",
+   "metadata": {},
+   "source": [
+    "# Install dependencies\n",
+    "\n",
+    "Install `txtai` and all dependencies."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "a246bb3a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%%capture\n",
+    "!pip install git+https://github.com/neuml/txtai"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "fc5c27ed",
+   "metadata": {},
+   "source": [
+    "# Medical Dataset\n",
+    "\n",
+    "For this example, we'll use a [PubMed subset of article metadata for H5N1](https://huggingface.co/datasets/NeuML/pubmed-h5n1). This dataset was created using [`paperetl`](https://github.com/neuml/paperetl), an open-source library for parsing medical and scientific papers.\n",
+    "\n",
+    "[PubMed](https://pubmed.ncbi.nlm.nih.gov/) has over 38 million article abstracts as of June 2025. `paperetl` supports loading the full dataset with all 38 million articles or just a smaller subset. The dataset link above has more details on how this can be changed for different codes and keywords. This link also has information on how the article abstracts can be loaded in addition to the metadata."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 80,
+   "id": "5dd4145d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from datasets import load_dataset\n",
+    "from txtai import Embeddings\n",
+    "\n",
+    "ds = load_dataset(\"neuml/pubmed-h5n1\", split=\"train\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "2d3b1cd8",
+   "metadata": {},
+   "source": [
+    "Next, we'll build a `txtai` embeddings index with the articles. We'll use a vector embeddings model that specializes in vectorizing medical papers: [PubMedBERT Embeddings](https://huggingface.co/NeuML/pubmedbert-base-embeddings). "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 81,
+   "id": "04e69829",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "embeddings = Embeddings(path=\"neuml/pubmedbert-base-embeddings\", content=True, columns={\"text\": \"title\"})\n",
+    "embeddings.index(x for x in ds if x[\"title\"])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 82,
+   "id": "a801391e",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "7865"
+      ]
+     },
+     "execution_count": 82,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "embeddings.count()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "907898a4",
+   "metadata": {},
+   "source": [
+    "# RAG Pipeline\n",
+    "\n",
+    "There are a number of [prior examples](https://neuml.github.io/txtai/examples/#llm) on how to run RAG with `txtai`. The [RAG pipeline](https://neuml.github.io/txtai/pipeline/text/rag/) takes two main parameters, an embeddings database and an LLM. The embeddings database is the one just created above. For this example, we'll use a [simple local LLM with 600M parameters](https://huggingface.co/Qwen/Qwen3-0.6B).\n",
+    "\n",
+    "Substitute your own embeddings database to change the knowledge base. `txtai` supports running local LLMs via [transformers](https://github.com/huggingface/transformers) or [llama.cpp](https://github.com/abetlen/llama-cpp-python). It also supports a wide variety of LLMs via [LiteLLM](https://github.com/BerriAI/litellm). For example, setting the 2nd RAG pipeline parameter below to `gpt-4o` along with the appropriate environment variables with access keys switches to a hosted LLM. See [this documentation page](https://neuml.github.io/txtai/pipeline/text/llm/) for more on this."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 83,
+   "id": "cd05fc41",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Device set to use cuda:0\n"
+     ]
+    }
+   ],
+   "source": [
+    "from txtai import RAG\n",
+    "\n",
+    "# Prompt templates\n",
+    "system = \"You are a friendly medical assistant that answers questions\"\n",
+    "template = \"\"\"\n",
+    "Answer the following question using the provided context.\n",
+    "\n",
+    "Question:\n",
+    "{question}\n",
+    "\n",
+    "Context:\n",
+    "{context}\n",
+    "\"\"\"\n",
+    "\n",
+    "# Create RAG pipeline\n",
+    "rag = RAG(embeddings, \"Qwen/Qwen3-0.6B\", system=system, template=template, output=\"flatten\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "98800220",
+   "metadata": {},
+   "source": [
+    "# RAG Queries\n",
+    "\n",
+    "Now that the pipeline is setup, let's run a query."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 85,
+   "id": "4d93ac4b",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "<think>\n",
+      "Okay, let's see. The user is asking about H5N1. The context provided starts with \"Why tell me now?\" and then goes into facts about H5N1. The first sentence mentions that people and healthcare providers are weighing in on pandemic messages. Then it says H5N1 is avian influenza, a potential pandemic.\n",
+      "\n",
+      "Wait, but the user's question is about H5N1. The context doesn't go into specifics about what H5N1 is, but it does state that it's avian influenza. So I need to make sure I answer based on that. The answer should be concise, maybe mention that H5N1 is avian flu and it's a potential pandemic. Also, note that people are weighing in on messages. But I need to check if there's any more information. The context ends there. So the answer should be straightforward.\n",
+      "</think>\n",
+      "\n",
+      "H5N1 influenza viruses are a type of avian influenza, a potential pandemic influenza virus that could cause widespread illness and death. While the context highlights the importance of public health and preparedness, it does not provide more specific details about its characteristics or risks.\n"
+     ]
+    }
+   ],
+   "source": [
+    "print(rag(\"Tell me about H5N1\"))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "39ab9946",
+   "metadata": {},
+   "source": [
+    "Notice that this LLM outputs a thinking or reasoning section then the answer.\n",
+    "\n",
+    "Let's review the context to validate this answer is derived from the knowledge base."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 79,
+   "id": "53de0f96",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "[{'id': '16775537',\n",
+       "  'text': '\"Why tell me now?\" the public and healthcare providers weigh in on pandemic influenza messages.',\n",
+       "  'score': 0.7156285643577576},\n",
+       " {'id': '22308474',\n",
+       "  'text': 'H5N1 influenza viruses: facts, not fear.',\n",
+       "  'score': 0.658343493938446},\n",
+       " {'id': '16440117',\n",
+       "  'text': 'Avian influenza--a pandemic waiting to happen?',\n",
+       "  'score': 0.5827972888946533},\n",
+       " {'id': '20667302',\n",
+       "  'text': 'The influenza A(H5N1) epidemic at six and a half years: 500 notified human cases and more to come.',\n",
+       "  'score': 0.5593500137329102},\n",
+       " {'id': '18936262',\n",
+       "  'text': 'What Australians know and believe about bird flu: results of a population telephone survey.',\n",
+       "  'score': 0.5568690299987793},\n",
+       " {'id': '30349811',\n",
+       "  'text': 'Back to the Future: Lessons Learned From the 1918 Influenza Pandemic.',\n",
+       "  'score': 0.5540266036987305},\n",
+       " {'id': '17276785',\n",
+       "  'text': 'Pandemic influenza: what infection control professionals should know.',\n",
+       "  'score': 0.5519200563430786},\n",
+       " {'id': '16681227',\n",
+       "  'text': 'A pandemic flu: not if, but when. SARS was the wake-up call we slept through.',\n",
+       "  'score': 0.5518345832824707},\n",
+       " {'id': '22402712',\n",
+       "  'text': 'Ferretting out the facts behind the H5N1 controversy.',\n",
+       "  'score': 0.5508109331130981},\n",
+       " {'id': '25546511',\n",
+       "  'text': \"One-way trip: influenza virus' adaptation to gallinaceous poultry may limit its pandemic potential.\",\n",
+       "  'score': 0.5494509339332581}]"
+      ]
+     },
+     "execution_count": 79,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "embeddings.search(\"Tell me about H5N1\", limit=10)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "025e6092",
+   "metadata": {},
+   "source": [
+    "The answer is doing a good job being based on the context above. Also keep in mind this is a small 600M parameter model, which is even more impressive.\n",
+    "\n",
+    "Let's try another query."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 86,
+   "id": "1899047a",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "<think>\n",
+      "Okay, let's see. The user is asking about the locations that have had H5N1 outbreaks, and the provided context mentions a few places: Indonesia and Bangladesh. The context also has a title about a decade of avian influenza in Bangladesh and mentions \"H5N1.\" \n",
+      "\n",
+      "Wait, the user's question is in English, so I need to make sure I'm interpreting the context correctly. The context includes two sentences: one about a decade in Bangladesh and another about H5N1. The user is probably looking for specific locations where H5N1 has been reported. \n",
+      "\n",
+      "Looking at the context again, it says \"Human avian influenza in Indonesia\" and \"A Decade of Avian Influenza in Bangladesh: Where Are We Now? Are we ready for pandemic influenza H5N1?\" So the outbreaks are in Indonesia and Bangladesh. \n",
+      "\n",
+      "I should confirm that there are no other mentions of other locations. The context doesn't provide more information beyond those two countries. Therefore, the answer should list Indonesia and Bangladesh as the locations with H5N1 outbreaks.\n",
+      "</think>\n",
+      "\n",
+      "The locations with H5N1 outbreaks are Indonesia and Bangladesh.\n"
+     ]
+    }
+   ],
+   "source": [
+    "print(rag(\"What locations have had H5N1 outbreaks?\"))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 87,
+   "id": "39d4efa6",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "[{'id': '21706937',\n",
+       "  'text': 'Human avian influenza in Indonesia: are they really clustered?',\n",
+       "  'score': 0.6269429326057434},\n",
+       " {'id': '31514405',\n",
+       "  'text': 'A Decade of Avian Influenza in Bangladesh: Where Are We Now?',\n",
+       "  'score': 0.5972536206245422},\n",
+       " {'id': '15889987',\n",
+       "  'text': 'Are we ready for pandemic influenza H5N1?',\n",
+       "  'score': 0.5863772630691528},\n",
+       " {'id': '17717543',\n",
+       "  'text': 'Commentary: From scarcity to abundance: pandemic vaccines and other agents for \"have not\" countries.',\n",
+       "  'score': 0.5844159126281738},\n",
+       " {'id': '22491771',\n",
+       "  'text': 'Two years after pandemic influenza A/2009/H1N1: what have we learned?',\n",
+       "  'score': 0.5812581777572632},\n",
+       " {'id': '39666804',\n",
+       "  'text': \"Why hasn't the bird flu pandemic started?\",\n",
+       "  'score': 0.5738048553466797},\n",
+       " {'id': '23402131',\n",
+       "  'text': 'Where do avian influenza viruses meet in the Americas?',\n",
+       "  'score': 0.5638074278831482},\n",
+       " {'id': '20667302',\n",
+       "  'text': 'The influenza A(H5N1) epidemic at six and a half years: 500 notified human cases and more to come.',\n",
+       "  'score': 0.560465395450592},\n",
+       " {'id': '17338983',\n",
+       "  'text': 'Human avian influenza: how ready are we?',\n",
+       "  'score': 0.555113673210144},\n",
+       " {'id': '24518630',\n",
+       "  'text': 'Recognizing true H5N1 infections in humans during confirmed outbreaks.',\n",
+       "  'score': 0.5501888990402222}]"
+      ]
+     },
+     "execution_count": 87,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "embeddings.search(\"What locations have had H5N1 outbreaks?\", limit=10)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "efa8e59e",
+   "metadata": {},
+   "source": [
+    "Once again the answer is based on the context which mentions the two countries in the answer. The context also discusses the Americas but it doesn't have as strong of language connecting H5N1 outbreaks to the location."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "18e04661",
+   "metadata": {},
+   "source": [
+    "# Add citations\n",
+    "\n",
+    "The last item we'll cover is citations. One of the most important aspects of a RAG process is being able to ensure the answer is based on reality. There are a number of ways to do this but in this example, we'll ask the LLM to perform this step."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 96,
+   "id": "ce10ff39",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Device set to use cuda:0\n"
+     ]
+    }
+   ],
+   "source": [
+    "# Prompt templates\n",
+    "system = \"You are a friendly medical assistant that answers questions\"\n",
+    "template = \"\"\"\n",
+    "Answer the following question using the provided context.\n",
+    "\n",
+    "After the answer, write a citation section with ALL the original article ids used for the answer.\n",
+    "\n",
+    "Question:\n",
+    "{question}\n",
+    "\n",
+    "Context:\n",
+    "{context}\n",
+    "\"\"\"\n",
+    "\n",
+    "def context(question):\n",
+    "    context = []\n",
+    "    for x in embeddings.search(question, limit=10):\n",
+    "        context.append(f\"ARTICLE ID: {x['id']}, TEXT: {x['text']}\")\n",
+    "\n",
+    "    return context\n",
+    "\n",
+    "# Create RAG pipeline\n",
+    "rag = RAG(embeddings, \"Qwen/Qwen3-0.6B\", system=system, template=template, output=\"flatten\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "f79bdc9a",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "H5N1 is a type of avian influenza virus.  \n",
+      "\n",
+      "**Citation Section:**  \n",
+      "- ARTICLE ID: 22010536, TEXT: Is avian influenza virus A(H5N1) a real threat to human health?\n"
+     ]
+    }
+   ],
+   "source": [
+    "question = \"What is H5N1?\"\n",
+    "print(rag(question, context(question), maxlength=2048, stripthink=True))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "15539c5d",
+   "metadata": {},
+   "source": [
+    "As expected, the answer adds a citation section. Also note that the RAG pipeline stripped the thinking section from the result."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "3bdf84b5",
+   "metadata": {},
+   "source": [
+    "# Wrapping up\n",
+    "\n",
+    "This notebook covered how to build a Medical RAG Research process with `txtai`. It also covered how to modify this logic to add in your own knowledge base or use a more sophisticated LLM.\n",
+    "\n",
+    "With an important space such as the medical domain, it's vital to ensure that answers are derived from reliable knowledge. This notebook shows how to add that reliability via RAG. But as with anything in an important domain, there should be a human in the loop and answers shouldn't be blindly relied upon. "
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "local",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.18"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

