chore: version packages (#991)

* Return partial result on failed job, especially job id

* Maintains NO_DATA_FOUND_IN_FILE throw behavior

.changeset/README.md

@@ -0,0 +1,8 @@
+# Changesets
+
+Hello and welcome! This folder has been automatically generated by `@changesets/cli`, a build tool that works
+with multi-package repos, or single-package repos to help you version and publish your code. You can
+find the full documentation for it [in our repository](https://github.com/changesets/changesets)
+
+We have a quick list of common questions to get you started engaging with this project in
+[our documentation](https://github.com/changesets/changesets/blob/main/docs/common-questions.md)

.changeset/config.json

@@ -0,0 +1,11 @@
+{
+  "$schema": "https://unpkg.com/@changesets/config@3.1.1/schema.json",
+  "changelog": "@changesets/cli/changelog",
+  "commit": false,
+  "fixed": [],
+  "linked": [],
+  "access": "restricted",
+  "baseBranch": "main",
+  "updateInternalDependencies": "patch",
+  "ignore": []
+}

.github/dependabot.yml

@@ -0,0 +1,11 @@
+# Please see the documentation for all configuration options:
+# https://docs.github.com/github/administering-a-repository/configuration-options-for-dependency-updates
+# and
+# https://docs.github.com/code-security/dependabot/dependabot-version-updates/configuration-options-for-the-dependabot.yml-file
+
+version: 2
+updates:
+  - package-ecosystem: "github-actions"
+    directory: "/"
+    schedule:
+      interval: "weekly"

.github/workflows/build_package.yml

@@ -1,48 +0,0 @@
-name: Build Package
-
-# Build package on its own without additional pip install
-
-on:
-  push:
-    branches:
-      - main
-  pull_request:
-
-env:
-  POETRY_VERSION: "1.6.1"
-
-jobs:
-  build:
-    runs-on: ${{ matrix.os }}
-    strategy:
-      # You can use PyPy versions in python-version.
-      # For example, pypy-2.7 and pypy-3.8
-      matrix:
-        os: [ubuntu-latest, windows-latest]
-        python-version: ["3.9"]
-    steps:
-      - uses: actions/checkout@v3
-      - name: Set up python ${{ matrix.python-version }}
-        uses: actions/setup-python@v4
-        with:
-          python-version: ${{ matrix.python-version }}
-      - name: Install Poetry
-        uses: snok/install-poetry@v1
-        with:
-          version: ${{ env.POETRY_VERSION }}
-      - name: Install deps
-        shell: bash
-        run: poetry install
-      - name: Ensure lock works
-        shell: bash
-        run: poetry lock
-      - name: Build
-        shell: bash
-        run: poetry build
-      - name: Test installing built package
-        shell: bash
-        run: python -m pip install .
-      - name: Test import
-        shell: bash
-        working-directory: ${{ vars.RUNNER_TEMP }}
-        run: python -c "import llama_cloud_services"

.github/workflows/build_package_py.yml

@@ -0,0 +1,53 @@
+name: Build Package - Python
+
+# Build package on its own without additional pip install
+
+on:
+  push:
+    branches:
+      - main
+    paths:
+      - "py/**"
+  pull_request:
+    paths:
+      - "py/**"
+env:
+  UV_VERSION: "0.7.20"
+
+jobs:
+  build:
+    runs-on: ${{ matrix.os }}
+    strategy:
+      # You can use PyPy versions in python-version.
+      # For example, pypy-2.7 and pypy-3.8
+      matrix:
+        os: [ubuntu-latest, windows-latest]
+        python-version: ["3.9"]
+    steps:
+      - uses: actions/checkout@v5
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v7
+        with:
+          version: ${{ env.UV_VERSION }}
+
+      - name: Set up Python
+        run: uv python install
+
+      - name: Display Python version
+        run: python --version
+
+      - name: Build
+        working-directory: py
+        run: uv build
+
+      - name: Test installing built package
+        shell: bash
+        working-directory: py
+        run: |
+          uv venv
+          uv pip install dist/*.whl
+
+      - name: Test import
+        working-directory: py
+        run: uv run -- python -c "import llama_cloud_services"

.github/workflows/build_package_ts.yml

@@ -0,0 +1,34 @@
+name: Build Package - TypeScript
+on:
+  push:
+    branches:
+      - main
+    paths:
+      - "ts/**"
+  pull_request:
+    paths:
+      - "ts/**"
+
+jobs:
+  pre_release:
+    name: Pre Release
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Checkout Repo
+        uses: actions/checkout@v5
+
+      - uses: pnpm/action-setup@v4
+
+      - name: Setup Node.js
+        uses: actions/setup-node@v5
+        with:
+          node-version-file: "ts/llama_cloud_services/.nvmrc"
+
+      - name: Install dependencies
+        working-directory: ts/llama_cloud_services/
+        run: pnpm install --no-frozen-lockfile
+
+      - name: Build
+        working-directory: ts/llama_cloud_services/
+        run: pnpm run build

.github/workflows/claude-code.yml

@@ -0,0 +1,95 @@
+name: Claude Code
+
+on:
+  issue_comment:
+    types: [created]
+  pull_request_review_comment:
+    types: [created]
+  issues:
+    types: [opened, assigned]
+  pull_request_review:
+    types: [submitted]
+
+jobs:
+  claude:
+    if: |
+      (github.event_name == 'issue_comment' && contains(github.event.comment.body, '@claude')) ||
+      (github.event_name == 'pull_request_review_comment' && contains(github.event.comment.body, '@claude')) ||
+      (github.event_name == 'pull_request_review' && contains(github.event.review.body, '@claude')) ||
+      (github.event_name == 'issues' && (contains(github.event.issue.body, '@claude') || contains(github.event.issue.title, '@claude')))
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+      pull-requests: read
+      issues: read
+      id-token: write
+    steps:
+      - name: Check repository access
+        id: check-access
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          # Get the user who triggered the event
+          case "${{ github.event_name }}" in
+            "issue_comment")
+              USER="${{ github.event.comment.user.login }}"
+              ;;
+            "pull_request_review_comment")
+              USER="${{ github.event.comment.user.login }}"
+              ;;
+            "pull_request_review")
+              USER="${{ github.event.review.user.login }}"
+              ;;
+            "issues")
+              USER="${{ github.event.issue.user.login }}"
+              ;;
+          esac
+
+          echo "Checking repository access for user: $USER"
+
+          # Check if user has write access to the repository
+          REPO="${{ github.repository }}"
+          if gh api repos/$REPO/collaborators/$USER/permission --jq '.permission' | grep -E "(admin|write)" > /dev/null 2>&1; then
+            echo "User $USER has write access to the repository"
+            echo "authorized=true" >> $GITHUB_OUTPUT
+          else
+            echo "User $USER does not have write access to the repository"
+            echo "authorized=false" >> $GITHUB_OUTPUT
+            exit 1
+          fi
+
+      - name: Checkout repository
+        if: steps.check-access.outputs.authorized == 'true'
+        uses: actions/checkout@v5
+        with:
+          fetch-depth: 1
+
+      - name: Run Claude Code
+        if: steps.check-access.outputs.authorized == 'true'
+        id: claude
+        uses: anthropics/claude-code-action@beta
+        with:
+          anthropic_api_key: ${{ secrets.ANTHROPIC_GITHUB_API_KEY }}
+
+          # Optional: Specify model (defaults to Claude Sonnet 4, uncomment for Claude Opus 4)
+          # model: "claude-opus-4-20250514"
+
+          # Optional: Customize the trigger phrase (default: @claude)
+          # trigger_phrase: "/claude"
+
+          # Optional: Trigger when specific user is assigned to an issue
+          # assignee_trigger: "claude-bot"
+
+          # Optional: Allow Claude to run specific commands
+          # Allow bash commands to be run, for things like running tests, linting, etc.
+          allowed_tools: "Bash(rg:*),Bash(find:*),Bash(grep:*),Bash(pnpm:*),Bash(npm:*),Bash(uv:*),Bash(pip:*),Bash(pipx:*),Bash(make:*),Bash(cd:*),WebFetch"
+
+          # Optional: Add custom instructions for Claude to customize its behavior for your project
+          # custom_instructions: |
+          #   Follow our coding standards
+          #   Ensure all new code has tests
+          #   Use TypeScript for new files
+
+          # Optional: Custom environment variables for Claude
+          # claude_env: |
+          #   NODE_ENV: test

.github/workflows/codeql.yml

@@ -1,14 +1,3 @@
-# For most projects, this workflow file will not need changing; you simply need
-# to commit it to your repository.
-#
-# You may wish to alter this file to override the set of languages analyzed,
-# or to provide custom queries or build logic.
-#
-# ******** NOTE ********
-# We have attempted to detect the languages in your repository. Please check
-# the `language` matrix defined below to confirm you have the correct set of
-# supported CodeQL languages.
-#
 name: "CodeQL"
 
 on:
@@ -28,54 +17,25 @@ jobs:
     #   - https://gh.io/supported-runners-and-hardware-resources
     #   - https://gh.io/using-larger-runners
     # Consider using larger runners for possible analysis time improvements.
-    runs-on: ${{ (matrix.language == 'swift' && 'macos-latest') || 'ubuntu-latest' }}
-    timeout-minutes: ${{ (matrix.language == 'swift' && 120) || 360 }}
+    runs-on: "ubuntu-latest"
+    timeout-minutes: 360
     permissions:
       actions: read
       contents: read
       security-events: write
 
-    strategy:
-      fail-fast: false
-      matrix:
-        language: ["python"]
-        # CodeQL supports [ 'cpp', 'csharp', 'go', 'java', 'javascript', 'python', 'ruby', 'swift' ]
-        # Use only 'java' to analyze code written in Java, Kotlin or both
-        # Use only 'javascript' to analyze code written in JavaScript, TypeScript or both
-        # Learn more about CodeQL language support at https://aka.ms/codeql-docs/language-support
-
     steps:
       - name: Checkout repository
-        uses: actions/checkout@v3
+        uses: actions/checkout@v5
 
       # Initializes the CodeQL tools for scanning.
       - name: Initialize CodeQL
-        uses: github/codeql-action/init@v2
+        uses: github/codeql-action/init@v4
         with:
-          languages: ${{ matrix.language }}
-          # If you wish to specify custom queries, you can do so here or in a config file.
-          # By default, queries listed here will override any specified in a config file.
-          # Prefix the list here with "+" to use these queries and those in the config file.
-
-          # For more details on CodeQL's query packs, refer to: https://docs.github.com/en/code-security/code-scanning/automatically-scanning-your-code-for-vulnerabilities-and-errors/configuring-code-scanning#using-queries-in-ql-packs
-          # queries: security-extended,security-and-quality
-
-      # Autobuild attempts to build any compiled languages (C/C++, C#, Go, Java, or Swift).
-      # If this step fails, then you should remove it and run the build manually (see below)
-      - name: Autobuild
-        uses: github/codeql-action/autobuild@v2
-
-      # ℹ️ Command-line programs to run using the OS shell.
-      # 📚 See https://docs.github.com/en/actions/using-workflows/workflow-syntax-for-github-actions#jobsjob_idstepsrun
-
-      #   If the Autobuild fails above, remove it and uncomment the following three lines.
-      #   modify them (or add more) to build your code if your project, please refer to the EXAMPLE below for guidance.
-
-      # - run: |
-      #     echo "Run, Build Application using script"
-      #     ./location_of_script_within_repo/buildscript.sh
+          languages: python
+          dependency-caching: true
 
       - name: Perform CodeQL Analysis
-        uses: github/codeql-action/analyze@v2
+        uses: github/codeql-action/analyze@v4
         with:
-          category: "/language:${{matrix.language}}"
+          category: "/language:python"

.github/workflows/lint.yml

@@ -1,4 +1,4 @@
-name: Linting
+name: Lint
 
 on:
   push:
@@ -7,7 +7,7 @@ on:
   pull_request:
 
 env:
-  POETRY_VERSION: "1.6.1"
+  UV_VERSION: "0.7.20"
 
 jobs:
   build:
@@ -18,20 +18,29 @@ jobs:
       matrix:
         python-version: ["3.9"]
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v5
         with:
           fetch-depth: ${{ github.event_name == 'pull_request' && 2 || 0 }}
-      - name: Set up python ${{ matrix.python-version }}
-        uses: actions/setup-python@v4
+      - name: Install uv
+        uses: astral-sh/setup-uv@v7
         with:
-          python-version: ${{ matrix.python-version }}
-      - name: Install Poetry
-        uses: snok/install-poetry@v1
+          version: ${{ env.UV_VERSION }}
+
+      - name: Set up Python
+        run: uv python install ${{ matrix.python-version }}
+
+      - uses: pnpm/action-setup@v4
+      - name: Setup Node.js
+        uses: actions/setup-node@v5
         with:
-          version: ${{ env.POETRY_VERSION }}
-      - name: Install pre-commit
-        shell: bash
-        run: poetry run pip install pre-commit
+          node-version-file: "ts/llama_cloud_services/.nvmrc"
+      - name: Install dependencies
+        run: pnpm install --no-frozen-lockfile
+
       - name: Run linter
         shell: bash
-        run: poetry run make lint
+        working-directory: py
+        run: uv run -- pre-commit run -a
+      # the js checks are run roundaboutly through lint-staged, and -a doesn't run it. Run them directly.
+      - run: pnpm -w --filter llama-cloud-services run lint
+      - run: pnpm -w --filter llama-cloud-services run format:check

.github/workflows/publish_release.yml

@@ -1,75 +0,0 @@
-name: Publish llama-parse to PyPI / GitHub
-
-on:
-  push:
-    tags:
-      - "v*"
-
-  workflow_dispatch:
-
-env:
-  POETRY_VERSION: "1.6.1"
-  PYTHON_VERSION: "3.9"
-
-jobs:
-  build-n-publish:
-    name: Build and publish to PyPI
-    if: github.repository == 'run-llama/llama_cloud_services'
-    runs-on: ubuntu-latest
-
-    steps:
-      - uses: actions/checkout@v3
-      - name: Set up python ${{ env.PYTHON_VERSION }}
-        uses: actions/setup-python@v4
-        with:
-          python-version: ${{ env.PYTHON_VERSION }}
-
-      - name: Install Poetry
-        uses: snok/install-poetry@v1
-        with:
-          version: ${{ env.POETRY_VERSION }}
-
-      - name: Install deps
-        shell: bash
-        run: pip install -e .
-
-      - name: Build and publish llama-cloud-services
-        uses: JRubics/poetry-publish@v2.1
-        with:
-          pypi_token: ${{ secrets.LLAMA_PARSE_PYPI_TOKEN }}
-          poetry_install_options: "--without dev"
-
-      - name: Build and publish llama-parse
-        uses: JRubics/poetry-publish@v2.1
-        with:
-          working_directory: "llama_parse"
-          pypi_token: ${{ secrets.LLAMA_PARSE_PYPI_TOKEN }}
-          poetry_install_options: "--without dev"
-
-      - name: Create GitHub Release
-        id: create_release
-        uses: actions/create-release@v1
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }} # This token is provided by Actions, you do not need to create your own token
-        with:
-          tag_name: ${{ github.ref }}
-          release_name: ${{ github.ref }}
-          draft: false
-          prerelease: false
-
-      - name: Get Asset name
-        run: |
-          export PKG=$(ls dist/ | grep tar)
-          set -- $PKG
-          echo "name=$1" >> $GITHUB_ENV
-
-      - name: Upload Release Asset (sdist) to GitHub
-        id: upload-release-asset
-        uses: actions/upload-release-asset@v1
-        env:
-          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        with:
-          upload_url: ${{ steps.create_release.outputs.upload_url }}
-          asset_path: dist/${{ env.name }}
-          asset_name: ${{ env.name }}
-          asset_content_type: application/zip

.github/workflows/test_e2e_py.yml

@@ -0,0 +1,38 @@
+name: Test end-to-end - Python
+
+on:
+  pull_request:
+    paths:
+      - "py/**"
+
+env:
+  UV_VERSION: "0.7.20"
+  LLAMA_CLOUD_API_KEY: ${{ secrets.LLAMA_CLOUD_API_KEY }}
+
+jobs:
+  test_e2e:
+    runs-on: ubuntu-latest
+    strategy:
+      # You can use PyPy versions in python-version.
+      # For example, pypy-2.7 and pypy-3.8
+      matrix:
+        python-version: ["3.12"]
+    steps:
+      - uses: actions/checkout@v5
+        with:
+          fetch-depth: 0
+      - name: Install uv
+        uses: astral-sh/setup-uv@v7
+        with:
+          version: ${{ env.UV_VERSION }}
+
+      - name: Set up Python
+        run: uv python install ${{ matrix.python-version }} && uv python pin ${{ matrix.python-version }}
+
+      - name: Run Tests
+        working-directory: py
+        run: make e2e
+
+      - name: Remove virtual environment
+        working-directory: py
+        run: rm -rf .venv/

.github/workflows/test_py.yml

@@ -0,0 +1,42 @@
+name: Test - Python
+
+on:
+  push:
+    branches:
+      - main
+    paths:
+      - "py/**"
+  pull_request:
+    paths:
+      - "py/**"
+
+env:
+  UV_VERSION: "0.7.20"
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    strategy:
+      # You can use PyPy versions in python-version.
+      # For example, pypy-2.7 and pypy-3.8
+      matrix:
+        python-version: ["3.9", "3.10", "3.11", "3.12"]
+    steps:
+      - uses: actions/checkout@v5
+        with:
+          fetch-depth: 0
+      - name: Install uv
+        uses: astral-sh/setup-uv@v7
+        with:
+          version: ${{ env.UV_VERSION }}
+
+      - name: Set up Python
+        run: uv python install ${{ matrix.python-version }} && uv python pin ${{ matrix.python-version }}
+
+      - name: Run Tests
+        working-directory: py
+        run: uv run pytest unit_tests/ -v
+
+      - name: Remove virtual environment
+        working-directory: py
+        run: rm -rf .venv/