examples/76_Whats_new_in_txtai_9_0.ipynb

@@ -0,0 +1,309 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {
+    "id": "e3wdiK5fGUoZ"
+   },
+   "source": [
+    "# 💡 What's new in txtai 9.0\n",
+    "\n",
+    "The 9.0 release adds first class support for sparse vector models (i.e. [SPLADE](https://en.wikipedia.org/wiki/Learned_sparse_retrieval)), late interaction models (i.e. [ColBERT](https://huggingface.co/colbert-ir/colbertv2.0)), fixed dimensional encoding (i.e. [MUVERA](https://arxiv.org/abs/2405.19504)) and reranking pipelines ✨ \n",
+    "\n",
+    "The embeddings framework was overhauled to seamlessly support both sparse and dense vector models. Previously, sparse vector support was limited to keyword/term indexes. Now learned sparse retrieval models such as SPLADE are supported. These models can help improve the accuracy of retrieval/search operations, which also improves RAG and Agents.\n",
+    "\n",
+    "Support for late interaction models, such as ColBERT, were also added to the embeddings framework. Unlike traditional vector models that pool outputs into single vector outputs, late interaction models produce multiple vectors. These models are paired with the MUVERA algorithm to transform multiple vectors into fixed dimensional single vectors for search.\n",
+    "\n",
+    "LLMs are quickly converging to produce similar outputs for similar inputs and becoming standard commodities. The retrieval or context layer makes or breaks projects. This is known as putting the R in RAG!\n",
+    "\n",
+    "**Standard upgrade disclaimer below**\n",
+    "\n",
+    "While everything is backwards compatible, it's prudent to backup production indexes before upgrading and test before deploying."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {
+    "id": "p8BbfjrhH-V2"
+   },
+   "source": [
+    "# Install dependencies\n",
+    "\n",
+    "Install `txtai` and all dependencies."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "metadata": {
+    "id": "-OXsTQgaGQPM"
+   },
+   "outputs": [],
+   "source": [
+    "%%capture\n",
+    "!pip install git+https://github.com/neuml/txtai#egg=txtai[ann,vectors]"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {
+    "id": "n4EXrtcYIIYE"
+   },
+   "source": [
+    "# Sparse vector indexes\n",
+    "\n",
+    "The first major change added with this release is `learned sparse retrieval` (aka sparse vector indexes) models. This effort was multi-faceted in that it required both changes to how vectors were generated as well as how they are stored.\n",
+    "\n",
+    "`txtai` uses approximate nearest neighbor (ANN) search for it's vector search operations. The default library is [Faiss](https://github.com/facebookresearch/faiss). There is support for other libraries but in all cases the existing ANN backends only supported dense (i.e. NumPy) vectors.\n",
+    "\n",
+    "There aren't many options out there for sparse ANN search that supports `txtai` requirements, so IVFSparse was introduced. IVFSparse is an Inverted file (IVF) index with flat vector file storage and sparse array support. There is also support for storing sparse vectors in Postgres via [pgvector](https://github.com/pgvector/pgvector).\n",
+    "\n",
+    "Let's see it in action.\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {
+    "id": "hzPD8_cQJNtN"
+   },
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "[{'id': '0',\n",
+       "  'text': 'US tops 5 million confirmed virus cases',\n",
+       "  'score': 0.019873601198196412},\n",
+       " {'id': '1',\n",
+       "  'text': \"Canada's last fully intact ice shelf has suddenly collapsed, forming a Manhattan-sized iceberg\",\n",
+       "  'score': 0.018737798929214476}]"
+      ]
+     },
+     "execution_count": 2,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "from txtai import Embeddings\n",
+    "\n",
+    "# Works with a list, dataset or generator\n",
+    "data = [\n",
+    "  \"US tops 5 million confirmed virus cases\",\n",
+    "  \"Canada's last fully intact ice shelf has suddenly collapsed, forming a Manhattan-sized iceberg\",\n",
+    "  \"Beijing mobilises invasion craft along coast as Taiwan tensions escalate\",\n",
+    "  \"The National Park Service warns against sacrificing slower friends in a bear attack\",\n",
+    "  \"Maine man wins $1M from $25 lottery ticket\",\n",
+    "  \"Make huge profits without work, earn up to $100,000 a day\"\n",
+    "]\n",
+    "\n",
+    "# Create an embeddings\n",
+    "embeddings = Embeddings(sparse=True, content=True)\n",
+    "embeddings.index(data)\n",
+    "embeddings.search(\"North America\", 10)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Late interaction models\n",
+    "\n",
+    "Late interaction models encode data into multi-vector outputs. In other words, multiple input tokens map to multiple output vectors. Then at search time, the maximum similarity algorithm is used to find the best matches between the corpus and a query. This algorithm has achieved excellent results on retrieval benchmarks such as [MTEB](https://github.com/embeddings-benchmark/mteb).\n",
+    "\n",
+    "The downside of this approach is that it produces multiple vectors as opposed a single vector for each input. For example, if a text element tokenizes to many input tokens, there will be many output vectors vs a single one as with standard pooled vector approaches.\n",
+    "\n",
+    "Starting with the 9.0 release, late interaction models are supported with embeddings instances. Late interaction vectors will be transformed into fixed dimensional vectors using the MUVERA algorithm. See below. "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "[{'id': '0',\n",
+       "  'text': 'US tops 5 million confirmed virus cases',\n",
+       "  'score': 0.04216160625219345},\n",
+       " {'id': '1',\n",
+       "  'text': \"Canada's last fully intact ice shelf has suddenly collapsed, forming a Manhattan-sized iceberg\",\n",
+       "  'score': 0.029944246634840965},\n",
+       " {'id': '3',\n",
+       "  'text': 'The National Park Service warns against sacrificing slower friends in a bear attack',\n",
+       "  'score': 0.015931561589241028}]"
+      ]
+     },
+     "execution_count": 3,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "from txtai import Embeddings\n",
+    "\n",
+    "# Works with a list, dataset or generator\n",
+    "data = [\n",
+    "  \"US tops 5 million confirmed virus cases\",\n",
+    "  \"Canada's last fully intact ice shelf has suddenly collapsed, forming a Manhattan-sized iceberg\",\n",
+    "  \"Beijing mobilises invasion craft along coast as Taiwan tensions escalate\",\n",
+    "  \"The National Park Service warns against sacrificing slower friends in a bear attack\",\n",
+    "  \"Maine man wins $1M from $25 lottery ticket\",\n",
+    "  \"Make huge profits without work, earn up to $100,000 a day\"\n",
+    "]\n",
+    "\n",
+    "# Create an embeddings\n",
+    "embeddings = Embeddings(path=\"colbert-ir/colbertv2.0\", content=True)\n",
+    "embeddings.index(data)\n",
+    "embeddings.search(\"North America\", 10)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Reranking pipeline\n",
+    "\n",
+    "Another major new component in this release is the Reranker pipeline. This pipeline takes an embeddings instance, a similarity instance and uses the similarity instance to rerank outputs. This is a key component of the MUVERA paper - using the standard vector index to retrieve candidates then reranking the outputs using the late interaction model."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "[{'id': '1',\n",
+       "  'text': \"Canada's last fully intact ice shelf has suddenly collapsed, forming a Manhattan-sized iceberg\",\n",
+       "  'score': 0.3324427008628845},\n",
+       " {'id': '0',\n",
+       "  'text': 'US tops 5 million confirmed virus cases',\n",
+       "  'score': 0.24423550069332123},\n",
+       " {'id': '3',\n",
+       "  'text': 'The National Park Service warns against sacrificing slower friends in a bear attack',\n",
+       "  'score': 0.16353240609169006}]"
+      ]
+     },
+     "execution_count": 4,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "from txtai import Embeddings\n",
+    "from txtai.pipeline import Reranker, Similarity\n",
+    "\n",
+    "# Works with a list, dataset or generator\n",
+    "data = [\n",
+    "  \"US tops 5 million confirmed virus cases\",\n",
+    "  \"Canada's last fully intact ice shelf has suddenly collapsed, forming a Manhattan-sized iceberg\",\n",
+    "  \"Beijing mobilises invasion craft along coast as Taiwan tensions escalate\",\n",
+    "  \"The National Park Service warns against sacrificing slower friends in a bear attack\",\n",
+    "  \"Maine man wins $1M from $25 lottery ticket\",\n",
+    "  \"Make huge profits without work, earn up to $100,000 a day\"\n",
+    "]\n",
+    "\n",
+    "# Create an embeddings\n",
+    "embeddings = Embeddings(path=\"colbert-ir/colbertv2.0\", content=True)\n",
+    "embeddings.index(data)\n",
+    "\n",
+    "similarity = Similarity(path=\"colbert-ir/colbertv2.0\", lateencode=True)\n",
+    "\n",
+    "ranker = Reranker(embeddings, similarity)\n",
+    "ranker(\"North America\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Notice that while the outputs are the same, the scoring and order is different.\n",
+    "\n",
+    "Let's try a more interesting example."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "[{'id': 'ChatGPT',\n",
+       "  'text': 'ChatGPT is a generative artificial intelligence chatbot developed by OpenAI and released on November 30, 2022. It uses large language models (LLMs) such as GPT-4o as well as other multimodal models to create human-like responses in text, speech, and images. It has access to features such as searching the web, using apps, and running programs. It is credited with accelerating the AI boom, an ongoing period of rapid investment in and public attention to the field of artificial intelligence (AI). Some observers have raised concern about the potential of ChatGPT and similar programs to displace human intelligence, enable plagiarism, or fuel misinformation.',\n",
+       "  'score': 0.6639302968978882},\n",
+       " {'id': 'ChatGPT Search',\n",
+       "  'text': 'ChatGPT Search (originally SearchGPT) is a search engine developed by OpenAI. It combines traditional search engine features with generative pretrained transformers (GPT) to generate responses, including citations to external websites.',\n",
+       "  'score': 0.6477508544921875},\n",
+       " {'id': 'ChatGPT in education',\n",
+       "  'text': 'The usage of ChatGPT in education has sparked considerable debate and exploration. ChatGPT is a chatbot based on large language models (LLMs) that was released by OpenAI in November 2022.',\n",
+       "  'score': 0.5918337106704712}]"
+      ]
+     },
+     "execution_count": 5,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "from txtai import Embeddings\n",
+    "from txtai.pipeline import Reranker, Similarity\n",
+    "\n",
+    "# Create an embeddings\n",
+    "embeddings = Embeddings()\n",
+    "embeddings.load(provider=\"huggingface-hub\", container=\"neuml/txtai-wikipedia\")\n",
+    "\n",
+    "similarity = Similarity(path=\"colbert-ir/colbertv2.0\", lateencode=True)\n",
+    "\n",
+    "ranker = Reranker(embeddings, similarity)\n",
+    "ranker(\"Tell me about ChatGPT\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {
+    "id": "tvnMO1Eai6Gy"
+   },
+   "source": [
+    "# Wrapping up\n",
+    "\n",
+    "This notebook gave a quick overview of txtai 9.0. Updated documentation and more examples will be forthcoming. There is much to cover and much to build on!\n",
+    "\n",
+    "See the following links for more information.\n",
+    "\n",
+    "- [9.0 Release on GitHub](https://github.com/neuml/txtai/releases/tag/v9.0.0)\n",
+    "- [Documentation site](https://neuml.github.io/txtai)"
+   ]
+  }
+ ],
+ "metadata": {
+  "accelerator": "GPU",
+  "colab": {
+   "gpuType": "T4",
+   "provenance": []
+  },
+  "kernelspec": {
+   "display_name": "local",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.18"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 0
+}