.github/workflows/test_ts.yml

@@ -0,0 +1,39 @@
+name: Test - TypeScript
+
+on:
+  push:
+    branches:
+      - main
+    paths:
+      - "ts/**"
+  pull_request:
+    paths:
+      - "ts/**"
+
+env:
+  TURBO_TOKEN: ${{ secrets.TURBO_TOKEN }}
+  TURBO_TEAM: ${{ vars.TURBO_TEAM }}
+  TURBO_REMOTE_ONLY: true
+  LLAMA_CLOUD_API_KEY: ${{ secrets.LLAMA_CLOUD_API_KEY }}
+
+jobs:
+  test:
+    name: Test - TypeScript
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v5
+      - uses: pnpm/action-setup@v4
+      - name: Setup Node.js
+        uses: actions/setup-node@v5
+        with:
+          node-version-file: "ts/llama_cloud_services/.nvmrc"
+      - name: Install dependencies
+        run: pnpm -r install --no-frozen-lockfile
+      - name: Build package
+        run: pnpm --filter llama-cloud-services build
+      - name: Run Tests
+        working-directory: ts/llama_cloud_services/
+        run: pnpm test
+      - name: Run e2e tests
+        working-directory: ts/e2e-tests/
+        run: pnpm test

.github/workflows/unit_test.yml

@@ -1,40 +0,0 @@
-name: Unit Testing
-
-on:
-  push:
-    branches:
-      - main
-  pull_request:
-
-env:
-  POETRY_VERSION: "1.6.1"
-  LLAMA_CLOUD_API_KEY: ${{ secrets.LLAMA_CLOUD_API_KEY }}
-
-jobs:
-  test:
-    runs-on: ubuntu-latest
-    strategy:
-      # You can use PyPy versions in python-version.
-      # For example, pypy-2.7 and pypy-3.8
-      matrix:
-        python-version: ["3.9", "3.10", "3.11", "3.12"]
-    steps:
-      - uses: actions/checkout@v3
-        with:
-          fetch-depth: 0
-      - name: Set up python ${{ matrix.python-version }}
-        uses: actions/setup-python@v4
-        with:
-          python-version: ${{ matrix.python-version }}
-      - name: Install Poetry
-        uses: snok/install-poetry@v1
-        with:
-          version: ${{ env.POETRY_VERSION }}
-      - name: Install deps
-        shell: bash
-        run: poetry install --with dev
-      - name: Run testing
-        env:
-          CI: true
-        shell: bash
-        run: poetry run pytest tests

.github/workflows/version_bump_and_release.yml

@@ -0,0 +1,61 @@
+name: Version Bump and Release
+
+on:
+  push:
+    branches:
+      - main
+
+concurrency: ${{ github.workflow }}-${{ github.ref }}
+
+jobs:
+  release:
+    name: Release
+    runs-on: ubuntu-latest
+    # Only run on main branch pushes
+    if: github.ref == 'refs/heads/main'
+    steps:
+      - name: Checkout Repo
+        uses: actions/checkout@v5
+
+      - uses: pnpm/action-setup@v4
+
+      - name: Setup Node.js
+        uses: actions/setup-node@v5
+        with:
+          node-version: "22"
+          cache: "pnpm"
+
+      - name: Setup Python
+        uses: actions/setup-python@v6
+        with:
+          python-version: "3.11"
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v7
+
+      - name: Install dependencies
+        run: pnpm install
+
+      - name: Add auth token to .npmrc file
+        run: |
+          cat << EOF >> ".npmrc"
+            //registry.npmjs.org/:_authToken=$NPM_TOKEN
+          EOF
+        env:
+          NPM_TOKEN: ${{ secrets.NPM_TOKEN }}
+
+      - name: Create Release Pull Request or Publish packages
+        id: changesets
+        uses: changesets/action@v1
+        with:
+          commit: "chore: version packages"
+          title: "chore: version packages"
+          # Custom version script
+          version: pnpm -w run version
+          # Custom publish script
+          publish: pnpm -w run publish
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          NPM_TOKEN: ${{ secrets.NPM_TOKEN }}
+          UV_PUBLISH_TOKEN: ${{ secrets.PYPI_TOKEN }}
+          LLAMA_PARSE_PYPI_TOKEN: ${{ secrets.LLAMA_PARSE_PYPI_TOKEN }}

.gitignore

@@ -5,3 +5,8 @@ __pycache__/
 .idea
 .env*
 .ipynb_checkpoints*
+*_cache/
+node_modules/
+.turbo/
+dist/
+.npmrc

.pre-commit-config.yaml

@@ -15,25 +15,26 @@ repos:
       - id: end-of-file-fixer
       - id: mixed-line-ending
       - id: trailing-whitespace
+        exclude: ^ts/llama_cloud_services/src/client/
   - repo: https://github.com/charliermarsh/ruff-pre-commit
     rev: v0.1.5
 
     hooks:
       - id: ruff
         args: [--fix, --exit-non-zero-on-fix]
-        exclude: ".*poetry.lock"
+        exclude: ".*uv.lock"
   - repo: https://github.com/psf/black-pre-commit-mirror
     rev: 23.10.1
     hooks:
       - id: black-jupyter
         name: black-src
         alias: black
-        exclude: ".*poetry.lock"
+        exclude: ".*uv.lock|examples/extract/solar_panel_e2e_comparison.ipynb"
   - repo: https://github.com/pre-commit/mirrors-mypy
     rev: v1.0.1
     hooks:
       - id: mypy
-        exclude: ^tests/
+        exclude: ^py/tests|^py/unit_tests
         additional_dependencies:
           [
             "types-requests",
@@ -59,17 +60,19 @@ repos:
         additional_dependencies: [black==23.10.1]
         # Using PEP 8's line length in docs prevents excess left/right scrolling
         args: [--line-length=79]
-  - repo: https://github.com/pre-commit/mirrors-prettier
-    rev: v3.0.3
+  - repo: local
     hooks:
-      - id: prettier
-        exclude: poetry.lock
+      - id: lint-staged
+        name: Run lint-staged for TS files
+        entry: pnpm -w exec lint-staged
+        language: system
+        pass_filenames: false
   - repo: https://github.com/codespell-project/codespell
     rev: v2.2.6
     hooks:
       - id: codespell
         additional_dependencies: [tomli]
-        exclude: ^(poetry.lock|examples)
+        exclude: ^(uv.lock|docs|ts|examples|pnpm-lock.yaml)
         args:
           [
             "--ignore-words-list",
@@ -84,6 +87,6 @@ repos:
     rev: v0.23.1
     hooks:
       - id: toml-sort-fix
-        exclude: ".*poetry.lock"
+        exclude: ".*uv.lock"
 
-exclude: .github/ISSUE_TEMPLATE
+exclude: ^(.github/ISSUE_TEMPLATE|ts/llama_cloud_services/src/client|pnpm-lock.yaml)

CONTRIBUTING.md

@@ -0,0 +1,33 @@
+# Python
+
+## Installation
+
+This project uses uv. Create a virtual environment, and run `uv sync`
+
+## Versioning (Maintainers only)
+
+Before merging your changes, make sure to bump the versions.
+
+Make a version bump to `pyproject.toml`. If the underlying dependency on the llamacloud platform OpenAPI
+sdk needs bumping, make sure to bring that in as well. If updating dependencies, run `uv lock`.
+
+The legacy `llama_parse` package re-exports some of `llama_cloud_services` in the old namespace. The
+versions need to be kept consistent to sidecar it with `llama_cloud_services`. Bump it's version in `llama_parse/pyproject.toml`, and also bump it's dependency version of `llama-cloud-services` to match.
+
+**Note**: Don't worry about updating the `llama_parse/poetry.lock` file when bumping versions. The GitHub action will automatically run `poetry lock` for the llama_parse package during the build process (though it doesn't commit the updated lockfile back to the repo).
+
+You can also do this with `./scripts/version-bump.py set 0.x.x` if you have `uv` installed.
+
+Once the change is merged, push a tag `git tag -a v0.x.x -m 0.x.x` and `git push origin v0.x.x`.
+
+This tagging step can be done with `./scripts/version-bump tag`.
+
+# Typescript
+
+## Installation
+
+...
+
+## Versioning
+
+...

Makefile

@@ -1,14 +0,0 @@
-GIT_ROOT ?= $(shell git rev-parse --show-toplevel)
-
-help:	## Show all Makefile targets.
-	@grep -E '^[a-zA-Z_-]+:.*?## .*$$' $(MAKEFILE_LIST) | awk 'BEGIN {FS = ":.*?## "}; {printf "\033[33m%-30s\033[0m %s\n", $$1, $$2}'
-
-format:	## Run code autoformatters (black).
-	pre-commit install
-	git ls-files | xargs pre-commit run black --files
-
-lint:	## Run linters: pre-commit (black, ruff, codespell) and mypy
-	pre-commit install && git ls-files | xargs pre-commit run --show-diff-on-failure --files
-
-test:	## Run tests via pytest
-	pytest tests

README.md

@@ -9,8 +9,8 @@ This repository contains the code for hand-written SDKs and clients for interact
 This includes:
 
 - [LlamaParse](./parse.md) - A GenAI-native document parser that can parse complex document data for any downstream LLM use case (Agents, RAG, data processing, etc.).
-- [LlamaReport (beta/invite-only)](./report.md) - A prebuilt agentic report builder that can be used to build reports from a variety of data sources.
-- [LlamaExtract (beta/invite-only)](./extract.md) - A prebuilt agentic data extractor that can be used to transform data into a structured JSON representation.
+- [LlamaExtract](./extract.md) - A prebuilt agentic data extractor that can be used to transform data into a structured JSON representation.
+- [LlamaCloud Index](./index.md) - A widely customizable and fully automated document ingestion pipeline that also serves retrieval purposes.
 