examples/78_Accessing_Low_Level_Vector_APIs.ipynb

@@ -0,0 +1,482 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "64be0343",
+   "metadata": {},
+   "source": [
+    "# Accessing Low Level Vector APIs\n",
+    "\n",
+    "The primary interface to build vector databases with `txtai` is through [Embeddings instances](https://neuml.github.io/txtai/embeddings/). `txtai` also supports accessing all of it's features through lower level APIs. \n",
+    "\n",
+    "Let's dive in.\n"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "e8198312",
+   "metadata": {},
+   "source": [
+    "# Install dependencies\n",
+    "\n",
+    "Install `txtai` and all dependencies."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "3f2d0792",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%%capture\n",
+    "!pip install git+https://github.com/neuml/txtai#egg=txtai[ann] gguf"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "46740a21",
+   "metadata": {},
+   "source": [
+    "# Load a dataset\n",
+    "\n",
+    "We'll use a [subset](https://huggingface.co/datasets/m-a-p/FineFineWeb-test) of the [FineFineWeb dataset](https://huggingface.co/datasets/m-a-p/FineFineWeb). This dataset is a domain-labeled version of the general purpose [FineWeb dataset](https://huggingface.co/datasets/HuggingFaceFW/fineweb). "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "3c13e554",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from datasets import load_dataset\n",
+    "\n",
+    "ds = load_dataset(\"m-a-p/FineFineWeb-test\", split=\"train\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "ae49af96",
+   "metadata": {},
+   "source": [
+    "# Building an Embeddings database\n",
+    "\n",
+    "Before going into the low-level API, let's recap how we build an Embeddings database."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 18,
+   "id": "2de3bf4c",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "0.6012564897537231 The National Aeronautics and Space Administration (NASA) is the United States’ civil space program. \n"
+     ]
+    }
+   ],
+   "source": [
+    "from txtai import Embeddings\n",
+    "\n",
+    "embeddings = Embeddings()\n",
+    "embeddings.index(ds[\"text\"][:10000])\n",
+    "for uid, score in embeddings.search(\"nasa\", 1):\n",
+    "    print(score, ds[uid][\"text\"][:100])"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "d6f196cb",
+   "metadata": {},
+   "source": [
+    "This simple example abstracts the heavy lifting behind the `Embeddings` interface. Behind the scenes, it defaults to vectorizing text using [all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2). Vectors are stored in a [Faiss index](https://github.com/facebookresearch/faiss).\n",
+    "\n",
+    "The first 10K records are vectorized and stored in the vector index. Then at query time, the query is vectorized and a vector similarity search is run.\n",
+    "\n",
+    "While the `Embeddings` interface is convenient, it's also possible to access lower level APIs. "
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "ecd576e8",
+   "metadata": {},
+   "source": [
+    "# Vectors Interface\n",
+    "\n",
+    "First, let's vectorize our data using the low level APIs. We'll use the default Hugging Face vectorizer available in `txtai`."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "03c85684",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from txtai.ann import ANNFactory\n",
+    "from txtai.vectors import VectorsFactory\n",
+    "\n",
+    "vectors = VectorsFactory.create({\"path\": \"sentence-transformers/all-MiniLM-L6-v2\"})\n",
+    "data = vectors.vectorize(ds[\"text\"])"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "36fac81a",
+   "metadata": {},
+   "source": [
+    "# ANN Interface\n",
+    "\n",
+    "Now that we have a NumPy array of vectors, let's store them in an Approximate Neighest Neighbor (ANN) backend. Recall earlier, we used the default Faiss interface. For this example, we're going to use the [PyTorch ANN](https://neuml.github.io/txtai/embeddings/configuration/ann/#torch). This will allow us to use new features that are available as of `txtai` 9.1."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "id": "34ebdc85",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "ann = ANNFactory.create({\n",
+    "    \"backend\": \"torch\",\n",
+    "    \"torch\": {\n",
+    "        \"safetensors\": True,\n",
+    "    }\n",
+    "})\n",
+    "ann.index(data)\n",
+    "ann.save(\"vectors.safetensors\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "98a26c9a",
+   "metadata": {},
+   "source": [
+    "This ANN builds a Torch tensor with the vectors and stores them in a [Safetensors](https://github.com/huggingface/safetensors) file.\n",
+    "\n",
+    "The code below shows how the file is simply a standard Safetensors file."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "id": "bcaa4832",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "data (1411868, 384)\n",
+      "Memory = 2068.17 MB\n"
+     ]
+    }
+   ],
+   "source": [
+    "from safetensors import safe_open\n",
+    "\n",
+    "def tensorinfo():\n",
+    "    memory = 0\n",
+    "    with safe_open(\"vectors.safetensors\", framework=\"np\") as f:\n",
+    "        for key in f.keys():\n",
+    "            array = f.get_tensor(key)\n",
+    "            print(key, array.shape)\n",
+    "            memory += array.nbytes\n",
+    "\n",
+    "    print(f\"Memory = {memory / 1024 / 1024:.2f} MB\")\n",
+    "\n",
+    "tensorinfo()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "d6525b1a",
+   "metadata": {},
+   "source": [
+    "# Vector search\n",
+    "\n",
+    "Now let's show how these low-level APIs can be used to implement vector search."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 11,
+   "id": "09b81958",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "The answer to your question, that how many miles is it from earth to mars, is very easy to know. Because of huge satellites which are being sent to\n",
+      "mars in search of life from many countries, we have discovered a lot about mars. According to experts, earth and mars reaches to their closest points\n",
+      "in every 26 months. This situation is considered as opposition of mars as the location of sun and mars in totally opposite to each other in relation\n",
+      "to earth. When this opposition takes place, the planet is visible with a red tint in the sky from earth. And this also gives mars a name, i.e. the red\n",
+      "planet. Mars is also the fourth planet from sun, which is located between Jupiter and earth. Its distance from sun is not only opposite but is also\n",
+      "much further away, than that of the earth and sun. The distance between the sun and mars is said to be 140 million miles. Mars can reach about 128\n",
+      "million miles closer to the sun whereas it can even travel around 154 million miles away from it. The assumed distance between mars and earth is said\n",
+      "to be between 40 to 225 million miles. The distance between these two planets keeps on changing throughout the year because of the elliptical path in\n",
+      "which all the planets rotate. As the distance between mars, sun and earth is so much high, it takes a Martian year, for mars to go around the sun. The\n",
+      "Martian period includes a time of around 687 earth days. This means that, it takes more than 2 years for the mars to reach its initial rotation point.\n",
+      "If we talk about one Martian day, it is the total time which is taken by a planet to spin around once. This day usually lasts longer than our regular\n",
+      "earth days. So this was the actual reason which states the distance between earth and mars. \n",
+      "\n",
+      " 0.7060051560401917\n"
+     ]
+    }
+   ],
+   "source": [
+    "import textwrap\n",
+    "\n",
+    "def search(text):\n",
+    "    result = ann.search(vectors.vectorize([text]), 1)\n",
+    "    index, score = result[0][0]\n",
+    "    print(textwrap.fill(ds[index][\"text\"], width=150), \"\\n\\n\", score)\n",
+    "\n",
+    "search(\"How far is earth from mars?\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "8adf64db",
+   "metadata": {},
+   "source": [
+    "# Torch 4-bit quantization\n",
+    "\n",
+    "`txtai` 9.1 adds a new feature: 4-bit vector quantization. This means that instead of using 32-bit floats for each vector dimension, this method uses 4 bits. This reduces memory usage to ~12-13% of the original size."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 12,
+   "id": "cbe5af13",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "absmax (8471208,)\n",
+      "code (16,)\n",
+      "data (271078656, 1)\n",
+      "shape (2,)\n",
+      "Memory = 290.84 MB\n"
+     ]
+    }
+   ],
+   "source": [
+    "ann = ANNFactory.create({\n",
+    "    \"backend\": \"torch\",\n",
+    "    \"torch\": {\n",
+    "        \"safetensors\": True,\n",
+    "        \"quantize\": {\n",
+    "            \"type\": \"nf4\"\n",
+    "        }\n",
+    "    }\n",
+    "})\n",
+    "ann.index(data)\n",
+    "ann.save(\"vectors.safetensors\")\n",
+    "\n",
+    "tensorinfo()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "ca7b1e0e",
+   "metadata": {},
+   "source": [
+    "Note how the unquantized vectors took 2068.17 MB and this only takes 290.84 MB! With quantization and ever growing GPUs, this opens the possibility of pinning your entire vector database in GPU memory!\n",
+    "\n",
+    "For example, let's extrapolate this dataset to 100M rows.\n",
+    "\n",
+    "```\n",
+    "(290.84 MB / 1,411,868) * 100,000,000 = 20,599.7 MB\n",
+    "```\n",
+    "\n",
+    "An entire 100M row dataset could fit into a single RTX 3090 / 4090 consumer GPU!\n",
+    "\n",
+    "Let's confirm search still works the same."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 13,
+   "id": "a02578dd",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "The answer to your question, that how many miles is it from earth to mars, is very easy to know. Because of huge satellites which are being sent to\n",
+      "mars in search of life from many countries, we have discovered a lot about mars. According to experts, earth and mars reaches to their closest points\n",
+      "in every 26 months. This situation is considered as opposition of mars as the location of sun and mars in totally opposite to each other in relation\n",
+      "to earth. When this opposition takes place, the planet is visible with a red tint in the sky from earth. And this also gives mars a name, i.e. the red\n",
+      "planet. Mars is also the fourth planet from sun, which is located between Jupiter and earth. Its distance from sun is not only opposite but is also\n",
+      "much further away, than that of the earth and sun. The distance between the sun and mars is said to be 140 million miles. Mars can reach about 128\n",
+      "million miles closer to the sun whereas it can even travel around 154 million miles away from it. The assumed distance between mars and earth is said\n",
+      "to be between 40 to 225 million miles. The distance between these two planets keeps on changing throughout the year because of the elliptical path in\n",
+      "which all the planets rotate. As the distance between mars, sun and earth is so much high, it takes a Martian year, for mars to go around the sun. The\n",
+      "Martian period includes a time of around 687 earth days. This means that, it takes more than 2 years for the mars to reach its initial rotation point.\n",
+      "If we talk about one Martian day, it is the total time which is taken by a planet to spin around once. This day usually lasts longer than our regular\n",
+      "earth days. So this was the actual reason which states the distance between earth and mars. \n",
+      "\n",
+      " 0.6982609033584595\n"
+     ]
+    }
+   ],
+   "source": [
+    "search(\"How far is earth from mars?\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "f0e62670",
+   "metadata": {},
+   "source": [
+    "Same result. Note the score is slightly different but this is expected."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "89abb301",
+   "metadata": {},
+   "source": [
+    "# GGUF Support\n",
+    "\n",
+    "`txtai` 9.1 also adds support for [GGML](https://github.com/ggml-org/ggml) / [GGUF](https://huggingface.co/docs/hub/en/gguf) popularized by [llama.cpp](https://github.com/ggml-org/llama.cpp)."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 19,
+   "id": "e04498fc",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "ann = ANNFactory.create({\n",
+    "    \"backend\": \"ggml\",\n",
+    "    \"ggml\": {\n",
+    "        \"quantize\": \"Q4_0\"\n",
+    "    }\n",
+    "})\n",
+    "ann.index(data)\n",
+    "ann.save(\"vectors.gguf\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "7bea0f58",
+   "metadata": {},
+   "source": [
+    "Now let's check out the generated file using the [gguf](https://github.com/ggml-org/llama.cpp/tree/master/gguf-py) package provided by llama.cpp."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 20,
+   "id": "b0e6f85e",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Tensor Name                    | Shape           | Size         | Quantization\n",
+      "--------------------------------------------------------------------------------\n",
+      "data                           | 384x1411868     | 258.52 MB    | Q4_0\n"
+     ]
+    }
+   ],
+   "source": [
+    "from gguf.gguf_reader import GGUFReader\n",
+    "\n",
+    "reader = GGUFReader(\"vectors.gguf\")\n",
+    "\n",
+    "# List all tensors\n",
+    "info = \"{:<30} | {:<15} | {:<12} | {}\"\n",
+    "print(info.format(\"Tensor Name\", \"Shape\", \"Size\", \"Quantization\"))\n",
+    "print(\"-\" * 80)\n",
+    "for tensor in reader.tensors:\n",
+    "    shape = \"x\".join(map(str, tensor.shape))\n",
+    "    size = f\"{tensor.n_elements / 2 / 1024 / 1024:.2f} MB\"\n",
+    "    quantization = tensor.tensor_type.name\n",
+    "    print(info.format(tensor.name, shape, size, quantization))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "7be79bf7",
+   "metadata": {},
+   "source": [
+    "And search like we did with Torch."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 18,
+   "id": "450dc024",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "The answer to your question, that how many miles is it from earth to mars, is very easy to know. Because of huge satellites which are being sent to\n",
+      "mars in search of life from many countries, we have discovered a lot about mars. According to experts, earth and mars reaches to their closest points\n",
+      "in every 26 months. This situation is considered as opposition of mars as the location of sun and mars in totally opposite to each other in relation\n",
+      "to earth. When this opposition takes place, the planet is visible with a red tint in the sky from earth. And this also gives mars a name, i.e. the red\n",
+      "planet. Mars is also the fourth planet from sun, which is located between Jupiter and earth. Its distance from sun is not only opposite but is also\n",
+      "much further away, than that of the earth and sun. The distance between the sun and mars is said to be 140 million miles. Mars can reach about 128\n",
+      "million miles closer to the sun whereas it can even travel around 154 million miles away from it. The assumed distance between mars and earth is said\n",
+      "to be between 40 to 225 million miles. The distance between these two planets keeps on changing throughout the year because of the elliptical path in\n",
+      "which all the planets rotate. As the distance between mars, sun and earth is so much high, it takes a Martian year, for mars to go around the sun. The\n",
+      "Martian period includes a time of around 687 earth days. This means that, it takes more than 2 years for the mars to reach its initial rotation point.\n",
+      "If we talk about one Martian day, it is the total time which is taken by a planet to spin around once. This day usually lasts longer than our regular\n",
+      "earth days. So this was the actual reason which states the distance between earth and mars. \n",
+      "\n",
+      " 0.7043964862823486\n"
+     ]
+    }
+   ],
+   "source": [
+    "search(\"How far is earth from mars?\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "af1a344c",
+   "metadata": {},
+   "source": [
+    "# Wrapping up\n",
+    "\n",
+    "While the `Embeddings` interface is the preferred way to build vector databases with `txtai`, it's entirely possible to also build with the low level APIs!"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "local",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.19"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

examples/79_RAG_is_more_than_Vector_Search.ipynb

@@ -0,0 +1,251 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "4c8ab3b0",
+   "metadata": {},
+   "source": [
+    "# RAG is more than Vector Search\n",
+    "\n",
+    "Retrieval Augmented Generation (RAG) is often associated with vector search. And while that is a primary use case, any search will do.\n",
+    "\n",
+    "- ✅ Vector Search\n",
+    "- ✅ Web Search\n",
+    "- ✅ SQL Query\n",
+    "\n",
+    "This notebook will go over a few RAG examples covering different retrieval methods. These examples require txtai 9.3+."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "ab8a77d1",
+   "metadata": {},
+   "source": [
+    "# Install dependencies\n",
+    "\n",
+    "Install `txtai` and all dependencies."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "21bebb46",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%%capture\n",
+    "!pip install git+https://github.com/neuml/txtai#egg=txtai[pipeline-data]\n",
+    "\n",
+    "# Download example SQL database\n",
+    "!wget https://huggingface.co/NeuML/txtai-wikipedia-slim/resolve/main/documents"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "1ece2b09",
+   "metadata": {},
+   "source": [
+    "# RAG with Late Interaction\n",
+    "\n",
+    "The first example will cover RAG with ColBERT / Late Interaction retrieval. TxtAI 9.0 added support for [MUVERA](https://arxiv.org/abs/2405.19504) and [ColBERT](https://arxiv.org/abs/2112.01488) multi-vector ranking. \n",
+    "\n",
+    "We'll build a pipeline that reads the ColBERT v2 paper, extracts the text into sections and builds an index with a ColBERT model. Then we'll wrap that as a [Reranker pipeline](https://neuml.github.io/txtai/pipeline/text/reranker/) using the same ColBERT model. Finally a RAG pipeline will utilize this for retrieval.\n",
+    "\n",
+    "_Note: This uses the custom [ColBERT Muvera Nano](https://huggingface.co/NeuML/colbert-muvera-nano) model which is only 970K parameters! That's right thousands. It's surprisingly effective._"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "id": "dae2e6dc",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "This paper introduces ColBERTv2, a neural information retrieval model that enhances the quality and efficiency of late interaction by combining an aggressive residual compression mechanism with a denoised supervision strategy, achieving state-of-the-art performance across diverse benchmarks while reducing the model's space footprint by 6–10× compared to previous methods.\n"
+     ]
+    }
+   ],
+   "source": [
+    "from txtai import Embeddings, RAG, Textractor\n",
+    "from txtai.pipeline import Reranker, Similarity\n",
+    "\n",
+    "# Get text from ColBERT v2 paper\n",
+    "textractor = Textractor(sections=True, backend=\"docling\")\n",
+    "data = textractor(\"https://arxiv.org/pdf/2112.01488\")\n",
+    "\n",
+    "# MUVERA fixed dimensional encodings\n",
+    "embeddings = Embeddings(content=True, path=\"neuml/colbert-muvera-nano\", vectors={\"trust_remote_code\": True})\n",
+    "embeddings.index(data)\n",
+    "\n",
+    "# Re-rank using same late interaction model\n",
+    "reranker = Reranker(embeddings, Similarity(\"neuml/colbert-muvera-nano\", lateencode=True, vectors={\"trust_remote_code\": True}))\n",
+    "\n",
+    "template = \"\"\"\n",
+    "  Answer the following question using the provided context.\n",
+    "\n",
+    "  Question:\n",
+    "  {question}\n",
+    "\n",
+    "  Context:\n",
+    "  {context}\n",
+    "\"\"\"\n",
+    "\n",
+    "# RAG with late interaction models\n",
+    "rag = RAG(reranker, \"Qwen/Qwen3-4B-Instruct-2507\", template=template, output=\"flatten\")\n",
+    "print(rag(\"Write a sentence abstract about this paper\", maxlength=2048))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "b31e46de",
+   "metadata": {},
+   "source": [
+    "# RAG with a Web Search\n",
+    "\n",
+    "Next we'll run a RAG pipeline using a web search as the retrieval method."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "b61b684b",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Artificial intelligence (AI) is the capability of computational systems to perform tasks typically associated with human intelligence, such as learning, reasoning, problem-solving, perception, and decision-making. It involves technologies like machine learning, deep learning, and natural language processing, and enables machines to simulate human-like learning, comprehension, problem solving, decision-making, creativity, and autonomy.\n"
+     ]
+    }
+   ],
+   "source": [
+    "from smolagents import WebSearchTool\n",
+    "\n",
+    "tool = WebSearchTool()\n",
+    "\n",
+    "def websearch(queries, limit):\n",
+    "    results = []\n",
+    "    for query in queries:\n",
+    "        result = [\n",
+    "            {\"id\": i, \"text\": f'{x[\"title\"]} {x[\"description\"]}', \"score\": 1.0} for i, x in enumerate(tool.search(query))\n",
+    "        ]\n",
+    "        results.append(result[:limit])\n",
+    "\n",
+    "    return results\n",
+    "\n",
+    "# RAG with a websearch\n",
+    "rag = RAG(websearch, \"Qwen/Qwen3-4B-Instruct-2507\", template=template, output=\"flatten\")\n",
+    "print(rag(\"What is AI?\", maxlength=2048))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "45c8a096",
+   "metadata": {},
+   "source": [
+    "# RAG with a SQL Query\n",
+    "\n",
+    "The last example we'll cover is running RAG with a SQL query. We'll use the SQL database that's a component of the [txtai-wikipedia-slim](https://huggingface.co/NeuML/txtai-wikipedia-slim) embeddings database.\n",
+    "\n",
+    "Since this is just a database with Wikipedia abstracts, we'll need a way to build a SQL query from a search query. For that we'll use an LLM to extract a keyword to use in a `LIKE` clause.\n",
+    "\n",
+    "Given that the LLM used was released in August 2025, let's ask it a question that can only be accurated answered with external data. `Who won the 2025 World Series?` which ended in November."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "c7ff2a35",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "In the 2025 World Series, the Los Angeles Dodgers defeated the Toronto Blue Jays in seven games to win the championship. The series took place from October 24 to November 1 (ending early on November 2, Toronto time). Dodgers pitcher Yoshinobu Yamamoto was named the World Series MVP. The series was televised by Fox in the United States and by Sportsnet in Canada.\n"
+     ]
+    }
+   ],
+   "source": [
+    "import sqlite3\n",
+    "\n",
+    "from txtai import LLM\n",
+    "\n",
+    "def keyword(query):\n",
+    "    return llm(f\"\"\"\n",
+    "        Extract a keyword for this search query: {query}.\n",
+    "        Return only text with no other formatting or explanation.\n",
+    "    \"\"\")\n",
+    "\n",
+    "def sqlsearch(queries, limit):\n",
+    "    results = []\n",
+    "    sql = \"SELECT id, text FROM sections WHERE id LIKE ? LIMIT ?\"\n",
+    "\n",
+    "    for query in queries:\n",
+    "        # Extract a keyword for this search\n",
+    "        query = keyword(query)\n",
+    "\n",
+    "        # Run the SQL Query\n",
+    "        results.append([\n",
+    "            {\"id\": uid, \"text\": text, \"score\": 1.0}\n",
+    "            for uid, text in cursor.execute(sql, [f\"%{query}%\", limit])\n",
+    "        ])\n",
+    "\n",
+    "    return results\n",
+    "\n",
+    "# Load the database\n",
+    "cursor = sqlite3.connect(\"documents\")\n",
+    "\n",
+    "# Load the LLM\n",
+    "llm = LLM(\"Qwen/Qwen3-4B-Instruct-2507\")\n",
+    "\n",
+    "# RAG with a SQL query\n",
+    "rag = RAG(sqlsearch, llm, template=template, output=\"flatten\")\n",
+    "print(rag(\"Tell me what happened in the 2025 World Series\", maxlength=2048))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "a9f78076",
+   "metadata": {},
+   "source": [
+    "And as we see, this answer is using the SQL database!"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "97b1e282",
+   "metadata": {},
+   "source": [
+    "# Wrapping up\n",
+    "\n",
+    "This notebook showed that RAG is about much more than vector search. With txtai 9.3+, any callable method is now supported for retrieval. Enjoy!"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "local",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.19"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

examples/80_Distilling_Knowledge_into_Tiny_LLMs.ipynb

@@ -0,0 +1,360 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "c65bdb85",
+   "metadata": {},
+   "source": [
+    "# Distilling Knowledge into Tiny LLMs\n",
+    "\n",
+    "Large Language Models (LLMs) are the magic behind AI. These massive billion and trillion parameter models have been shown to generalize well when trained on enough data.\n",
+    "\n",
+    "A big problem is that they are hard to run and expensive. So many just call LLMs through APIs such as OpenAI or Claude. Additionally, in many instances, developers spend a lot of time with complex prompt logic hoping to cover all the edge cases and believe they need a model that's large enough to handle all the rules.\n",
+    "\n",
+    "If you truly want control over your business processes, running a local model is a better choice. And the good news is that it doesn't have to be a giant and expensive multi-billion parameter model. We can finetune LLMs to handle our specific business logic, which helps us take control and limit prompt complexity. \n",
+    "\n",
+    "This notebook will show how we can distill knowledge into tiny LLMs."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "279ecd4b",
+   "metadata": {},
+   "source": [
+    "# Install dependencies\n",
+    "\n",
+    "Install `txtai` and all dependencies."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "4f7bca3a",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%%capture\n",
+    "!pip install git+https://github.com/neuml/txtai#egg=txtai[pipeline-train] datasets"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "c7b1b4be",
+   "metadata": {},
+   "source": [
+    "# The LLM\n",
+    "\n",
+    "We'll use a [600M parameter Qwen3 model](https://hf.co/qwen/qwen3-0.6b) for this example. Our target task will be translating user requests into linux commands."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "id": "e1d95dd8",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from txtai import LLM\n",
+    "\n",
+    "llm = LLM(\"Qwen/Qwen3-0.6B\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "8f276cfc",
+   "metadata": {},
+   "source": [
+    "Let's try one with the base model as it is."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 11,
+   "id": "686c3983",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "'ps -e'"
+      ]
+     },
+     "execution_count": 11,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "llm(\"\"\"\n",
+    "Translate the following request into a linux command. Only print the command.\n",
+    "\n",
+    "Find number of logged in users\n",
+    "\"\"\", maxlength=1024)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "dcd92414",
+   "metadata": {},
+   "source": [
+    "As we can see, the model actually has a good understanding and at least prints a command. But in this case it's not correct. Let's get to fine-tuning!"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "73666982",
+   "metadata": {},
+   "source": [
+    "# Finetuning the LLM with knowledge\n",
+    "\n",
+    "Yes, 600M parameters is small and we can't possibly expect it to do well with everything. But the good news is that we can distill knowledge into this tiny LLM and make it better. We'll use this [linux commands dataset](https://huggingface.co/datasets/mecha-org/linux-command-dataset) from the Hugging Face Hub. We'll also use this [training pipeline from txtai](https://neuml.github.io/txtai/pipeline/train/trainer).\n",
+    "\n",
+    "First, we'll create the training dataset. We'll use the same prompt strategy from above.\n",
+    "\n",
+    "```python\n",
+    "\"\"\"\n",
+    "Translate the following request into a linux command. Only print the command.\n",
+    "\n",
+    "{user request}\n",
+    "\"\"\"\n",
+    "```"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "id": "43d6b563",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from datasets import load_dataset\n",
+    "from transformers import AutoTokenizer\n",
+    "\n",
+    "# LLM path\n",
+    "path = \"Qwen/Qwen3-0.6B\"\n",
+    "tokenizer = AutoTokenizer.from_pretrained(path)\n",
+    "\n",
+    "# Load the training dataset\n",
+    "dataset = load_dataset(\"mecha-org/linux-command-dataset\", split=\"train\")\n",
+    "\n",
+    "def prompt(row):\n",
+    "    text = tokenizer.apply_chat_template([\n",
+    "        {\"role\": \"system\", \"content\": \"Translate the following request into a linux command. Only print the command.\"},\n",
+    "        {\"role\": \"user\", \"content\": row[\"input\"]},\n",
+    "        {\"role\": \"assistant\", \"content\": row[\"output\"]}\n",
+    "    ], tokenize=False, enable_thinking=False)\n",
+    "\n",
+    "    return {\"text\": text}\n",
+    "\n",
+    "# Map to training prompts\n",
+    "train = dataset.map(prompt, remove_columns=[\"input\", \"output\"])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "7f71dce0",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/html": [
+       "\n",
+       "    <div>\n",
+       "      \n",
+       "      <progress value='210' max='210' style='width:300px; height:20px; vertical-align: middle;'></progress>\n",
+       "      [210/210 01:12, Epoch 1/1]\n",
+       "    </div>\n",
+       "    <table border=\"1\" class=\"dataframe\">\n",
+       "  <thead>\n",
+       " <tr style=\"text-align: left;\">\n",
+       "      <th>Step</th>\n",
+       "      <th>Training Loss</th>\n",
+       "    </tr>\n",
+       "  </thead>\n",
+       "  <tbody>\n",
+       "    <tr>\n",
+       "      <td>50</td>\n",
+       "      <td>0.625300</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <td>100</td>\n",
+       "      <td>0.490200</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <td>150</td>\n",
+       "      <td>0.403300</td>\n",
+       "    </tr>\n",
+       "    <tr>\n",
+       "      <td>200</td>\n",
+       "      <td>0.391800</td>\n",
+       "    </tr>\n",
+       "  </tbody>\n",
+       "</table><p>"
+      ],
+      "text/plain": [
+       "<IPython.core.display.HTML object>"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "from txtai.pipeline import HFTrainer\n",
+    "\n",
+    "# Load the training pipeline\n",
+    "trainer = HFTrainer()\n",
+    "\n",
+    "# Train the model\n",
+    "# Set output_dir to save, trained in memory for this example\n",
+    "model = trainer(\n",
+    "    \"Qwen/Qwen3-0.6B\",\n",
+    "    train,\n",
+    "    task=\"language-generation\",\n",
+    "    maxlength=512,\n",
+    "    bf16=True,\n",
+    "    per_device_train_batch_size=4,\n",
+    "    num_train_epochs=1,\n",
+    "    logging_steps=50,\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "68b5bc4e",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "'who | wc -l'"
+      ]
+     },
+     "execution_count": 17,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "from txtai import LLM\n",
+    "\n",
+    "llm = LLM(model)\n",
+    "\n",
+    "llm([\n",
+    "    {\"role\": \"system\", \"content\": \"Translate the following request into a linux command. Only print the command.\"},\n",
+    "    {\"role\": \"user\", \"content\": \"Find number of logged in users\"}\n",
+    "])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 14,
+   "id": "7427186d",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "'ls ~/'"
+      ]
+     },
+     "execution_count": 14,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "llm([\n",
+    "    {\"role\": \"system\", \"content\": \"Translate the following request into a linux command. Only print the command.\"},\n",
+    "    {\"role\": \"user\", \"content\": \"List the files in my home directory\"}\n",
+    "])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 16,
+   "id": "a141d63b",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "'zip -r data.zip data'"
+      ]
+     },
+     "execution_count": 16,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "llm([\n",
+    "    {\"role\": \"system\", \"content\": \"Translate the following request into a linux command. Only print the command.\"},\n",
+    "    {\"role\": \"user\", \"content\": \"Zip the data directory with all it's contents\"}\n",
+    "])"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "fc7d0fda",
+   "metadata": {},
+   "source": [
+    "It even works well without the system prompt."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 20,
+   "id": "a7d0c671",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "'du -sh ~'"
+      ]
+     },
+     "execution_count": 20,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "llm(\"Calculate the total amount of disk space used for my home directory. Only print the total.\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "48cfcacf",
+   "metadata": {},
+   "source": [
+    "# Wrapping up\n",
+    "\n",
+    "This notebook demonstrated how it's very straightforward to distill knowledge into LLMs with `txtai`. Don't always go for the giant LLM, spend a little time finetuning a tiny LLM, it is well worth it!"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "local",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.19"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

examples/81_OpenCode_as_a_txtai_LLM.ipynb

@@ -0,0 +1,377 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "f95d2373",
+   "metadata": {},
+   "source": [
+    "# OpenCode as a txtai LLM\n",
+    "\n",
+    "[OpenCode](https://github.com/anomalyco/opencode) is an open source AI coding agent. It's rapidly growing in popularity as an open alternative to [Claude Code](https://code.claude.com/docs/en/overview). \n",
+    "\n",
+    "OpenCode shows the power of permissive open source. It's rapidly undergoing mass adoption and as of January 2026 sits at `81K+` GitHub ⭐'s. OpenCode supports running as a [local server](https://opencode.ai/docs/server/) via `opencode serve`.\n",
+    "\n",
+    "This notebook will explore how TxtAI integrates this as a LLM pipeline."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "a8719b92",
+   "metadata": {},
+   "source": [
+    "# Install dependencies\n",
+    "\n",
+    "Install `txtai` and all dependencies."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "e00c85bc",
+   "metadata": {
+    "vscode": {
+     "languageId": "shellscript"
+    }
+   },
+   "outputs": [],
+   "source": [
+    "%%capture\n",
+    "!pip install git+https://github.com/neuml/txtai#egg=txtai[api,pipeline-llm]\n",
+    "\n",
+    "# Install OpenCode if not already installed and run serve\n",
+    "!curl -fsSL https://opencode.ai/install | bash\n",
+    "!opencode serve &"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "73ffe60e",
+   "metadata": {},
+   "source": [
+    "# Create a OpenCode LLM\n",
+    "\n",
+    "Now that everything is installed, let's test it out!"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "id": "0e35a304",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/markdown": [
+       "```python\n",
+       "print(\"Hello, World!\")\n",
+       "```"
+      ],
+      "text/plain": [
+       "<IPython.core.display.Markdown object>"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "from IPython.display import Markdown, display\n",
+    "\n",
+    "from txtai import LLM\n",
+    "\n",
+    "def generate(request):\n",
+    "    display(Markdown(llm(request)))\n",
+    "\n",
+    "# Connect to OpenCode server and use default LLM provider\n",
+    "llm = LLM(\"opencode\")\n",
+    "\n",
+    "generate(\"Show a Python Hello World example\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "55dd0fad",
+   "metadata": {},
+   "source": [
+    "OK, now let's try this for a more advanced use case."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 8,
+   "id": "b3080841",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/markdown": [
+       "```python\n",
+       "from txtai import Embeddings, LLM\n",
+       "\n",
+       "# Create embeddings model\n",
+       "embeddings = Embeddings(path=\"sentence-transformers/all-MiniLM-L6-v2\")\n",
+       "\n",
+       "# Create LLM\n",
+       "llm = LLM(\"huggingface/Qwen/Qwen2.5-1.5B-Instruct\")\n",
+       "\n",
+       "# Sample documents\n",
+       "documents = [\n",
+       "    \"Python is a programming language created by Guido van Rossum\",\n",
+       "    \"Machine learning is a subset of artificial intelligence\",\n",
+       "    \"Deep learning uses neural networks with multiple layers\"\n",
+       "]\n",
+       "\n",
+       "# Build the index\n",
+       "embeddings.index([(i, doc) for i, doc in enumerate(documents)])\n",
+       "\n",
+       "# RAG function\n",
+       "def rag_query(question):\n",
+       "    # Search for relevant documents\n",
+       "    results = embeddings.search(question, 2)\n",
+       "    \n",
+       "    # Build context from retrieved documents\n",
+       "    context = \" \".join([documents[result[0]] for result in results])\n",
+       "    \n",
+       "    # Generate response using LLM with context\n",
+       "    prompt = f\"Context: {context}\\n\\nQuestion: {question}\\n\\nAnswer:\"\n",
+       "    return llm(prompt)\n",
+       "\n",
+       "# Example usage\n",
+       "question = \"What is Python?\"\n",
+       "answer = rag_query(question)\n",
+       "print(f\"Question: {question}\")\n",
+       "print(f\"Answer: {answer}\")\n",
+       "```"
+      ],
+      "text/plain": [
+       "<IPython.core.display.Markdown object>"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "generate(\"Show a simple and basic TxtAI RAG example\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "b44e0309",
+   "metadata": {},
+   "source": [
+    "Interesting! It's not limited to Python though."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "id": "0a2ec1b4",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/markdown": [
+       "```assembly\n",
+       "section .data\n",
+       "    hello db 'Hello, World!', 10    ; String with newline\n",
+       "    hello_len equ $ - hello         ; Length of string\n",
+       "\n",
+       "section .text\n",
+       "    global _start\n",
+       "\n",
+       "_start:\n",
+       "    ; sys_write system call\n",
+       "    mov eax, 4                      ; sys_write\n",
+       "    mov ebx, 1                      ; stdout\n",
+       "    mov ecx, hello                  ; message\n",
+       "    mov edx, hello_len              ; message length\n",
+       "    int 0x80                        ; kernel interrupt\n",
+       "\n",
+       "    ; sys_exit system call\n",
+       "    mov eax, 1                      ; sys_exit\n",
+       "    mov ebx, 0                      ; exit code 0\n",
+       "    int 0x80                        ; kernel interrupt\n",
+       "```\n",
+       "\n",
+       "Compile and run:\n",
+       "```bash\n",
+       "nasm -f elf32 hello.asm -o hello.o\n",
+       "ld -m elf_i386 hello.o -o hello\n",
+       "./hello\n",
+       "```"
+      ],
+      "text/plain": [
+       "<IPython.core.display.Markdown object>"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "generate(\"Show a x86 assembly hello world\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 11,
+   "id": "4e28d5c4",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/markdown": [
+       "```python\n",
+       "a = None\n",
+       "if a is not None:\n",
+       "    a.startswith(\"hello\")\n",
+       "else:\n",
+       "    print(\"a is None, cannot call startswith\")\n",
+       "```\n",
+       "\n",
+       "**What's wrong:** The variable `a` is `None`, which is a `NoneType` object. The `startswith()` method only exists for strings, not for `NoneType`. You need to check if `a` is not `None` before calling string methods on it."
+      ],
+      "text/plain": [
+       "<IPython.core.display.Markdown object>"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "generate(\"\"\"\n",
+    "Fix the following error in the code below and explain what's wrong.\n",
+    "\n",
+    "Error:\n",
+    "AttributeError: 'NoneType' object has no attribute 'startswith'\n",
+    "\n",
+    "Code:\n",
+    "a = None\n",
+    "a.startswith(\"hello\")\n",
+    "\"\"\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "156122f0",
+   "metadata": {},
+   "source": [
+    "# OpenAI-compatible Endpoint\n",
+    "\n",
+    "Now that OpenCode is integrated into the TxtAI ecosystem, it opens a lot of different opportunities. Let's say we want to host a local OpenAI compatible endpoint, no problem!\n",
+    "\n",
+    "Write the following file.\n",
+    "\n",
+    "## config.yml\n",
+    "\n",
+    "```yaml\n",
+    "# Enable OpenAI compat endpoint\n",
+    "openai: True\n",
+    "\n",
+    "llm:\n",
+    "  path: opencode\n",
+    "```\n",
+    "\n",
+    "and start the following process.\n",
+    "\n",
+    "```\n",
+    "CONFIG=config.yml uvicorn \"txtai.api:app\"\n",
+    "```"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 16,
+   "id": "33057f59",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/markdown": [
+       "```python\n",
+       "def factorial(n):\n",
+       "    if n < 0:\n",
+       "        raise ValueError(\"Factorial is not defined for negative numbers\")\n",
+       "    if n == 0 or n == 1:\n",
+       "        return 1\n",
+       "    result = 1\n",
+       "    for i in range(2, n + 1):\n",
+       "        result *= i\n",
+       "    return result\n",
+       "\n",
+       "# Example usage:\n",
+       "# print(factorial(5))  # Output: 120\n",
+       "```"
+      ],
+      "text/plain": [
+       "<IPython.core.display.Markdown object>"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "from IPython.display import Markdown, display\n",
+    "\n",
+    "from openai import OpenAI\n",
+    "\n",
+    "client = OpenAI(\n",
+    "    base_url=\"http://localhost:8000/v1\",\n",
+    "    api_key=\"api-key\",\n",
+    ")\n",
+    "\n",
+    "response = client.chat.completions.create(\n",
+    "    messages=[{\n",
+    "        \"role\": \"user\",\n",
+    "        \"content\": \"Show a Python factorial function\",\n",
+    "    }],\n",
+    "    model=\"opencode\",\n",
+    ")\n",
+    "\n",
+    "display(Markdown((response.choices[0].message.content)))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "58a9169b",
+   "metadata": {},
+   "source": [
+    "Just like before we get an answer! This time via the OpenAI client. Fun times."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "47f9dbbd",
+   "metadata": {},
+   "source": [
+    "# Wrapping up\n",
+    "\n",
+    "This notebook showed how to connect TxtAI and OpenCode. This will lead to some interesting integrations. One such example is [ncoder](https://github.com/neuml/ncoder), an AI coding agent that integrates with Jupyter Notebooks. Stay tuned for more!"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "local",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.19"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