 ## Getting Started
 
@@ -25,18 +25,47 @@ Then, get your API key from [LlamaCloud](https://cloud.llamaindex.ai/).
 Then, you can use the services in your code:
 
 ```python
-from llama_cloud_services import LlamaParse, LlamaReport, LlamaExtract
+from llama_cloud_services import (
+    LlamaParse,
+    LlamaExtract,
+    LlamaCloudIndex,
+)
 
 parser = LlamaParse(api_key="YOUR_API_KEY")
-report = LlamaReport(api_key="YOUR_API_KEY")
 extract = LlamaExtract(api_key="YOUR_API_KEY")
+index = LlamaCloudIndex(
+    "my_first_index", project_name="default", api_key="YOUR_API_KEY"
+)
 ```
 
 See the quickstart guides for each service for more information:
 
 - [LlamaParse](./parse.md)
-- [LlamaReport (beta/invite-only)](./report.md)
-- [LlamaExtract (beta/invite-only)](./extract.md)
+- [LlamaExtract](./extract.md)
+- [LlamaCloud Index](./index.md)
+
+## Switch to EU SaaS 🇪🇺
+
+If you are interested in using LlamaCloud services in the EU, you can adjust your base URL to `https://api.cloud.eu.llamaindex.ai`.
+
+You can also create your API key in the EU region [here](https://cloud.eu.llamaindex.ai).
+
+```python
+from llama_cloud_services import (
+    LlamaParse,
+    LlamaExtract,
+    EU_BASE_URL,
+)
+
+parser = LlamaParse(api_key="YOUR_API_KEY", base_url=EU_BASE_URL)
+extract = LlamaExtract(api_key="YOUR_API_KEY", base_url=EU_BASE_URL)
+index = LlamaCloudIndex(
+    "my_first_index",
+    project_name="default",
+    api_key="YOUR_API_KEY",
+    base_url=EU_BASE_URL,
+)
+```
 
 ## Documentation
 

examples-ts/README.md

@@ -0,0 +1,8 @@
+# LlamaCloud Services Examples - Python
+
+In this folder you will find several TypeScript end-to-end applications that contain examples regarding:
+
+- [LlamaParse](./parse/)
+- [LlamaCloud Index](./index/)
+
+Follow the instructions in each example folder to get started!

examples-ts/classify/.gitignore

@@ -0,0 +1,21 @@
+node_modules
+package-lock.json
+yarn.lock
+
+.DS_Store
+.cache
+.env
+.vercel
+.output
+.nitro
+/build/
+/api/
+/server/build
+/public/build# Sentry Config File
+.env.sentry-build-plugin
+/test-results/
+/playwright-report/
+/blob-report/
+/playwright/.cache/
+.tanstack
+.vscode

examples-ts/classify/.prettierignore

@@ -0,0 +1,4 @@
+**/build
+**/public
+pnpm-lock.yaml
+routeTree.gen.ts

examples-ts/classify/README.md

@@ -0,0 +1,88 @@
+# LlamaClassify Demo
+
+A TypeScript demo application showcasing the power of **LlamaClassify** - an agentic documents classification service from [LlamaCloud](https://cloud.llamaindex.ai). This demo allows you to classify financial documents among three different types (Cash flow statement, Income Statement and Balance Sheet).
+
+## Table of Contents
+
+- [Features](#features)
+- [Prerequisites](#prerequisites)
+- [Installation](#installation)
+- [Usage](#usage)
+  - [Start the Demo](#start-the-demo)
+- [How It Works](#how-it-works)
+- [Troubleshooting](#troubleshooting)
+  - [Common Issues](#common-issues)
+- [License](#license)
+- [Contributing](#contributing)
+
+## Features
+
+- 📄 **Documemt Classification**: Classify files based on well-defined rules you can customized and play around with.
+- 🤖 **Reasoning-based Actionable Insights**: Get in-depth, reasoning based insights on the document classification, accompanied by confidence scores.
+- 🎨 **Beautiful UI**: [DaisyUI](https://daisyui.com)-based interface powered by [TanStack](https://tanstack.com)
+- ⚡ **Fast Development**: Hot reload support with development mode
+- 🛠️ **TypeScript**: Full TypeScript support with strict type checking
+
+## Prerequisites
+
+- Node.js (version 22 or higher)
+- pnpm package manager
+- LlamaCloud API key
+
+## Installation
+
+1. Clone the repository:
+
+```bash
+git clone https://github.com/run-llama/llama_cloud_services
+cd lama_cloud_services/examples-ts/classify/
+```
+
+2. Install dependencies:
+
+```bash
+npm install
+```
+
+3. Set up your environment variables:
+
+```bash
+# Add your API key to your environment
+export LLAMA_CLOUD_API_KEY="your-llamacloud-api-key"
+```
+
+## Usage
+
+### Start the Demo
+
+```bash
+npm run dev
+```
+
+The application will be up and running on http://localhost:3000
+
+## How It Works
+
+1. **Document Input**: Enter the path to your document when prompted
+2. **Parsing**: LlamaClassify, based on the rules you can find [here](./src/utils/classifier.ts), processes the document and classifies it
+3. **Results**: The classification outcome, as well as the reasoning behind it and the confidence score, are displayed in the UI.
+
+## Troubleshooting
+
+### Common Issues
+
+1. **Module Resolution Errors**: Ensure you're using Node.js 22+ and have all dependencies installed
+2. **API Key Issues**: Verify your LlamaCloud API key is correctly set
+3. **File Path Errors**: Use absolute paths or ensure relative paths are correct from the project root
+
+## License
+
+MIT License - see the [LICENSE](../../LICENSE) file for details.
+
+## Contributing
+
+1. Fork the repository
+2. Create a feature branch
+3. Make your changes
+4. Run `npm run format` and `npm run lint`
+5. Submit a pull request

examples-ts/classify/package.json

@@ -0,0 +1,34 @@
+{
+  "name": "tanstack-start-example-basic",
+  "private": true,
+  "sideEffects": false,
+  "type": "module",
+  "scripts": {
+    "dev": "vite dev",
+    "build": "vite build && tsc --noEmit",
+    "start": "node .output/server/index.mjs"
+  },
+  "dependencies": {
+    "@tanstack/react-router": "^1.133.22",
+    "@tanstack/react-router-devtools": "^1.133.22",
+    "@tanstack/react-start": "^1.133.22",
+    "llama-cloud-services": "^0.3.10",
+    "react": "^19.0.0",
+    "react-dom": "^19.0.0",
+    "tailwind-merge": "^2.6.0",
+    "zod": "^3.24.2"
+  },
+  "devDependencies": {
+    "@tailwindcss/postcss": "^4.1.15",
+    "@types/node": "^22.5.4",
+    "@types/react": "^19.0.8",
+    "@types/react-dom": "^19.0.3",
+    "@vitejs/plugin-react": "^4.6.0",
+    "daisyui": "^5.3.7",
+    "postcss": "^8.5.1",
+    "tailwindcss": "^4.1.15",
+    "typescript": "^5.7.2",
+    "vite": "^7.1.7",
+    "vite-tsconfig-paths": "^5.1.4"
+  }
+}

examples-ts/classify/postcss.config.mjs

@@ -0,0 +1,5 @@
+export default {
+  plugins: {
+    '@tailwindcss/postcss': {},
+  },
+}

examples-ts/classify/public/site.webmanifest

@@ -0,0 +1,19 @@
+{
+  "name": "",
+  "short_name": "",
+  "icons": [
+    {
+      "src": "/android-chrome-192x192.png",
+      "sizes": "192x192",
+      "type": "image/png"
+    },
+    {
+      "src": "/android-chrome-512x512.png",
+      "sizes": "512x512",
+      "type": "image/png"
+    }
+  ],
+  "theme_color": "#ffffff",
+  "background_color": "#ffffff",
+  "display": "standalone"
+}

examples-ts/classify/src/components/DefaultCatchBoundary.tsx

@@ -0,0 +1,53 @@
+import {
+  ErrorComponent,
+  Link,
+  rootRouteId,
+  useMatch,
+  useRouter,
+} from '@tanstack/react-router'
+import type { ErrorComponentProps } from '@tanstack/react-router'
+
+export function DefaultCatchBoundary({ error }: ErrorComponentProps) {
+  const router = useRouter()
+  const isRoot = useMatch({
+    strict: false,
+    select: (state) => state.id === rootRouteId,
+  })
+
+  console.error('DefaultCatchBoundary Error:', error)
+
+  return (
+    <div className="min-w-0 flex-1 p-4 flex flex-col items-center justify-center gap-6">
+      <ErrorComponent error={error} />
+      <div className="flex gap-2 items-center flex-wrap">
+        <button
+          onClick={() => {
+            router.invalidate()
+          }}
+          className={`px-2 py-1 bg-gray-600 dark:bg-gray-700 rounded-sm text-white uppercase font-extrabold`}
+        >
+          Try Again
+        </button>
+        {isRoot ? (
+          <Link
+            to="/"
+            className={`px-2 py-1 bg-gray-600 dark:bg-gray-700 rounded-sm text-white uppercase font-extrabold`}
+          >
+            Home
+          </Link>
+        ) : (
+          <Link
+            to="/"
+            className={`px-2 py-1 bg-gray-600 dark:bg-gray-700 rounded-sm text-white uppercase font-extrabold`}
+            onClick={(e) => {
+              e.preventDefault()
+              window.history.back()
+            }}
+          >
+            Go Back
+          </Link>
+        )}
+      </div>
+    </div>
+  )
+}

examples-ts/classify/src/components/NotFound.tsx

@@ -0,0 +1,25 @@
+import { Link } from '@tanstack/react-router'
+
+export function NotFound({ children }: { children?: any }) {
+  return (
+    <div className="space-y-2 p-2">
+      <div className="text-gray-600 dark:text-gray-400">
+        {children || <p>The page you are looking for does not exist.</p>}
+      </div>
+      <p className="flex items-center gap-2 flex-wrap">
+        <button
+          onClick={() => window.history.back()}
+          className="bg-emerald-500 text-white px-2 py-1 rounded-sm uppercase font-black text-sm"
+        >
+          Go back
+        </button>
+        <Link
+          to="/"
+          className="bg-cyan-600 text-white px-2 py-1 rounded-sm uppercase font-black text-sm"
+        >
+          Start Over
+        </Link>
+      </p>
+    </div>
+  )
+}

examples-ts/classify/src/routeTree.gen.ts

@@ -0,0 +1,225 @@
+/* eslint-disable */
+
+// @ts-nocheck
+
+// noinspection JSUnusedGlobalSymbols
+
+// This file was automatically generated by TanStack Router.
+// You should NOT make any changes in this file as it will be overwritten.
+// Additionally, you should also exclude this file from your linter and/or formatter to prevent it from being checked or modified.
+
+import { Route as rootRouteImport } from './routes/__root'
+import { Route as UsersRouteImport } from './routes/users'
+import { Route as IndexRouteImport } from './routes/index'
+import { Route as UsersIndexRouteImport } from './routes/users.index'
+import { Route as PostsIndexRouteImport } from './routes/posts.index'
+import { Route as UsersUserIdRouteImport } from './routes/users.$userId'
+import { Route as PostsPostIdRouteImport } from './routes/posts.$postId'
+import { Route as ApiClassifyRouteImport } from './routes/api/classify'
+import { Route as PostsPostIdDeepRouteImport } from './routes/posts_.$postId.deep'
+
+const UsersRoute = UsersRouteImport.update({
+  id: '/users',
+  path: '/users',
+  getParentRoute: () => rootRouteImport,
+} as any)
+const IndexRoute = IndexRouteImport.update({
+  id: '/',
+  path: '/',
+  getParentRoute: () => rootRouteImport,
+} as any)
+const UsersIndexRoute = UsersIndexRouteImport.update({
+  id: '/',
+  path: '/',
+  getParentRoute: () => UsersRoute,
+} as any)
+const PostsIndexRoute = PostsIndexRouteImport.update({
+  id: '/posts/',
+  path: '/posts/',
+  getParentRoute: () => rootRouteImport,
+} as any)
+const UsersUserIdRoute = UsersUserIdRouteImport.update({
+  id: '/$userId',
+  path: '/$userId',
+  getParentRoute: () => UsersRoute,
+} as any)
+const PostsPostIdRoute = PostsPostIdRouteImport.update({
+  id: '/posts/$postId',
+  path: '/posts/$postId',
+  getParentRoute: () => rootRouteImport,
+} as any)
+const ApiClassifyRoute = ApiClassifyRouteImport.update({
+  id: '/api/classify',
+  path: '/api/classify',
+  getParentRoute: () => rootRouteImport,
+} as any)
+const PostsPostIdDeepRoute = PostsPostIdDeepRouteImport.update({
+  id: '/posts_/$postId/deep',
+  path: '/posts/$postId/deep',
+  getParentRoute: () => rootRouteImport,
+} as any)
+
+export interface FileRoutesByFullPath {
+  '/': typeof IndexRoute
+  '/users': typeof UsersRouteWithChildren
+  '/api/classify': typeof ApiClassifyRoute
+  '/posts/$postId': typeof PostsPostIdRoute
+  '/users/$userId': typeof UsersUserIdRoute
+  '/posts': typeof PostsIndexRoute
+  '/users/': typeof UsersIndexRoute
+  '/posts/$postId/deep': typeof PostsPostIdDeepRoute
+}
+export interface FileRoutesByTo {
+  '/': typeof IndexRoute
+  '/api/classify': typeof ApiClassifyRoute
+  '/posts/$postId': typeof PostsPostIdRoute
+  '/users/$userId': typeof UsersUserIdRoute
+  '/posts': typeof PostsIndexRoute
+  '/users': typeof UsersIndexRoute
+  '/posts/$postId/deep': typeof PostsPostIdDeepRoute
+}
+export interface FileRoutesById {
+  __root__: typeof rootRouteImport
+  '/': typeof IndexRoute
+  '/users': typeof UsersRouteWithChildren
+  '/api/classify': typeof ApiClassifyRoute
+  '/posts/$postId': typeof PostsPostIdRoute
+  '/users/$userId': typeof UsersUserIdRoute
+  '/posts/': typeof PostsIndexRoute
+  '/users/': typeof UsersIndexRoute
+  '/posts_/$postId/deep': typeof PostsPostIdDeepRoute
+}
+export interface FileRouteTypes {
+  fileRoutesByFullPath: FileRoutesByFullPath
+  fullPaths:
+    | '/'
+    | '/users'
+    | '/api/classify'
+    | '/posts/$postId'
+    | '/users/$userId'
+    | '/posts'
+    | '/users/'
+    | '/posts/$postId/deep'
+  fileRoutesByTo: FileRoutesByTo
+  to:
+    | '/'
+    | '/api/classify'
+    | '/posts/$postId'
+    | '/users/$userId'
+    | '/posts'
+    | '/users'
+    | '/posts/$postId/deep'
+  id:
+    | '__root__'
+    | '/'
+    | '/users'
+    | '/api/classify'
+    | '/posts/$postId'
+    | '/users/$userId'
+    | '/posts/'
+    | '/users/'
+    | '/posts_/$postId/deep'
+  fileRoutesById: FileRoutesById
+}
+export interface RootRouteChildren {
+  IndexRoute: typeof IndexRoute
+  UsersRoute: typeof UsersRouteWithChildren
+  ApiClassifyRoute: typeof ApiClassifyRoute
+  PostsPostIdRoute: typeof PostsPostIdRoute
+  PostsIndexRoute: typeof PostsIndexRoute
+  PostsPostIdDeepRoute: typeof PostsPostIdDeepRoute
+}
+
+declare module '@tanstack/react-router' {
+  interface FileRoutesByPath {
+    '/users': {
+      id: '/users'
+      path: '/users'
+      fullPath: '/users'
+      preLoaderRoute: typeof UsersRouteImport
+      parentRoute: typeof rootRouteImport
+    }
+    '/': {
+      id: '/'
+      path: '/'
+      fullPath: '/'
+      preLoaderRoute: typeof IndexRouteImport
+      parentRoute: typeof rootRouteImport
+    }
+    '/users/': {
+      id: '/users/'
+      path: '/'
+      fullPath: '/users/'
+      preLoaderRoute: typeof UsersIndexRouteImport
+      parentRoute: typeof UsersRoute
+    }
+    '/posts/': {
+      id: '/posts/'
+      path: '/posts'
+      fullPath: '/posts'
+      preLoaderRoute: typeof PostsIndexRouteImport
+      parentRoute: typeof rootRouteImport
+    }
+    '/users/$userId': {
+      id: '/users/$userId'
+      path: '/$userId'
+      fullPath: '/users/$userId'
+      preLoaderRoute: typeof UsersUserIdRouteImport
+      parentRoute: typeof UsersRoute
+    }
+    '/posts/$postId': {
+      id: '/posts/$postId'
+      path: '/posts/$postId'
+      fullPath: '/posts/$postId'
+      preLoaderRoute: typeof PostsPostIdRouteImport
+      parentRoute: typeof rootRouteImport
+    }
+    '/api/classify': {
+      id: '/api/classify'
+      path: '/api/classify'
+      fullPath: '/api/classify'
+      preLoaderRoute: typeof ApiClassifyRouteImport
+      parentRoute: typeof rootRouteImport
+    }
+    '/posts_/$postId/deep': {
+      id: '/posts_/$postId/deep'
+      path: '/posts/$postId/deep'
+      fullPath: '/posts/$postId/deep'
+      preLoaderRoute: typeof PostsPostIdDeepRouteImport
+      parentRoute: typeof rootRouteImport
+    }
+  }
+}
+
+interface UsersRouteChildren {
+  UsersUserIdRoute: typeof UsersUserIdRoute
+  UsersIndexRoute: typeof UsersIndexRoute
+}
+
+const UsersRouteChildren: UsersRouteChildren = {
+  UsersUserIdRoute: UsersUserIdRoute,
+  UsersIndexRoute: UsersIndexRoute,
+}
+
+const UsersRouteWithChildren = UsersRoute._addFileChildren(UsersRouteChildren)
+
+const rootRouteChildren: RootRouteChildren = {
+  IndexRoute: IndexRoute,
+  UsersRoute: UsersRouteWithChildren,
+  ApiClassifyRoute: ApiClassifyRoute,
+  PostsPostIdRoute: PostsPostIdRoute,
+  PostsIndexRoute: PostsIndexRoute,
+  PostsPostIdDeepRoute: PostsPostIdDeepRoute,
+}
+export const routeTree = rootRouteImport
+  ._addFileChildren(rootRouteChildren)
+  ._addFileTypes<FileRouteTypes>()
+
+import type { getRouter } from './router.tsx'
+import type { createStart } from '@tanstack/react-start'
+declare module '@tanstack/react-start' {
+  interface Register {
+    ssr: true
+    router: Awaited<ReturnType<typeof getRouter>>
+  }
+}