examples/82_Agentic_College_Search.ipynb

@@ -0,0 +1,358 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "5a13697b",
+   "metadata": {},
+   "source": [
+    "# Agentic College Search\n",
+    "\n",
+    "This example will demonstrate how to use `txtai` agents to identify list of strong engineering colleges with baseball programs 🎓⚙️⚾.\n",
+    "\n",
+    "We'll setup a default [agents.md](https://github.com/agentsmd/agents.md) file and then give an LLM access to the web.\n",
+    "\n",
+    "Let's get started!"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "face0f9e",
+   "metadata": {},
+   "source": [
+    "# Install dependencies\n",
+    "\n",
+    "Install `txtai` and all dependencies."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "f5da3842",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%%capture\n",
+    "!pip install git+https://github.com/neuml/txtai#egg=txtai[agent]"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "f9a6a500",
+   "metadata": {},
+   "source": [
+    "# Define an `agents.md` file\n",
+    "\n",
+    "Next, we'll define an agents.md file. This file defines default behavior for our college search agent and is inserted into the system prompt of every agent run."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "4e0c500d",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%%writefile agents.md\n",
+    "# College Researcher\n",
+    "\n",
+    "You are a college research agent. Your job is to help find the best matching schools given a set of criteria.\n",
+    "\n",
+    "## Academics\n",
+    "\n",
+    "Work to match the request with a set of schools that match the academic requirements\n",
+    "\n",
+    "## Athletics\n",
+    "\n",
+    "Some requests may be interested in collegiate athletics. Make sure to balance that with other parts of the request."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "83dee7de",
+   "metadata": {},
+   "source": [
+    "# Identify the Top 20 schools\n",
+    "\n",
+    "Next, let's run a prompt to identify the best 20 colleges given a set of criteria.\n",
+    "\n",
+    "First, we'll setup the scaffolding code to create and run an agent. We'll use a [Qwen3 4B non-thinking LLM](https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507) as the agent's model. We'll add the `websearch` and `webview` tools to the agent along with the `agents.md` file previously created.\n",
+    "\n",
+    "Additionally, we'll add a sliding window of the last 5 responses as \"agent memory\". This will help create a rolling dialogue. "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "410f2585",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from txtai import Agent\n",
+    "from IPython.display import display, Markdown\n",
+    "\n",
+    "def run(query, reset=False):\n",
+    "    answer = agent(query, maxlength=60000, reset=reset)\n",
+    "    display(Markdown(answer))\n",
+    "\n",
+    "agent = Agent(model=\"Qwen/Qwen3-4B-Instruct-2507\", tools=[\"websearch\", \"webview\"], memory=5, instructions=\"agents.md\", max_steps=50)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "53aa302b",
+   "metadata": {},
+   "source": [
+    "Now we can run the search. To come up with our list of colleges, a series of web searches were executed by the agent. It's like if you had the superpower 🚀 to kick off 10+ web searches with slight variations each time as you go. It's a rapid fire set of operations to collect information for the LLM for analysis.\n",
+    "\n",
+    "_Note: The output below has the `verbosity_level` set to 0 for brevity. If you run this again, you'll get the full details with each step._"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "f2d1ef2e",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/html": [
+       "<pre style=\"white-space:pre;overflow-x:auto;line-height:normal;font-family:Menlo,'DejaVu Sans Mono',consolas,'Courier New',monospace\"><span style=\"color: #d4b702; text-decoration-color: #d4b702\">╭──────────────────────────────────────────────────── </span><span style=\"color: #d4b702; text-decoration-color: #d4b702; font-weight: bold\">New run</span><span style=\"color: #d4b702; text-decoration-color: #d4b702\"> ────────────────────────────────────────────────────╮</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>                                                                                                                 <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">List 20 universities located in the Mid-Atlantic or Northeast that would be good candidates to consider with </span>   <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">top 50 engineering programs and an active baseball team. Visit websites covering the top engineering schools in</span> <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">the Mid-Atlantic and Northeast. Add a rationale for each of the picks in a few words.</span>                           <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>                                                                                                                 <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">╰─ PipelineModel - Qwen/Qwen3-4B-Instruct-2507 ───────────────────────────────────────────────────────────────────╯</span>\n",
+       "</pre>\n"
+      ],
+      "text/plain": [
+       "\u001b[38;2;212;183;2m╭─\u001b[0m\u001b[38;2;212;183;2m───────────────────────────────────────────────────\u001b[0m\u001b[38;2;212;183;2m \u001b[0m\u001b[1;38;2;212;183;2mNew run\u001b[0m\u001b[38;2;212;183;2m \u001b[0m\u001b[38;2;212;183;2m───────────────────────────────────────────────────\u001b[0m\u001b[38;2;212;183;2m─╮\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m                                                                                                                 \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1mList 20 universities located in the Mid-Atlantic or Northeast that would be good candidates to consider with \u001b[0m   \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1mtop 50 engineering programs and an active baseball team. Visit websites covering the top engineering schools in\u001b[0m \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1mthe Mid-Atlantic and Northeast. Add a rationale for each of the picks in a few words.\u001b[0m                           \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m                                                                                                                 \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m╰─\u001b[0m\u001b[38;2;212;183;2m PipelineModel - Qwen/Qwen3-4B-Instruct-2507 \u001b[0m\u001b[38;2;212;183;2m──────────────────────────────────────────────────────────────────\u001b[0m\u001b[38;2;212;183;2m─╯\u001b[0m\n"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "text/markdown": [
+       "Here are 20 universities in the Mid-Atlantic or Northeast with top 50 engineering programs and active baseball teams, along with a brief rationale for each:\n",
+       "\n",
+       "1. **Massachusetts Institute of Technology (MIT)** – Top-ranked engineering program in the Northeast with a strong research focus and active baseball team.\n",
+       "2. **Princeton University** – Renowned for engineering excellence and competitive baseball in the Ivy League.\n",
+       "3. **Columbia University** – Strong engineering programs and a consistently active baseball team.\n",
+       "4. **University of Pennsylvania** – Top-tier engineering and a well-established baseball program in the Ivy League.\n",
+       "5. **Carnegie Mellon University** – Leading in computer and engineering fields with a competitive baseball team.\n",
+       "6. **Cornell University** – Excellent engineering programs and a strong baseball presence.\n",
+       "7. **Northeastern University** – Strong engineering programs and active baseball team.\n",
+       "8. **New York University** – Top engineering programs with a competitive baseball team.\n",
+       "9. **Rensselaer Polytechnic Institute (RPI)** – Strong engineering focus and active baseball in the Liberty League.\n",
+       "10. **Rutgers University** – Top engineering programs and a well-established baseball team.\n",
+       "11. **Lehigh University** – Strong engineering and active baseball in the Patriot League.\n",
+       "12. **University of Connecticut** – Top engineering programs and a competitive baseball team.\n",
+       "13. **University of Massachusetts Amherst** – Solid engineering programs and active baseball.\n",
+       "14. **University of Massachusetts Lowell** – Strong engineering and active baseball in the America East Conference.\n",
+       "15. **Syracuse University** – Top engineering programs and a competitive baseball team.\n",
+       "16. **Boston University** – Strong engineering and active baseball.\n",
+       "17. **Dartmouth College** – Excellent engineering programs and competitive baseball.\n",
+       "18. **Brown University** – Top engineering and active baseball.\n",
+       "19. **Bucknell University** – Solid engineering and active baseball.\n",
+       "20. **University of Rochester** – Strong engineering and competitive baseball.\n",
+       "\n",
+       "These institutions combine top-tier engineering programs with active collegiate baseball teams, making them excellent candidates for students seeking both academic excellence and athletic engagement."
+      ],
+      "text/plain": [
+       "<IPython.core.display.Markdown object>"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "run((\n",
+    "    \"List 20 universities located in the Mid-Atlantic or Northeast that would be good candidates to consider \"\n",
+    "    \"with top 50 engineering programs and an active baseball team. \"\n",
+    "    \"Visit websites covering the top engineering schools in the Mid-Atlantic and Northeast. \"\n",
+    "    \"Add a rationale for each of the picks in a few words. \"\n",
+    "))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "a7533137",
+   "metadata": {},
+   "source": [
+    "Interesting! Looks like quite an impressive list of colleges. This output is the product of taking the results of 10+ web searches and analyzing the results with an LLM. A very practical and realistic use of \"Agentic AI\". This is a great list of strong engineering schools with baseball programs 🎓⚙️⚾.\n",
+    "\n",
+    "Let's narrow it down to the top 5 engineering schools."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 105,
+   "id": "5e261825",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/html": [
+       "<pre style=\"white-space:pre;overflow-x:auto;line-height:normal;font-family:Menlo,'DejaVu Sans Mono',consolas,'Courier New',monospace\"><span style=\"color: #d4b702; text-decoration-color: #d4b702\">╭──────────────────────────────────────────────────── </span><span style=\"color: #d4b702; text-decoration-color: #d4b702; font-weight: bold\">New run</span><span style=\"color: #d4b702; text-decoration-color: #d4b702\"> ────────────────────────────────────────────────────╮</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>                                                                                                                 <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">Read the list below and pick the Top 5 engineering programs.</span>                                                    <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>                                                                                                                 <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">Use the following conversation history to help answer the question above.</span>                                       <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>                                                                                                                 <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">User: List 20 universities located in the Mid-Atlantic or Northeast that would be good candidates to consider </span>  <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">with top 50 engineering programs and an active baseball team. Visit websites covering the top engineering </span>      <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">schools in the Mid-Atlantic and Northeast. Add a rationale for each of the picks in a few words. </span>               <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">Assistant: Here are 20 universities in the Mid-Atlantic or Northeast with top 50 engineering programs and </span>      <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">active baseball teams, along with a brief rationale for each:</span>                                                   <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>                                                                                                                 <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">1. **Massachusetts Institute of Technology (MIT)** – Top-ranked engineering program in the Northeast with a </span>    <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">strong research focus and active baseball team.</span>                                                                 <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">2. **Princeton University** – Renowned for engineering excellence and competitive baseball in the Ivy League.</span>   <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">3. **Columbia University** – Strong engineering programs and a consistently active baseball team.</span>               <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">4. **University of Pennsylvania** – Top-tier engineering and a well-established baseball program in the Ivy </span>    <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">League.</span>                                                                                                         <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">5. **Carnegie Mellon University** – Leading in computer and engineering fields with a competitive baseball </span>     <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">team.</span>                                                                                                           <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">6. **Cornell University** – Excellent engineering programs and a strong baseball presence.</span>                      <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">7. **Northeastern University** – Strong engineering programs and active baseball team.</span>                          <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">8. **New York University** – Top engineering programs with a competitive baseball team.</span>                         <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">9. **Rensselaer Polytechnic Institute (RPI)** – Strong engineering focus and active baseball in the Liberty </span>    <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">League.</span>                                                                                                         <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">10. **Rutgers University** – Top engineering programs and a well-established baseball team.</span>                     <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">11. **Lehigh University** – Strong engineering and active baseball in the Patriot League.</span>                       <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">12. **University of Connecticut** – Top engineering programs and a competitive baseball team.</span>                   <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">13. **University of Massachusetts Amherst** – Solid engineering programs and active baseball.</span>                   <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">14. **University of Massachusetts Lowell** – Strong engineering and active baseball in the America East </span>        <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">Conference.</span>                                                                                                     <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">15. **Syracuse University** – Top engineering programs and a competitive baseball team.</span>                         <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">16. **Boston University** – Strong engineering and active baseball.</span>                                             <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">17. **Dartmouth College** – Excellent engineering programs and competitive baseball.</span>                            <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">18. **Brown University** – Top engineering and active baseball.</span>                                                 <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">19. **Bucknell University** – Solid engineering and active baseball.</span>                                            <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">20. **University of Rochester** – Strong engineering and competitive baseball.</span>                                  <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>                                                                                                                 <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">These institutions combine top-tier engineering programs with active collegiate baseball teams, making them </span>    <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">excellent candidates for students seeking both academic excellence and athletic engagement.</span>                     <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>                                                                                                                 <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span> <span style=\"font-weight: bold\">If the history is irrelevant, forget it and use other tools to answer the question.</span>                             <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>                                                                                                                 <span style=\"color: #d4b702; text-decoration-color: #d4b702\">│</span>\n",
+       "<span style=\"color: #d4b702; text-decoration-color: #d4b702\">╰─ PipelineModel - Qwen/Qwen3-4B-Instruct-2507 ───────────────────────────────────────────────────────────────────╯</span>\n",
+       "</pre>\n"
+      ],
+      "text/plain": [
+       "\u001b[38;2;212;183;2m╭─\u001b[0m\u001b[38;2;212;183;2m───────────────────────────────────────────────────\u001b[0m\u001b[38;2;212;183;2m \u001b[0m\u001b[1;38;2;212;183;2mNew run\u001b[0m\u001b[38;2;212;183;2m \u001b[0m\u001b[38;2;212;183;2m───────────────────────────────────────────────────\u001b[0m\u001b[38;2;212;183;2m─╮\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m                                                                                                                 \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1mRead the list below and pick the Top 5 engineering programs.\u001b[0m                                                    \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m                                                                                                                 \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1mUse the following conversation history to help answer the question above.\u001b[0m                                       \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m                                                                                                                 \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1mUser: List 20 universities located in the Mid-Atlantic or Northeast that would be good candidates to consider \u001b[0m  \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1mwith top 50 engineering programs and an active baseball team. Visit websites covering the top engineering \u001b[0m      \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1mschools in the Mid-Atlantic and Northeast. Add a rationale for each of the picks in a few words. \u001b[0m               \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1mAssistant: Here are 20 universities in the Mid-Atlantic or Northeast with top 50 engineering programs and \u001b[0m      \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1mactive baseball teams, along with a brief rationale for each:\u001b[0m                                                   \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m                                                                                                                 \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m1. **Massachusetts Institute of Technology (MIT)** – Top-ranked engineering program in the Northeast with a \u001b[0m    \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1mstrong research focus and active baseball team.\u001b[0m                                                                 \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m2. **Princeton University** – Renowned for engineering excellence and competitive baseball in the Ivy League.\u001b[0m   \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m3. **Columbia University** – Strong engineering programs and a consistently active baseball team.\u001b[0m               \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m4. **University of Pennsylvania** – Top-tier engineering and a well-established baseball program in the Ivy \u001b[0m    \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1mLeague.\u001b[0m                                                                                                         \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m5. **Carnegie Mellon University** – Leading in computer and engineering fields with a competitive baseball \u001b[0m     \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1mteam.\u001b[0m                                                                                                           \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m6. **Cornell University** – Excellent engineering programs and a strong baseball presence.\u001b[0m                      \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m7. **Northeastern University** – Strong engineering programs and active baseball team.\u001b[0m                          \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m8. **New York University** – Top engineering programs with a competitive baseball team.\u001b[0m                         \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m9. **Rensselaer Polytechnic Institute (RPI)** – Strong engineering focus and active baseball in the Liberty \u001b[0m    \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1mLeague.\u001b[0m                                                                                                         \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m10. **Rutgers University** – Top engineering programs and a well-established baseball team.\u001b[0m                     \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m11. **Lehigh University** – Strong engineering and active baseball in the Patriot League.\u001b[0m                       \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m12. **University of Connecticut** – Top engineering programs and a competitive baseball team.\u001b[0m                   \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m13. **University of Massachusetts Amherst** – Solid engineering programs and active baseball.\u001b[0m                   \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m14. **University of Massachusetts Lowell** – Strong engineering and active baseball in the America East \u001b[0m        \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1mConference.\u001b[0m                                                                                                     \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m15. **Syracuse University** – Top engineering programs and a competitive baseball team.\u001b[0m                         \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m16. **Boston University** – Strong engineering and active baseball.\u001b[0m                                             \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m17. **Dartmouth College** – Excellent engineering programs and competitive baseball.\u001b[0m                            \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m18. **Brown University** – Top engineering and active baseball.\u001b[0m                                                 \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m19. **Bucknell University** – Solid engineering and active baseball.\u001b[0m                                            \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1m20. **University of Rochester** – Strong engineering and competitive baseball.\u001b[0m                                  \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m                                                                                                                 \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1mThese institutions combine top-tier engineering programs with active collegiate baseball teams, making them \u001b[0m    \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1mexcellent candidates for students seeking both academic excellence and athletic engagement.\u001b[0m                     \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m                                                                                                                 \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m \u001b[1mIf the history is irrelevant, forget it and use other tools to answer the question.\u001b[0m                             \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m│\u001b[0m                                                                                                                 \u001b[38;2;212;183;2m│\u001b[0m\n",
+       "\u001b[38;2;212;183;2m╰─\u001b[0m\u001b[38;2;212;183;2m PipelineModel - Qwen/Qwen3-4B-Instruct-2507 \u001b[0m\u001b[38;2;212;183;2m──────────────────────────────────────────────────────────────────\u001b[0m\u001b[38;2;212;183;2m─╯\u001b[0m\n"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    },
+    {
+     "data": {
+      "text/markdown": [
+       "The top 5 engineering programs from the list of universities in the Mid-Atlantic or Northeast with active baseball teams are:\n",
+       "\n",
+       "1. **Massachusetts Institute of Technology (MIT)** – Top-ranked engineering program in the Northeast with a strong research focus and active baseball team.\n",
+       "2. **Princeton University** – Renowned for engineering excellence and competitive baseball in the Ivy League.\n",
+       "3. **Carnegie Mellon University** – Leading in computer and engineering fields with a competitive baseball team.\n",
+       "4. **Cornell University** – Excellent engineering programs and a strong baseball presence.\n",
+       "5. **University of Pennsylvania** – Top-tier engineering and a well-established baseball program in the Ivy League.\n",
+       "\n",
+       "These schools stand out due to their exceptional engineering programs and active collegiate baseball teams, offering a balanced academic and athletic experience."
+      ],
+      "text/plain": [
+       "<IPython.core.display.Markdown object>"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "run(\"Read the list below and pick the Top 5 engineering programs.\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "e6c27f16",
+   "metadata": {},
+   "source": [
+    "# Wrapping up\n",
+    "\n",
+    "This example showed how to run an Agentic Loop that identified the best set of colleges to research further. It's an agent with memory and can be used to build an ongoing dialogue with an agent. Give it a try!"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "local",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.19"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