examples-ts/classify/src/router.tsx

@@ -0,0 +1,15 @@
+import { createRouter } from '@tanstack/react-router'
+import { routeTree } from './routeTree.gen'
+import { DefaultCatchBoundary } from './components/DefaultCatchBoundary'
+import { NotFound } from './components/NotFound'
+
+export function getRouter() {
+  const router = createRouter({
+    routeTree,
+    defaultPreload: 'intent',
+    defaultErrorComponent: DefaultCatchBoundary,
+    defaultNotFoundComponent: () => <NotFound />,
+    scrollRestoration: true,
+  })
+  return router
+}

examples-ts/classify/src/routes/__root.tsx

@@ -0,0 +1,128 @@
+/// <reference types="vite/client" />
+import {
+  HeadContent,
+  Scripts,
+  createRootRoute,
+} from '@tanstack/react-router'
+import * as React from 'react'
+import { DefaultCatchBoundary } from '~/components/DefaultCatchBoundary'
+import { NotFound } from '~/components/NotFound'
+import { seo } from '~/utils/seo'
+
+export const Route = createRootRoute({
+  head: () => ({
+    meta: [
+      {
+        charSet: 'utf-8',
+      },
+      {
+        name: 'viewport',
+        content: 'width=device-width, initial-scale=1',
+      },
+      ...seo({
+        title:
+          'Financial Documents Classification Agent',
+        description: `Classify financial documents as balance sheets, income statements and cash flow statemets. `,
+      }),
+    ],
+    links: [
+      { rel: 'stylesheet', href: "https://cdn.jsdelivr.net/npm/daisyui@5" },
+      {
+        rel: 'apple-touch-icon',
+        sizes: '180x180',
+        href: '/apple-touch-icon.png',
+      },
+      {
+        rel: 'icon',
+        type: 'image/png',
+        sizes: '32x32',
+        href: '/favicon-32x32.png',
+      },
+      {
+        rel: 'icon',
+        type: 'image/png',
+        sizes: '16x16',
+        href: '/favicon-16x16.png',
+      },
+      { rel: 'manifest', href: '/site.webmanifest', color: '#fffff' },
+      { rel: 'icon', href: '/favicon.ico' },
+    ],
+    scripts: [
+      {
+        src: '/customScript.js',
+        type: 'text/javascript',
+      },
+      {
+        src: "https://cdn.jsdelivr.net/npm/@tailwindcss/browser@4",
+        type: "text/javascript",
+      }
+    ],
+  }),
+  errorComponent: DefaultCatchBoundary,
+  notFoundComponent: () => <NotFound />,
+  shellComponent: RootDocument,
+})
+
+function RootDocument({ children }: { children: React.ReactNode }) {
+  return (
+    <html>
+      <head>
+        <HeadContent />
+      </head>
+      <body>
+        <div className="navbar bg-base-100 shadow-sm">
+          <div className="navbar-start">
+            <div className="dropdown">
+              <div tabIndex={0} role="button" className="btn btn-ghost btn-circle">
+                <svg
+                  xmlns="http://www.w3.org/2000/svg"
+                  className="h-5 w-5"
+                  fill="none"
+                  viewBox="0 0 24 24"
+                  stroke="currentColor"
+                >
+                  <path
+                    strokeLinecap="round"
+                    strokeLinejoin="round"
+                    strokeWidth="2"
+                    d="M4 6h16M4 12h16M4 18h7"
+                  />
+                </svg>
+              </div>
+              <ul
+                tabIndex={0}
+                className="menu menu-lg dropdown-content bg-base-100 rounded-box z-1 mt-3 w-80 p-2 shadow"
+              >
+                <li><a href="/">Home</a></li>
+                <li><a href="https://cloud.llamaindex.ai">Get Started with LlamaCloud</a></li>
+                <li><a href="https://developers.llamaindex.ai/python/cloud/llamaclassify/getting_started/">LlamaClassify Docs</a></li>
+              </ul>
+            </div>
+          </div>
+          <div className="navbar-center">
+            <a className="btn btn-ghost text-xl" href="/">Financial Documents Classification Agent</a>
+          </div>
+          <div className="navbar-end">
+            <a href="https://github.com/run-llama/llama_cloud_services/main/blob/examples-ts/classify">
+              <button className="btn btn-ghost btn-circle">
+                <div className="indicator">
+                  <svg
+                    xmlns="http://www.w3.org/2000/svg"
+                    className="h-10 w-10"
+                    fill="currentColor"
+                    viewBox="0 0 640 512"
+                  >
+                    <path d="M237.9 461.4C237.9 463.4 235.6 465 232.7 465C229.4 465.3 227.1 463.7 227.1 461.4C227.1 459.4 229.4 457.8 232.3 457.8C235.3 457.5 237.9 459.1 237.9 461.4zM206.8 456.9C206.1 458.9 208.1 461.2 211.1 461.8C213.7 462.8 216.7 461.8 217.3 459.8C217.9 457.8 216 455.5 213 454.6C210.4 453.9 207.5 454.9 206.8 456.9zM251 455.2C248.1 455.9 246.1 457.8 246.4 460.1C246.7 462.1 249.3 463.4 252.3 462.7C255.2 462 257.2 460.1 256.9 458.1C256.6 456.2 253.9 454.9 251 455.2zM316.8 72C178.1 72 72 177.3 72 316C72 426.9 141.8 521.8 241.5 555.2C254.3 557.5 258.8 549.6 258.8 543.1C258.8 536.9 258.5 502.7 258.5 481.7C258.5 481.7 188.5 496.7 173.8 451.9C173.8 451.9 162.4 422.8 146 415.3C146 415.3 123.1 399.6 147.6 399.9C147.6 399.9 172.5 401.9 186.2 425.7C208.1 464.3 244.8 453.2 259.1 446.6C261.4 430.6 267.9 419.5 275.1 412.9C219.2 406.7 162.8 398.6 162.8 302.4C162.8 274.9 170.4 261.1 186.4 243.5C183.8 237 175.3 210.2 189 175.6C209.9 169.1 258 202.6 258 202.6C278 197 299.5 194.1 320.8 194.1C342.1 194.1 363.6 197 383.6 202.6C383.6 202.6 431.7 169 452.6 175.6C466.3 210.3 457.8 237 455.2 243.5C471.2 261.2 481 275 481 302.4C481 398.9 422.1 406.6 366.2 412.9C375.4 420.8 383.2 435.8 383.2 459.3C383.2 493 382.9 534.7 382.9 542.9C382.9 549.4 387.5 557.3 400.2 555C500.2 521.8 568 426.9 568 316C568 177.3 455.5 72 316.8 72zM169.2 416.9C167.9 417.9 168.2 420.2 169.9 422.1C171.5 423.7 173.8 424.4 175.1 423.1C176.4 422.1 176.1 419.8 174.4 417.9C172.8 416.3 170.5 415.6 169.2 416.9zM158.4 408.8C157.7 410.1 158.7 411.7 160.7 412.7C162.3 413.7 164.3 413.4 165 412C165.7 410.7 164.7 409.1 162.7 408.1C160.7 407.5 159.1 407.8 158.4 408.8zM190.8 444.4C189.2 445.7 189.8 448.7 192.1 450.6C194.4 452.9 197.3 453.2 198.6 451.6C199.9 450.3 199.3 447.3 197.3 445.4C195.1 443.1 192.1 442.8 190.8 444.4zM179.4 429.7C177.8 430.7 177.8 433.3 179.4 435.6C181 437.9 183.7 438.9 185 437.9C186.6 436.6 186.6 434 185 431.7C183.6 429.4 181 428.4 179.4 429.7z" />
+                  </svg>
+                </div>
+              </button>
+            </a>
+          </div>
+        </div>
+        <hr />
+        {children}
+        <Scripts />
+      </body>
+    </html>
+  )
+}

examples-ts/classify/src/routes/api/classify.ts

@@ -0,0 +1,45 @@
+import { createFileRoute } from '@tanstack/react-router'
+import { classifier, classificationRules, parsingConfig } from '~/utils/classifier'
+
+export const Route = createFileRoute('/api/classify')({
+  component: RouteComponent,
+  server: {
+      handlers: {
+        POST: async ({ request }) => {
+          const body = await request.formData()
+          const fl = body.get("file") as File;
+          if (!fl) {
+                return new Response(JSON.stringify({"result": "you need to provide a file"}))
+            }
+          const buff = await fl.arrayBuffer()
+          const rawRes = await classifier.classify(
+            classificationRules,
+            parsingConfig,
+            [new Uint8Array(buff)],
+          )
+          const results = rawRes.items
+          let classification = ""
+
+          for (const result of results) {
+            if ("result" in result && result.result) {
+                classification += `
+                    <div class="card bg-base-100 shadow-xl p-6 mb-4">
+                        <div class="space-y-3">
+                        <p><span class="font-semibold">📄 Document:</span> ${fl.name}</p>
+                        <p><span class="font-semibold">🏷️ Type:</span> <span class="badge badge-primary">${result.result.type}</span></p>
+                        <p><span class="font-semibold">📊 Confidence:</span> ${result.result.confidence*100}%</p>
+                        <p><span class="font-semibold">💭 Reasoning:</span> ${result.result.reasoning}</p>
+                        </div>
+                    </div>
+                `
+            }
+          }
+          return new Response(JSON.stringify({"result": classification}))
+        },
+      },
+    },
+})
+
+function RouteComponent() {
+  return
+}