examples/83_TxtAI_got_skills.ipynb

@@ -0,0 +1,410 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "8d8cb3ad",
+   "metadata": {},
+   "source": [
+    "# TxtAI got skills\n",
+    "\n",
+    "This example will demonstrate how to use `txtai` agents with [`skill.md`](https://agentskills.io/specification) files.\n",
+    "\n",
+    "We'll setup a `skill.md` file with details on how to use TxtAI and run a series of agent requests.\n",
+    "\n",
+    "Let's get started!"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "00a5a0ea",
+   "metadata": {},
+   "source": [
+    "# Install dependencies\n",
+    "\n",
+    "Install `txtai` and all dependencies."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "0ff7aa19",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%%capture\n",
+    "!pip install git+https://github.com/neuml/txtai#egg=txtai[agent]"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "98bcd5c3",
+   "metadata": {},
+   "source": [
+    "# Define a `skill.md` file\n",
+    "\n",
+    "Next, we'll create our `skill.md` file. This file has examples on how to build embeddings databases, how to use re-ranker pipelines, RAG pipelines and more.\n",
+    "\n",
+    "The upside of a `skill.md` file vs an `agents.md` file is that it can be dynamically added to the agent context. The description helps the agent decide if the `skill` is necessary given the request. Think of it like a dynamic knowledge base that's easy to modify."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "49179643",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%%writefile skill.md\n",
+    "---\n",
+    "name: txtai\n",
+    "description: Examples on how to build txtai embeddings databases, txtai RAG pipelines, txtai reranker pipelines and txtai translation pipelines\n",
+    "---\n",
+    "\n",
+    "# Build an embeddings database\n",
+    "\n",
+    "```python\n",
+    "from txtai import Embeddings\n",
+    "\n",
+    "# Create embeddings model, backed by sentence-transformers & transformers\n",
+    "embeddings = Embeddings(path=\"sentence-transformers/nli-mpnet-base-v2\")\n",
+    "\n",
+    "data = [\n",
+    "  \"US tops 5 million confirmed virus cases\",\n",
+    "  \"Canada's last fully intact ice shelf has suddenly collapsed, \" +\n",
+    "  \"forming a Manhattan-sized iceberg\",\n",
+    "  \"Beijing mobilises invasion craft along coast as Taiwan tensions escalate\",\n",
+    "  \"The National Park Service warns against sacrificing slower friends \" +\n",
+    "  \"in a bear attack\",\n",
+    "  \"Maine man wins $1M from $25 lottery ticket\",\n",
+    "  \"Make huge profits without work, earn up to $100,000 a day\"\n",
+    "]\n",
+    "\n",
+    "# Index the list of text\n",
+    "embeddings.index(data)\n",
+    "```\n",
+    "\n",
+    "# Search an embeddings database\n",
+    "\n",
+    "```python\n",
+    "embeddings.search(\"Search query\")\n",
+    "```\n",
+    "\n",
+    "# Build a RAG pipeline\n",
+    "\n",
+    "```python\n",
+    "from txtai import Embeddings, RAG\n",
+    "\n",
+    "# Input data\n",
+    "data = [\n",
+    "  \"US tops 5 million confirmed virus cases\",\n",
+    "  \"Canada's last fully intact ice shelf has suddenly collapsed, \" +\n",
+    "  \"forming a Manhattan-sized iceberg\",\n",
+    "  \"Beijing mobilises invasion craft along coast as Taiwan tensions escalate\",\n",
+    "  \"The National Park Service warns against sacrificing slower friends \" +\n",
+    "  \"in a bear attack\",\n",
+    "  \"Maine man wins $1M from $25 lottery ticket\",\n",
+    "  \"Make huge profits without work, earn up to $100,000 a day\"\n",
+    "]\n",
+    "\n",
+    "# Build embeddings index\n",
+    "embeddings = Embeddings(content=True)\n",
+    "embeddings.index(data)\n",
+    "\n",
+    "# Create the RAG pipeline\n",
+    "rag = RAG(embeddings, \"Qwen/Qwen3-0.6B\", template=\"\"\"\n",
+    "  Answer the following question using the provided context.\n",
+    "\n",
+    "  Question:\n",
+    "  {question}\n",
+    "\n",
+    "  Context:\n",
+    "  {context}\n",
+    "\"\"\")\n",
+    "\n",
+    "# Run RAG pipeline\n",
+    "rag(\"What was won?\")\n",
+    "```\n",
+    "\n",
+    "# Translate text from English into French\n",
+    "\n",
+    "```python\n",
+    "from txtai.pipeline import Translation\n",
+    "\n",
+    "# Create and run pipeline\n",
+    "translate = Translation()\n",
+    "translate(\"This is a test translation\", \"fr\")\n",
+    "```\n",
+    "\n",
+    "# Re-ranker pipeline\n",
+    "\n",
+    "```python\n",
+    "from txtai import Embeddings\n",
+    "from txtai.pipeline import Reranker, Similarity\n",
+    "\n",
+    "# Embeddings instance\n",
+    "embeddings = Embeddings()\n",
+    "embeddings.load(provider=\"huggingface-hub\", container=\"neuml/txtai-wikipedia\")\n",
+    "\n",
+    "# Similarity instance\n",
+    "similarity = Similarity(path=\"colbert-ir/colbertv2.0\", lateencode=True)\n",
+    "\n",
+    "# Reranking pipeline\n",
+    "reranker = Reranker(embeddings, similarity)\n",
+    "reranker(\"Tell me about AI\")\n",
+    "```"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "4ed26969",
+   "metadata": {},
+   "source": [
+    "# Query for TxtAI questions\n",
+    "\n",
+    "Now, let's try this out and see if the LLM is smart enough to use the defined skill vs. going out on the web.\n",
+    "\n",
+    "Let's setup the scaffolding code to create and run an agent. We'll use a [Qwen3 4B non-thinking LLM](https://huggingface.co/Qwen/Qwen3-4B-Instruct-2507) as the agent's model. We'll add the `websearch` and `webview` tools to the agent along with the `skill.md` file previously created.\n",
+    "\n",
+    "Additionally, we'll add a sliding window of the last 2 responses as \"agent memory\". This will help create a rolling dialogue. "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "b6a1aaf1",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from txtai import Agent\n",
+    "from IPython.display import display, Markdown\n",
+    "\n",
+    "def run(query, reset=False):\n",
+    "    answer = agent(query, maxlength=50000, reset=reset)\n",
+    "    display(Markdown(answer))\n",
+    "\n",
+    "agent = Agent(\n",
+    "    model=\"Qwen/Qwen3-4B-Instruct-2507\",\n",
+    "    tools=[\"websearch\", \"webview\", \"skill.md\"],\n",
+    "    memory=2,\n",
+    "    verbosity_level=0\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "19302f54",
+   "metadata": {},
+   "source": [
+    "First, we'll ask how to build a TxtAI embeddings database."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "id": "121d0b8a",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/markdown": [
+       "To create a txtai embeddings program that indexes data, use the following Python code:\n",
+       "\n",
+       "```python\n",
+       "from txtai import Embeddings\n",
+       "\n",
+       "# Create embeddings model using sentence-transformers\n",
+       "embeddings = Embeddings(path=\"sentence-transformers/nli-mpnet-base-v2\")\n",
+       "\n",
+       "# Sample data to index\n",
+       "data = [\n",
+       "  \"US tops 5 million confirmed virus cases\",\n",
+       "  \"Canada's last fully intact ice shelf has suddenly collapsed, \" +\n",
+       "  \"forming a Manhattan-sized iceberg\",\n",
+       "  \"Beijing mobilises invasion craft along coast as Taiwan tensions escalate\",\n",
+       "  \"The National Park Service warns against sacrificing slower friends \" +\n",
+       "  \"in a bear attack\",\n",
+       "  \"Maine man wins $1M from $25 lottery ticket\",\n",
+       "  \"Make huge profits without work, earn up to $100,000 a day\"\n",
+       "]\n",
+       "\n",
+       "# Index the data\n",
+       "embeddings.index(data)\n",
+       "```\n",
+       "\n",
+       "This program initializes an embeddings database using the `sentence-transformers/nli-mpnet-base-v2` model and indexes a list of text data. The indexed data can later be searched or used in other applications like retrieval, RAG, or translation."
+      ],
+      "text/plain": [
+       "<IPython.core.display.Markdown object>"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "run(\"Write a txtai embeddings program that indexes data\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "f723e07f",
+   "metadata": {},
+   "source": [
+    "The Agent pulled the correct section from the `skill.md` file. Now, let's look for an example on how to use a re-ranker pipeline."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "id": "221a590c",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/markdown": [
+       "```python\n",
+       "from txtai import Embeddings\n",
+       "from txtai.pipeline import Reranker, Similarity\n",
+       "\n",
+       "# Embeddings instance\n",
+       "embeddings = Embeddings()\n",
+       "embeddings.load(provider=\"huggingface-hub\", container=\"neuml/txtai-wikipedia\")\n",
+       "\n",
+       "# Similarity instance\n",
+       "similarity = Similarity(path=\"colbert-ir/colbertv2.0\", lateencode=True)\n",
+       "\n",
+       "# Reranking pipeline\n",
+       "reranker = Reranker(embeddings, similarity)\n",
+       "reranker(\"Tell me about AI\")\n",
+       "```"
+      ],
+      "text/plain": [
+       "<IPython.core.display.Markdown object>"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "run(\"Write a txtai re-ranker pipeline\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "4af3a6b6",
+   "metadata": {},
+   "source": [
+    "Great! This also works. But remember we have access to an LLM here. It doesn't have to blindly just pull the text. Let's ask it to modify the last example."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "id": "c9f1eed4",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/markdown": [
+       "```python\n",
+       "from txtai import Embeddings\n",
+       "from txtai.pipeline import Reranker, Similarity\n",
+       "\n",
+       "# Embeddings instance\n",
+       "embeddings = Embeddings()\n",
+       "embeddings.load(provider=\"huggingface-hub\", container=\"neuml/txtai-wikipedia\")\n",
+       "\n",
+       "# Similarity instance with a different reranker model and lateencode disabled\n",
+       "similarity = Similarity(path=\"BAAI/bge-reranker-base\", lateencode=False)\n",
+       "\n",
+       "# Reranking pipeline\n",
+       "reranker = Reranker(embeddings, similarity)\n",
+       "reranker(\"Tell me about AI\")\n",
+       "```"
+      ],
+      "text/plain": [
+       "<IPython.core.display.Markdown object>"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "run(\"Update the similarity path to use another reranker model. Disable lateencode.\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "4a3bb21e",
+   "metadata": {},
+   "source": [
+    "Notice it just edited the code with a different reranker model and even added a comment to note this change.\n",
+    "\n",
+    "We can also clear the rolling dialogue and start fresh. "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "id": "ad53f7a8",
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/markdown": [
+       "```python\n",
+       "from txtai.pipeline import Translation\n",
+       "\n",
+       "# Create and run pipeline for English to Spanish translation\n",
+       "translate = Translation()\n",
+       "translated_text = translate(\"This is a test translation\", \"es\")\n",
+       "print(translated_text)\n",
+       "```"
+      ],
+      "text/plain": [
+       "<IPython.core.display.Markdown object>"
+      ]
+     },
+     "metadata": {},
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "run(\"Write a txtai program that translate text from English to Spanish\", reset=True)"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "007daee0",
+   "metadata": {},
+   "source": [
+    "# Wrapping up\n",
+    "\n",
+    "This example shows how to add a `skill.md` file to `txtai` agents. Go give it a try!"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "local",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.19"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

examples/85_Progressive_Distillation.ipynb

@@ -0,0 +1,337 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "40f124c0",
+   "metadata": {},
+   "source": [
+    "# Progressive Distillation\n",
+    "\n",
+    "Now that almost everyone has thought about or is actively integrating AI workflows into their projects, some might ask is this all worth the cost? Many think the current economics of the AI space don't scale and that there will be upward price movement. Others still might not be comfortable with sending their data to remote services for processing. Then there is the crowd that wants to deploy models in small spaces with limited compute.\n",
+    "\n",
+    "Are there ways we can deploy small models locally and run at a lower cost? Yes with [Knowledge Distillation](https://en.wikipedia.org/wiki/Knowledge_distillation). Knowledge distillation can get a bad rap due to it's questionable use in training some Large Language Models (LLMs). But it's a perfectly valid way to transfer performance from a larger model to a smaller one. Especially when both models are yours and/or open. \n",
+    "\n",
+    "This notebook will explore progressive distillation which is a technique to incrementally transfer knowledge from a series of larger teacher models into a smaller student."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "f4aad867",
+   "metadata": {},
+   "source": [
+    "# Install dependencies\n",
+    "\n",
+    "Install `txtai` and all dependencies."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 1,
+   "id": "16e2c760",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%%capture\n",
+    "!pip install git+https://github.com/neuml/txtai#egg=txtai[pipeline-train] datasets"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "e3a23082",
+   "metadata": {},
+   "source": [
+    "# Setup the Training Pipeline\n",
+    "\n",
+    "The first step we need to do is setup up the training pipeline. We'll use the Hugging Face Training framework to build a series of models.\n",
+    "\n",
+    "The following code establishes a `train` method, `test` method and loads the classification training data."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "ea7ee576",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from datasets import load_dataset\n",
+    "from transformers import AutoModelForSequenceClassification, AutoTokenizer\n",
+    "from txtai.pipeline import HFTrainer, Labels\n",
+    "\n",
+    "def train(teacher, student, distillation, **kwargs):\n",
+    "    trainer = HFTrainer()\n",
+    "\n",
+    "    model = AutoModelForSequenceClassification.from_pretrained(student, trust_remote_code=True)\n",
+    "    tokenizer = AutoTokenizer.from_pretrained(student, trust_remote_code=True)\n",
+    "\n",
+    "    return trainer(\n",
+    "        (model, tokenizer),\n",
+    "        ds[\"train\"],\n",
+    "        columns=(\"sentence\", \"label\"),\n",
+    "        maxlength=maxlength,\n",
+    "        teacher=teacher,\n",
+    "        distillation=distillation,\n",
+    "        **kwargs\n",
+    "    )\n",
+    "\n",
+    "def test(model):\n",
+    "    labels = Labels(model, dynamic=False, trust_remote_code=True)\n",
+    "\n",
+    "    # Determine accuracy on test set\n",
+    "    results = [row[\"label\"] == labels(row[\"sentence\"], max_length=maxlength)[0][0] for row in ds[\"validation\"]]\n",
+    "    print(sum(results) / len(ds[\"validation\"]))\n",
+    "\n",
+    "# Hugging Face dataset\n",
+    "ds = load_dataset(\"nyu-mll/glue\", \"sst2\")\n",
+    "maxlength = 128"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "732fad7d",
+   "metadata": {},
+   "source": [
+    "# Progressive Distillation\n",
+    "\n",
+    "We're going to build a [bert-hash-femto](https://huggingface.co/NeuML/bert-hash-femto) classifier which is a extremely small 250K parameter model. This model was pretrained using the same recipe as [BERT](https://arxiv.org/abs/1810.04805). The paper [Well-Read Students Learn Better: On the Importance of Pre-training Compact Models](https://arxiv.org/abs/1908.08962) established that small models perform better with Knowledge Distillation tasks when they are pretrained. [This article](https://huggingface.co/blog/NeuML/bert-hash-embeddings) is also a good source for more information on the topic.\n",
+    "\n",
+    "The next series of steps will do the following:\n",
+    "\n",
+    "- Train a [BERT Mini](https://huggingface.co/google/bert_uncased_L-4_H-256_A-4) model using a [BERT Large](https://huggingface.co/assemblyai/bert-large-uncased-sst2) teacher\n",
+    "- Train a [BERT Tiny](https://huggingface.co/google/bert_uncased_L-2_H-128_A-2) model using the BERT Mini teacher above\n",
+    "- Train a [BERT Hash Femto](https://huggingface.co/NeuML/bert-hash-femto) model using the BERT Tiny teacher above\n",
+    "\n",
+    "This path goes from a 11M parameter model -> 4.4M parameter model -> 250K parameter model\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "id": "43baccd6",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "[transformers] \u001b[1mBertForSequenceClassification LOAD REPORT\u001b[0m from: google/bert_uncased_L-4_H-256_A-4\n",
+      "Key                                        | Status     | \n",
+      "-------------------------------------------+------------+-\n",
+      "cls.predictions.transform.dense.weight     | UNEXPECTED | \n",
+      "cls.predictions.decoder.bias               | UNEXPECTED | \n",
+      "cls.predictions.transform.LayerNorm.bias   | UNEXPECTED | \n",
+      "cls.predictions.decoder.weight             | UNEXPECTED | \n",
+      "cls.seq_relationship.bias                  | UNEXPECTED | \n",
+      "cls.seq_relationship.weight                | UNEXPECTED | \n",
+      "cls.predictions.transform.dense.bias       | UNEXPECTED | \n",
+      "cls.predictions.bias                       | UNEXPECTED | \n",
+      "cls.predictions.transform.LayerNorm.weight | UNEXPECTED | \n",
+      "classifier.weight                          | MISSING    | \n",
+      "classifier.bias                            | MISSING    | \n",
+      "\n",
+      "Notes:\n",
+      "- UNEXPECTED:\tcan be ignored when loading from different task/architecture; not ok if you expect identical arch.\n",
+      "- MISSING:\tthose params were newly initialized because missing from the checkpoint. Consider training on your downstream task.\n"
+     ]
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "0.875\n"
+     ]
+    }
+   ],
+   "source": [
+    "test(train(\n",
+    "    \"assemblyai/bert-large-uncased-sst2\",\n",
+    "    \"google/bert_uncased_L-4_H-256_A-4\",\n",
+    "    (1.0, 1.0),\n",
+    "    learning_rate=1e-4,\n",
+    "    num_train_epochs=5,\n",
+    "    per_device_train_batch_size=32,\n",
+    "    output_dir=\"bert-mini-sst2\"\n",
+    "))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "id": "9ba3158e",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "[transformers] \u001b[1mBertForSequenceClassification LOAD REPORT\u001b[0m from: google/bert_uncased_L-2_H-128_A-2\n",
+      "Key                                        | Status     | \n",
+      "-------------------------------------------+------------+-\n",
+      "cls.predictions.transform.dense.weight     | UNEXPECTED | \n",
+      "cls.predictions.transform.LayerNorm.bias   | UNEXPECTED | \n",
+      "cls.seq_relationship.weight                | UNEXPECTED | \n",
+      "cls.seq_relationship.bias                  | UNEXPECTED | \n",
+      "cls.predictions.transform.dense.bias       | UNEXPECTED | \n",
+      "cls.predictions.bias                       | UNEXPECTED | \n",
+      "cls.predictions.transform.LayerNorm.weight | UNEXPECTED | \n",
+      "classifier.weight                          | MISSING    | \n",
+      "classifier.bias                            | MISSING    | \n",
+      "\n",
+      "Notes:\n",
+      "- UNEXPECTED:\tcan be ignored when loading from different task/architecture; not ok if you expect identical arch.\n",
+      "- MISSING:\tthose params were newly initialized because missing from the checkpoint. Consider training on your downstream task.\n"
+     ]
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "0.8325688073394495\n"
+     ]
+    }
+   ],
+   "source": [
+    "test(train(\n",
+    "    \"bert-mini-sst2\",\n",
+    "    \"google/bert_uncased_L-2_H-128_A-2\",\n",
+    "    (1.0, 1.0),\n",
+    "    learning_rate=1e-4,\n",
+    "    num_train_epochs=5,\n",
+    "    per_device_train_batch_size=32,\n",
+    "    output_dir=\"bert-tiny-sst2\"\n",
+    "))"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 9,
+   "id": "e98d950e",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "[transformers] \u001b[1mBertHashForSequenceClassification LOAD REPORT\u001b[0m from: neuml/bert-hash-femto\n",
+      "Key                      | Status  | \n",
+      "-------------------------+---------+-\n",
+      "bert.pooler.dense.bias   | MISSING | \n",
+      "bert.pooler.dense.weight | MISSING | \n",
+      "classifier.bias          | MISSING | \n",
+      "classifier.weight        | MISSING | \n",
+      "\n",
+      "Notes:\n",
+      "- MISSING:\tthose params were newly initialized because missing from the checkpoint. Consider training on your downstream task.\n"
+     ]
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "0.8084862385321101\n"
+     ]
+    }
+   ],
+   "source": [
+    "test(train(\n",
+    "    \"bert-tiny-sst2\",\n",
+    "    \"neuml/bert-hash-femto\",\n",
+    "    (1.0, 1.0),\n",
+    "    learning_rate=3e-4,\n",
+    "    num_train_epochs=5,\n",
+    "    per_device_train_batch_size=32,\n",
+    "    output_dir=\"bert-femto-sst2\"\n",
+    "))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "5f1c94e7",
+   "metadata": {},
+   "source": [
+    "Let's look at the performance. The 11M parameter model registered an accuracy score of `0.8750` on the SST2 dev set. Then the 4M parameter model scored `0.8326` and finally the tiny femto model scored `0.8085`. As we can see each score got progressively worse but each model has less capability. Let's train a femto model directly to compare."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "id": "b89d8292",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "[transformers] \u001b[1mBertHashForSequenceClassification LOAD REPORT\u001b[0m from: neuml/bert-hash-femto\n",
+      "Key                      | Status  | \n",
+      "-------------------------+---------+-\n",
+      "bert.pooler.dense.bias   | MISSING | \n",
+      "bert.pooler.dense.weight | MISSING | \n",
+      "classifier.bias          | MISSING | \n",
+      "classifier.weight        | MISSING | \n",
+      "\n",
+      "Notes:\n",
+      "- MISSING:\tthose params were newly initialized because missing from the checkpoint. Consider training on your downstream task.\n"
+     ]
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "0.801605504587156\n"
+     ]
+    }
+   ],
+   "source": [
+    "test(train(\n",
+    "    None,\n",
+    "    \"neuml/bert-hash-femto\",\n",
+    "    None,\n",
+    "    learning_rate=3e-4,\n",
+    "    num_train_epochs=5,\n",
+    "    per_device_train_batch_size=32,\n",
+    "    output_dir=\"bert-femto-sst2\"\n",
+    "))"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "4864b8d2",
+   "metadata": {},
+   "source": [
+    "This scored `0.8016` a sizable shift down from the progressively distilled model. Now keep in mind the femto model only has `250K` parameters but we gave it a sizable accuracy boost within it's capabilities."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "e4e04226",
+   "metadata": {},
+   "source": [
+    "# Wrapping up\n",
+    "\n",
+    "This example showed how progressive distillation can boost overall model performance, especially for tiny models. Incrementally compressing knowledge into a series of smaller subsets enabled the final model to learn more efficiently than directly training the model on the dataset without distillation. Put progressive distillation in the toolkit when working with tiny models!"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "local",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.10.20"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

examples/agent_quickstart.py

@@ -0,0 +1,80 @@
+"""
+Agent Quick Start
+Easy to use way to get started with AI Agents.
+
+TxtAI has many example notebooks covering everything the framework provides
+Examples: https://neuml.github.io/txtai/examples
+
+Install TxtAI
+  pip install txtai[agent]
+"""
+
+# pylint: disable=C0103
+from datetime import datetime
+from txtai import Agent
+
+# Step 1: Define your Embeddings database
+#
+# Replace provider/container with a path to a local Embeddings database
+# See RAG Quickstart for an example of building your own custom database
+embeddings = {
+    "name": "wikipedia",
+    "description": "Searches a Wikipedia database",
+    # "path": "path to your embeddings database"
+    "provider": "huggingface-hub",
+    "container": "neuml/txtai-wikipedia",
+}
+
+
+# Step 2: Define other tools
+#
+# Add any Python function. Just need to describe it.
+def today() -> str:
+    """
+    Gets the current date and time
+
+    Returns:
+        current date and time
+    """
+
+    return datetime.today().isoformat()
+
+
+# Step 3: Create a list of available tools
+#
+# Combine defined tools with default tools
+tools = [
+    embeddings,  # Embeddings database with YOUR data
+    today,  # Python function
+    "websearch",  # Runs a websearch using default engine
+    "webview",  # Loads a web page
+]
+
+# Step 4: Set LLM configuration
+#
+# LLM APIs
+#  model = "gpt-5.1"
+#  model = "claude-opus-4-5-20251101"
+#  model = "gemini/gemini-3-pro-preview"
+#
+# Local LLMs
+#  model = "ollama/gpt-oss
+#  model = "openai/gpt-oss-20b"
+#  model = "unsloth/gpt-oss-20b-GGUF/gpt-oss-20b-Q4_K_M.gguf"
+#
+# Pass multiple options as a dictionary
+#  model = {
+#    "path": "unsloth/Qwen3-30B-A3B-Instruct-2507-GGUF/Qwen3-30B-A3B-Instruct-2507-Q4_K_M.gguf",
+#    "n_ctx": 25000
+#  }
+model = "Qwen/Qwen3-4B-Instruct-2507"
+
+# Step 4: Create an Agent
+#
+# Set LLM, tools and other configuration
+# See this for more options: https://huggingface.co/docs/smolagents/reference/agents#agents
+agent = Agent(model=model, tools=tools, max_steps=10)
+
+print(agent("Tell me about the Roman Empire"))
+print(agent("What is the current date?"))
+print(agent("Get the 5 top news stories for today", maxlength=25000))

examples/baseball.py

@@ -15,7 +15,7 @@ import numpy as np
 import pandas as pd
 import streamlit as st
 
-from txtai.embeddings import Embeddings
+from txtai import Embeddings
 
 
 class Stats:
@@ -38,7 +38,7 @@ class Stats:
         self.names = self.loadnames()
 
         # Build index
-        self.vectors, self.data, self.embeddings = self.index()
+        self.vectors, self.data, self.maxyear, self.embeddings = self.index()
 
     def loadcolumns(self):
         """
@@ -122,16 +122,12 @@ class Stats:
         # Build data dictionary
         vectors = {f'{row["yearID"]}{row["playerID"]}': self.transform(row) for _, row in self.stats.iterrows()}
         data = {f'{row["yearID"]}{row["playerID"]}': dict(row) for _, row in self.stats.iterrows()}
+        maxyear = max(row["yearID"] for _, row in self.stats.iterrows())
 
-        embeddings = Embeddings(
-            {
-                "transform": self.transform,
-            }
-        )
-
+        embeddings = Embeddings({"transform": Stats.transform})
         embeddings.index((uid, vectors[uid], None) for uid in vectors)
 
-        return vectors, data, embeddings
+        return vectors, data, maxyear, embeddings
 
     def metrics(self, name):
         """
@@ -159,13 +155,14 @@ class Stats:
 
         return range(1871, datetime.datetime.today().year), 1950, None
 
-    def search(self, name=None, year=None, row=None, limit=10):
+    def search(self, name=None, year=None, window=None, row=None, limit=10):
         """
         Runs an embeddings search. This method takes either a player-year or stats row as input.
 