examples-ts/classify/src/routes/index.tsx

@@ -0,0 +1,99 @@
+import { createFileRoute } from '@tanstack/react-router'
+import { useRef, useState } from 'react'
+
+export const Route = createFileRoute('/')({
+  component: Home,
+})
+
+function Home() {
+  const [file, setFile] = useState<null | File>(null)
+  const fileInputRef = useRef<HTMLInputElement>(null)
+  const [reply, setReply] = useState<null | string>(null)
+  const [loading, setLoading] = useState<boolean>(false)
+  const handleFileChange = (event: React.ChangeEvent<HTMLInputElement>) => {
+    const selectedFile = event.target.files?.[0]
+    if (selectedFile) {
+      setFile(selectedFile)
+    }
+  }
+  const handleClearFile = () => {
+    if (file) {
+      setFile(null)
+    }
+    if (fileInputRef.current) {
+      fileInputRef.current.value = ''
+    }
+    if (reply) {
+      setReply(null)
+    }
+  }
+
+  const handleClassify = async () => {
+    if (!file) return
+
+    if (reply) {
+      setReply(null)
+    }
+    setLoading(true)
+    try {
+      const formData = new FormData()
+      formData.append('file', file)
+
+      const res = await fetch('/api/classify', {
+        method: 'POST',
+        body: formData,
+      })
+
+      const data = await res.json()
+      setReply(data.result)
+    } catch (error) {
+      console.error('Error:', error)
+    } finally {
+      setLoading(false)
+    }
+  }
+
+  return (
+    <div className="flex flex-col justify-center items-center gap-y-8">
+        <br />
+        <h1 className="text-xl font-bold text-gray-700">AI-Powered finacial document classification</h1>
+        <h2 className="text-lg font-semibold text-gray-500">Need help sorting out the financial documents jungle? Let our classification agent handle it!</h2>
+        <fieldset className="fieldset bg-base-100 border-base-300 rounded-box w-200 border p-4">
+            <legend className="fieldset-legend text-lg">Upload your financial document here</legend>
+            <label className="label flex justify-center">
+                <input type="file" className="file-input" onChange={handleFileChange} accept='application/pdf' ref={fileInputRef} />
+            </label>
+        </fieldset>
+        {file && (
+          <div className="flex flex-col justify-center items-center gap-y-8">
+          <p className="text-sm text-gray-600">Selected file: {file.name}</p>
+          <div className='grid grid-cols-2 gap-x-6'>
+            <button
+                type="button"
+                className='btn bg-gray-500 text-white shadow-lg hover:bg-gray-600 hover:shadow-xl rounded'
+                onClick={handleClassify}
+            >
+            Classify
+            </button>
+            <button
+                  onClick={handleClearFile}
+                  type="button"
+                  className="px-4 py-2 bg-red-300 text-black rounded hover:bg-red-400 hover:shadow-xl shadow-lg"
+              >
+              Clear
+            </button>
+          </div>
+          </div>
+        )}
+        {loading && (
+          <span className="loading loading-spinner text-primary"></span>
+        )}
+        {reply && (
+          <div
+            className="max-w-2xl w-full"
+            dangerouslySetInnerHTML={{ __html: reply }}
+          />
+        )}
+    </div>
+  )
+}

examples-ts/classify/src/utils/classifier.ts

@@ -0,0 +1,23 @@
+import { LlamaClassify, ClassifierRule, ClassifyParsingConfiguration } from "llama-cloud-services"
+
+export const classifier = new LlamaClassify(process.env.LLAMA_CLOUD_API_KEY);
+
+export const classificationRules: ClassifierRule[] = [
+    {
+        description: "Shows a company's assets, liabilities, and shareholders' equity at a specific point in time, providing a snapshot of financial position.",
+        type: "balance_sheet"
+    },
+    {
+        description: "Reports cash inflows and outflows from operating, investing, and financing activities, highlighting liquidity and cash management.",
+        type: "cash_flow_statement"
+    },
+    {
+        description: "Summarizes revenues, expenses, and profits over a period, indicating financial performance and profitability.",
+        type: "income_statement"
+    },
+];
+
+export const parsingConfig: ClassifyParsingConfiguration = {
+    lang: "en",
+    max_pages: 20,
+}

examples-ts/classify/src/utils/seo.ts

@@ -0,0 +1,33 @@
+export const seo = ({
+  title,
+  description,
+  keywords,
+  image,
+}: {
+  title: string
+  description?: string
+  image?: string
+  keywords?: string
+}) => {
+  const tags = [
+    { title },
+    { name: 'description', content: description },
+    { name: 'keywords', content: keywords },
+    { name: 'twitter:title', content: title },
+    { name: 'twitter:description', content: description },
+    { name: 'twitter:creator', content: '@tannerlinsley' },
+    { name: 'twitter:site', content: '@tannerlinsley' },
+    { name: 'og:type', content: 'website' },
+    { name: 'og:title', content: title },
+    { name: 'og:description', content: description },
+    ...(image
+      ? [
+          { name: 'twitter:image', content: image },
+          { name: 'twitter:card', content: 'summary_large_image' },
+          { name: 'og:image', content: image },
+        ]
+      : []),
+  ]
+
+  return tags
+}

examples-ts/classify/tsconfig.json

@@ -0,0 +1,22 @@
+{
+  "include": ["**/*.ts", "**/*.tsx"],
+  "compilerOptions": {
+    "strict": true,
+    "esModuleInterop": true,
+    "jsx": "react-jsx",
+    "module": "ESNext",
+    "moduleResolution": "Bundler",
+    "lib": ["DOM", "DOM.Iterable", "ES2022"],
+    "isolatedModules": true,
+    "resolveJsonModule": true,
+    "skipLibCheck": true,
+    "target": "ES2022",
+    "allowJs": true,
+    "forceConsistentCasingInFileNames": true,
+    "baseUrl": ".",
+    "paths": {
+      "~/*": ["./src/*"]
+    },
+    "noEmit": true
+  }
+}

examples-ts/classify/vite.config.ts

@@ -0,0 +1,19 @@
+import { tanstackStart } from '@tanstack/react-start/plugin/vite'
+import { defineConfig } from 'vite'
+import tsConfigPaths from 'vite-tsconfig-paths'
+import viteReact from '@vitejs/plugin-react'
+
+export default defineConfig({
+  server: {
+    port: 3000,
+  },
+  plugins: [
+    tsConfigPaths({
+      projects: ['./tsconfig.json'],
+    }),
+    tanstackStart({
+      srcDirectory: 'src',
+    }),
+    viteReact(),
+  ],
+})

examples-ts/extract/README.md

@@ -0,0 +1,122 @@
+# LlamaExtract Demo
+
+A TypeScript demo application showcasing the power of **LlamaExract** - a structured data extraction agentic service from [LlamaCloud](https://cloud.llamaindex.ai). This demo allows you to extract structured information from scientific papers and get them into a nice markdown format.
+
+## Table of Contents
+
+- [Features](#features)
+- [Prerequisites](#prerequisites)
+- [Installation](#installation)
+- [Usage](#usage)
+  - [Start the Demo](#start-the-demo)
+  - [Development Mode](#development-mode)
+  - [Build the Project](#build-the-project)
+  - [Code Quality](#code-quality)
+  - [Quick Commands Reference](#quick-commands-reference)
+- [How It Works](#how-it-works)
+- [API Dependencies](#api-dependencies)
+- [Troubleshooting](#troubleshooting)
+  - [Common Issues](#common-issues)
+- [License](#license)
+- [Contributing](#contributing)
+
+## Features
+
+- 📄 **Structured Data Extraction**: Extract data from your files effortlessly, and structure them the way you want!
+- 🤖 **Markdown Rendering**: Generate markdown directly from your extracted data
+- 🎨 **Beautiful CLI**: Styled console interface with colors and ASCII art
+- ⚡ **Fast Development**: Hot reload support with watch mode
+- 🛠️ **TypeScript**: Full TypeScript support with strict type checking
+
+## Prerequisites
+
+- Node.js (version 18 or higher)
+- pnpm package manager
+- LlamaCloud API key
+
+## Installation
+
+1. Clone the repository:
+
+```bash
+git clone https://github.com/run-llama/llama_cloud_services
+cd lama_cloud_services/examples-ts/extract/
+```
+
+2. Install dependencies:
+
+```bash
+npm install
+```
+
+3. Set up your environment variables:
+
+```bash
+# Add your API key to your environment
+export LLAMA_CLOUD_API_KEY="your-llamacloud-api-key"
+```
+
+## Usage
+
+### Start the Demo
+
+```bash
+npm run start
+```
+
+The application will display a welcome screen and prompt you to enter the path to a document you'd like to process.
+
+### Development Mode
+
+For development with hot reload:
+
+```bash
+npm run dev
+```
+
+### Build the Project
+
+```bash
+npm run build
+```
+
+### Code Quality
+
+Format code:
+
+```bash
+npm run format
+```
+
+Lint code:
+
+```bash
+npm run lint
+```
+
+## How It Works
+
+1. **Document Input**: Enter the path to your document when prompted
+2. **Parsing**: LlamaExtract, based on the schema you can find [here](./src/schema.ts), processes the document and extracts structured data
+3. **Markdown Rendering**: The extracted content is rendered into beautiful markdown
+4. **Results**: View the results directly in your terminal
+
+## Troubleshooting
+
+### Common Issues
+
+1. **Module Resolution Errors**: Ensure you're using Node.js 18+ and have all dependencies installed
+2. **API Key Issues**: Verify your LlamaCloud API key is correctly set
+3. **File Path Errors**: Use absolute paths or ensure relative paths are correct from the project root
+
+## License
+
+MIT License - see the [LICENSE](../../LICENSE) file for details.
+
+## Contributing
+
+1. Fork the repository
+2. Create a feature branch
+3. Make your changes
+4. Run `npm run format` and `npm run lint`
+5. Submit a pull request