         Args:
             name: player name to search
             year: year to search
+            window: limit to window recent seasons
             row: row of stats to search
             limit: max results to return
 
@@ -183,13 +180,17 @@ class Stats:
 
         results, ids = [], set()
         if query is not None:
-            for uid, _ in self.embeddings.search(query, limit * 5):
+            candidates = limit * 100 if window else limit * 5
+            for uid, _ in self.embeddings.search(query, candidates):
                 # Only add unique players
                 if uid[4:] not in ids:
                     result = self.data[uid].copy()
-                    result["link"] = f'https://www.baseball-reference.com/players/{result["nameLast"].lower()[0]}/{result["bbrefID"]}.shtml'
-                    results.append(result)
-                    ids.add(uid[4:])
+
+                    # Add first player if this is a player comparison. Limit results to window, if necessary
+                    if (not ids and not row) or not window or result["yearID"] > self.maxyear - window:
+                        result["link"] = f'https://www.baseball-reference.com/players/{result["nameLast"].lower()[0]}/{result["bbrefID"]}.shtml'
+                        results.append(result)
+                        ids.add(uid[4:])
 
                     if len(ids) >= limit:
                         break
@@ -253,10 +254,10 @@ class Batting(Stats):
         ]
 
     def load(self):
-        # Retrieve raw data from GitHub
-        players = pd.read_csv("https://raw.githubusercontent.com/chadwickbureau/baseballdatabank/master/core/People.csv")
-        batting = pd.read_csv("https://raw.githubusercontent.com/chadwickbureau/baseballdatabank/master/core/Batting.csv")
-        fielding = pd.read_csv("https://raw.githubusercontent.com/chadwickbureau/baseballdatabank/master/core/Fielding.csv")
+        # Retrieve raw data
+        players = pd.read_csv("https://hf.co/datasets/neuml/baseballdata/resolve/main/People.csv")
+        batting = pd.read_csv("https://hf.co/datasets/neuml/baseballdata/resolve/main/Batting.csv")
+        fielding = pd.read_csv("https://hf.co/datasets/neuml/baseballdata/resolve/main/Fielding.csv")
 
         # Merge player data in
         batting = pd.merge(players, batting, how="inner", on=["playerID"])
@@ -387,9 +388,9 @@ class Pitching(Stats):
         ]
 
     def load(self):
-        # Retrieve raw data from GitHub
-        players = pd.read_csv("https://raw.githubusercontent.com/chadwickbureau/baseballdatabank/master/core/People.csv")
-        pitching = pd.read_csv("https://raw.githubusercontent.com/chadwickbureau/baseballdatabank/master/core/Pitching.csv")
+        # Retrieve raw data
+        players = pd.read_csv("https://hf.co/datasets/neuml/baseballdata/resolve/main/People.csv")
+        pitching = pd.read_csv("https://hf.co/datasets/neuml/baseballdata/resolve/main/Pitching.csv")
 
         # Merge player data in
         pitching = pd.merge(players, pitching, how="inner", on=["playerID"])
@@ -435,13 +436,14 @@ class Application:
         Runs a Streamlit application.
         """
 
-        st.title("⚾ Baseball Statistics")
+        st.set_page_config(layout="wide", page_title="Baseball Stats", page_icon="⚾")
+        st.title("⚾ Baseball Stats")
         st.markdown(
             """
-            This application finds the best matching historical players using vector search with [txtai](https://github.com/neuml/txtai).
-            Raw data is from the [Baseball Databank](https://github.com/chadwickbureau/baseballdatabank) GitHub project. Read [this
+            This application finds the best matching players using vector search with [txtai](https://github.com/neuml/txtai).
+            Raw data is from the [Lahman Baseball Database](https://sabr.org/lahman-database/). Read [this
             article](https://medium.com/neuml/explore-baseball-history-with-vector-search-5778d98d6846) for more details.
-        """
+            """
         )
 
         player, search = st.tabs(["Player", "Search"])
@@ -471,6 +473,9 @@ class Application:
         # Player name
         name = self.name(stats.names, params.get("name"))
 
+        # Limit player-year comparisons using this window
+        window = self.window(params.get("window"), "window")
+
         # Player metrics
         active, best, metrics = stats.metrics(name)
 
@@ -482,13 +487,16 @@ class Application:
             self.chart(category, metrics)
 
         # Run search
-        results = stats.search(name, year)
+        results = stats.search(name, year, window)
 
         # Display results
         self.table(results, ["link", "nameFirst", "nameLast", "teamID"] + stats.columns[1:])
 
-        # Save parameters
-        st.experimental_set_query_params(category=category, name=name, year=year)
+        # Save query parameters
+        st.query_params.clear()
+        for key, value in [("category", category), ("name", name), ("window", window), ("year", year)]:
+            if value:
+                st.query_params[key] = value
 
     def search(self):
         """
@@ -498,6 +506,10 @@ class Application:
         st.markdown("Find players with similar statistics.")
 
         stats, category = None, self.category("Batting", "searchcategory")
+
+        # Limit player-year comparisons using this window
+        window = self.window(None, "searchwindow")
+
         with st.form("search"):
             if category == "Batting":
                 stats, columns = self.batting, self.batting.columns[:-6]
@@ -510,7 +522,7 @@ class Application:
             submitted = st.form_submit_button("Search")
             if submitted:
                 # Run search
-                results = stats.search(row=inputs.to_dict(orient="records")[0])
+                results = stats.search(window=window, row=inputs.to_dict(orient="records")[0])
 
                 # Display table
                 self.table(results, ["link", "nameFirst", "nameLast", "teamID"] + stats.columns[1:])
@@ -523,18 +535,18 @@ class Application:
             parameters
         """
 
-        # Get parameters
-        params = st.experimental_get_query_params()
-        params = {x: params[x][0] for x in params}
+        # Get query parameters
+        params = {x: st.query_params.get(x) for x in ["category", "name", "window", "year"]}
 
         # Sync parameters with session state
         if all(x in st.session_state for x in ["category", "name", "year"]):
             # Copy session year if category and name are unchanged
             params["year"] = str(st.session_state["year"]) if all(params.get(x) == st.session_state[x] for x in ["category", "name"]) else None
 
-            # Copy category and name from session state
+            # Copy category, name and window from session state
             params["category"] = st.session_state["category"]
             params["name"] = st.session_state["name"]
+            params["window"] = st.session_state["window"]
 
         return params
 
@@ -559,6 +571,29 @@ class Application:
         # Radio box component
         return st.radio("Stat", categories, index=default, horizontal=True, key=key)
 
+    def window(self, window, key):
+        """
+        Limit results to last N seasons.
+
+        Args:
+            window: limit to window seasons
+            key: widget key
+
+        Returns:
+            window component
+        """
+
+        # Get window size
+        window = st.text_input("Limit to last N seasons", value=window, key=key)
+
+        # Clear invalid input
+        if window and (not window.isnumeric() or int(window) < 1):
+            st.error("Window must be a number greater or equal to 1")
+            return None
+
+        # Convert to int
+        return int(window) if window else window
+
     def name(self, names, name):
         """
         Builds name input widget.
@@ -627,7 +662,7 @@ class Application:
         chart = (chart + rule).encode(y=alt.Y(title=metric)).properties(height=200).configure_axis(grid=False)
 
         # Draw chart
-        st.altair_chart(chart + rule, theme="streamlit", use_container_width=True)
+        st.altair_chart(chart + rule, theme="streamlit", width="stretch")
 
     def table(self, results, columns):
         """
@@ -643,7 +678,7 @@ class Application:
                 results,
                 column_order=columns,
                 column_config={
-                    "link": st.column_config.LinkColumn("Link", width="small"),
+                    "link": st.column_config.LinkColumn("Link", width="small", display_text=":material/open_in_new:"),
                     "yearID": st.column_config.NumberColumn("Year", format="%d"),
                     "nameFirst": "First",
                     "nameLast": "Last",

examples/benchmarks.py

@@ -23,9 +23,10 @@ from elasticsearch import Elasticsearch
 from elasticsearch.helpers import bulk
 from pytrec_eval import RelevanceEvaluator
 from rank_bm25 import BM25Okapi
+from tqdm.auto import tqdm
 
 from txtai.embeddings import Embeddings
-from txtai.pipeline import LLM, RAG, Tokenizer
+from txtai.pipeline import LLM, RAG, Similarity, Tokenizer
 from txtai.scoring import ScoringFactory
 
 
@@ -111,8 +112,16 @@ class Index:
         Iterates over the dataset yielding a row at a time for indexing.
         """
 
-        with open(f"{self.path}/corpus.jsonl", encoding="utf-8") as f:
-            for line in f:
+        # Data file
+        path = f"{self.path}/corpus.jsonl"
+
+        # Get total count
+        with open(path, encoding="utf-8") as f:
+            total = sum(1 for _ in f)
+
+        # Yield data
+        with open(path, encoding="utf-8") as f:
+            for line in tqdm(f, total=total):
                 row = json.loads(line)
                 text = f'{row["title"]}. {row["text"]}' if row.get("title") else row["text"]
                 if text:
@@ -266,6 +275,70 @@ class Score(Index):
         return scoring
 
 
+class Similar(Index):
+    """
+    Search data using a similarity pipeline.
+    """
+
+    def index(self):
+        # Load similarity pipeline
+        model = Similarity(**self.readconfig("similar", {}))
+
+        # Get datasets
+        data = list(self.rows())
+        ids = [x[0] for x in data]
+        texts = [x[1] for x in data]
+
+        # Encode texts
+        data = model.encode(texts, "data")
+
+        return (ids, data, model)
+
+    def search(self, queries, limit):
+        # Unpack backend
+        ids, data, model = self.backend
+
+        # Run model inference
+        results = []
+        for result in model(queries, data, limit=limit):
+            results.append([(ids[x], score) for x, score in result])
+
+        return results
+
+
+class Rerank(Embed):
+    """
+    Embeddings index using txtai combined with a similarity pipeline
+    """
+
+    def index(self):
+        # Build embeddings index
+        embeddings = super().index()
+
+        # Combine similar pipeline with embeddings
+        model = Similar(self.path, self.config, self.output, self.refresh)
+        return model.index() + (embeddings,)
+
+    def search(self, queries, limit):
+        # Unpack backend
+        ids, data, model, embeddings = self.backend
+
+        # Run initial query
+        indices = []
+        for r in embeddings.batchsearch(queries, limit * 10):
+            indices.append({x: ids.index(uid) for x, (uid, _) in enumerate(r)})
+
+        # Run model inference
+        results = []
+        for x, query in enumerate(queries):
+            queue = data[list(indices[x].values())]
+            if len(queue) > 0:
+                result = model(query, queue, limit=limit)
+                results.append([(ids[indices[x][i]], score) for i, score in result])
+
+        return results
+
+
 class RankBM25(Index):
     """
     BM25 index using rank-bm25.
@@ -491,6 +564,10 @@ def create(method, path, config, output, refresh):
         return SQLiteFTS(path, config, output, refresh)
     if method == "es":
         return Elastic(path, config, output, refresh)
+    if method == "similar":
+        return Similar(path, config, output, refresh)
+    if method == "rerank":
+        return Rerank(path, config, output, refresh)
 
     # Default
     return Embed(path, config, output, refresh)
@@ -614,7 +691,7 @@ def benchmarks(args):
             "climate-fever",
             "scifact",
         ]
-        methods = ["embed", "hybrid", "rag", "scoring", "rank", "bm25s", "sqlite", "es"]
+        methods = ["embed", "hybrid", "rag", "scoring", "rank", "bm25s", "sqlite", "es", "similar", "rerank"]
         mode = "w"
 
     # Run and save benchmarks

examples/rag_quickstart.py

@@ -0,0 +1,70 @@
+"""
+RAG Quick Start
+Easy to use way to get started with RAG using YOUR data
+
+For a complete application see this: https://github.com/neuml/rag
+
+TxtAI has many example notebooks covering everything the framework provides
+Examples: https://neuml.github.io/txtai/examples
+
+Install TxtAI
+  pip install txtai[pipeline-data]
+"""
+
+# pylint: disable=C0103
+import os
+
+from txtai import Embeddings, RAG
+from txtai.pipeline import Textractor
+
+# Step 1: Collect files from local directory
+#
+# Defaults to "data". Set to whereever your files are.
+path = "data"
+files = [os.path.join(path, f) for f in os.listdir(path) if os.path.isfile(os.path.join(path, f))]
+
+# Step 2: Text Extraction / Chunking
+#
+# Using section based chunking here. More complex options available such as semantic chunking, iterative chunking etc.
+# Documentation: https://neuml.github.io/txtai/pipeline/data/textractor
+# Supports Chonkie chunking as well: https://docs.chonkie.ai/oss/chunkers/overview
+textractor = Textractor(backend="docling", sections=True)
+chunks = []
+for f in files:
+    for chunk in textractor(f):
+        chunks.append((f, chunk))
+
+# Step 3: Build an embeddings database
+#
+# The `path` parameter sets the vector embeddings model. Supports Hugging Face models, llama.cpp, Ollama, vLLM and more.
+# Documentation: https://neuml.github.io/txtai/embeddings/
+embeddings = Embeddings(content=True, path="Qwen/Qwen3-Embedding-0.6B", maxlength=2048)
+embeddings.index(chunks)
+
+# Step 4: Create RAG pipeline
+#
+# Combines an embeddings database and an LLM.
+# Supports Hugging Face models, llama.cpp, Ollama, vLLM and more
+# Documentation: https://neuml.github.io/txtai/pipeline/text/rag
+
+# User prompt template
+template = """
+  Answer the following question using the provided context.
+
+  Question:
+  {question}
+
+  Context:
+  {context}
+"""
+
+rag = RAG(
+    embeddings,
+    "Qwen/Qwen3-0.6B",
+    system="You are a friendly assistant",
+    template=template,
+    output="flatten",
+)
+
+question = "Summarize the main advancements made by BERT"
+print(rag(question, maxlength=2048, stripthink=True))