examples-ts/extract/eslint.config.mts

@@ -0,0 +1,14 @@
+import js from "@eslint/js";
+import globals from "globals";
+import tseslint from "typescript-eslint";
+import { defineConfig } from "eslint/config";
+
+export default defineConfig([
+  {
+    files: ["**/*.{js,mjs,cjs,ts,mts,cts}"],
+    plugins: { js },
+    extends: ["js/recommended"],
+    languageOptions: { globals: globals.browser },
+  },
+  tseslint.configs.recommended,
+]);

examples-ts/extract/package.json

@@ -0,0 +1,37 @@
+{
+  "name": "llama-extract-demo",
+  "version": "0.1.0",
+  "description": "Demo for LlamaExtract in TypeScript",
+  "main": "index.js",
+  "scripts": {
+    "test": "echo \"There are no tests\"",
+    "start": "npm exec tsx src/index.ts",
+    "lint": "eslint ./src/",
+    "format": "prettier --write ./src/",
+    "build": "tsc",
+    "dev": "npm exec tsx --watch src/index.ts"
+  },
+  "author": "LlamaIndex",
+  "license": "MIT",
+  "dependencies": {
+    "cli-markdown": "^3.5.1",
+    "consola": "^3.4.2",
+    "figlet": "^1.8.2",
+    "llama-cloud-services": "file:../../ts/llama_cloud_services",
+    "marked": "^15.0.12",
+    "marked-terminal": "^7.3.0",
+    "picocolors": "^1.1.1"
+  },
+  "devDependencies": {
+    "@eslint/js": "^9.32.0",
+    "@types/figlet": "^1.7.0",
+    "@types/marked-terminal": "^6.1.1",
+    "@types/node": "^24.2.0",
+    "eslint": "^9.32.0",
+    "globals": "^16.3.0",
+    "jiti": "^2.5.1",
+    "prettier": "^3.6.2",
+    "typescript": "^5.9.2",
+    "typescript-eslint": "^8.39.0"
+  }
+}

examples-ts/extract/src/index.ts

@@ -0,0 +1,47 @@
+import { LlamaExtract, ExtractConfig } from "llama-cloud-services";
+import cliMarkdown from "cli-markdown";
+import { logger } from "./logger";
+import pc from "picocolors";
+import { consoleInput, renderLogo } from "./utils";
+import { dataSchema } from "./schema";
+import { renderMarkdown, ResearchData } from "./markdown";
+
+export async function main(): Promise<number> {
+  const extractClient = new LlamaExtract(
+    process.env.LLAMA_CLOUD_API_KEY!,
+    "https://api.cloud.llamaindex.ai",
+  );
+  await renderLogo();
+  logger.log(
+    `Welcome to ${pc.bold(
+      pc.magentaBright("LlamaExtract Demo✨"),
+    )}, our demo for ${pc.bold(pc.green("LlamaExtract"))}, a ${pc.bold(
+      pc.cyan("LlamaCloud☁️"),
+    )}  (https://cloud.llamaindex.ai) product!.\nIn this demo we are going to try extracting relevant information  ${pc.bold(
+      pc.yellowBright("from scientific papers"),
+    )}. Type the path to the paper you would like to process below👇\nIf you wish to exit, just type ${pc.bold(
+      pc.gray("quit"),
+    )}.\n`,
+  );
+  while (true) {
+    const userInput = await consoleInput();
+    if (userInput.toLowerCase() == "quit") {
+      break;
+    }
+    try {
+      const generatedData = await extractClient.extract(
+        dataSchema,
+        {} as ExtractConfig,
+        userInput,
+      );
+      const research = renderMarkdown(generatedData?.data as ResearchData); // Added await here
+      logger.log(`${pc.bold(pc.cyan("Extracted information:✨"))}:\n`);
+      logger.log(cliMarkdown(research));
+    } catch (error) {
+      logger.error(`Error processing file: ${error}`);
+    }
+  }
+  return 0;
+}
+
+main().catch(console.error);

examples-ts/extract/src/logger.ts

@@ -0,0 +1,8 @@
+import { createConsola } from "consola";
+import type { ConsolaInstance } from "consola";
+
+export const logger: ConsolaInstance = createConsola({
+  formatOptions: {
+    date: false,
+  },
+});

examples-ts/extract/src/markdown.ts

@@ -0,0 +1,172 @@
+type Author = {
+  name: string;
+  affiliation?: string;
+  email?: string;
+};
+
+type Methodology = {
+  approach?: string;
+  participants?: string;
+  methods?: string[];
+};
+
+type Result = {
+  finding?: string;
+  significance?: string;
+  supportingData?: string;
+};
+
+type Reference = {
+  title: string;
+  authors: string;
+  year?: string;
+  relevance?: string;
+};
+
+type Discussion = {
+  implications?: string[];
+  limitations?: string[];
+  futureWork?: string[];
+};
+
+type Publication = {
+  journal?: string;
+  year: string;
+  doi?: string;
+  url?: string;
+};
+
+export type ResearchData = {
+  title: string;
+  authors: Author[];
+  abstract: string;
+  keywords?: string[];
+  mainFindings: string[];
+  methodology?: Methodology;
+  results?: Result[];
+  discussion?: Discussion;
+  references?: Reference[];
+  publication?: Publication;
+};
+
+export function renderMarkdown(data: ResearchData): string {
+  const {
+    title,
+    authors,
+    abstract,
+    keywords,
+    mainFindings,
+    methodology,
+    results,
+    discussion,
+    references,
+    publication,
+  } = data;
+
+  const md: string[] = [];
+
+  md.push(`# ${title}\n`);
+
+  // Authors
+  md.push(`## Authors`);
+  md.push(
+    authors
+      .map(
+        (author) =>
+          `- **${author.name}**${
+            author.affiliation ? `, *${author.affiliation}*` : ""
+          }${author.email ? ` (${author.email})` : ""}`,
+      )
+      .join("\n"),
+  );
+
+  // Abstract
+  md.push(`\n## Abstract\n${abstract}`);
+
+  // Keywords
+  if (keywords && keywords.length > 0) {
+    md.push(`\n## Keywords\n${keywords.map((k) => `- ${k}`).join("\n")}`);
+  }
+
+  // Main Findings
+  md.push(
+    `\n## Main Findings\n${mainFindings.map((f) => `- ${f}`).join("\n")}`,
+  );
+
+  // Methodology
+  if (methodology) {
+    md.push(`\n## Methodology`);
+    if (methodology.approach) md.push(`**Approach:** ${methodology.approach}`);
+    if (methodology.participants)
+      md.push(`**Participants:** ${methodology.participants}`);
+    if (methodology.methods?.length) {
+      md.push(
+        `**Methods:**\n${methodology.methods.map((m) => `- ${m}`).join("\n")}`,
+      );
+    }
+  }
+
+  // Results
+  if (results?.length) {
+    md.push(`\n## Results`);
+    results.forEach((result, i) => {
+      md.push(`\n### Result ${i + 1}`);
+      if (result.finding) md.push(`- **Finding:** ${result.finding}`);
+      if (result.significance)
+        md.push(`- **Significance:** ${result.significance}`);
+      if (result.supportingData)
+        md.push(`- **Supporting Data:** ${result.supportingData}`);
+    });
+  }
+
+  // Discussion
+  if (discussion) {
+    md.push(`\n## Discussion`);
+    if (discussion.implications?.length) {
+      md.push(
+        `### Implications\n${discussion.implications
+          .map((d) => `- ${d}`)
+          .join("\n")}`,
+      );
+    }
+    if (discussion.limitations?.length) {
+      md.push(
+        `### Limitations\n${discussion.limitations
+          .map((d) => `- ${d}`)
+          .join("\n")}`,
+      );
+    }
+    if (discussion.futureWork?.length) {
+      md.push(
+        `### Future Work\n${discussion.futureWork
+          .map((d) => `- ${d}`)
+          .join("\n")}`,
+      );
+    }
+  }
+
+  // References
+  if (references?.length) {
+    md.push(`\n## References`);
+    references.forEach((ref, i) => {
+      md.push(
+        `\n**[${i + 1}]** ${ref.title} — *${ref.authors}*${
+          ref.year ? ` (${ref.year})` : ""
+        }`,
+      );
+      if (ref.relevance) md.push(`> ${ref.relevance}`);
+    });
+  }
+
+  // Publication Info
+  if (publication) {
+    md.push(`\n## Publication`);
+    if (publication.journal) md.push(`- **Journal:** ${publication.journal}`);
+    if (publication.year) md.push(`- **Year:** ${publication.year}`);
+    if (publication.doi) md.push(`- **DOI:** ${publication.doi}`);
+    if (publication.url)
+      md.push(`- **URL:** [${publication.url}](${publication.url})`);
+  }
+
+  return md.join("\n");
+}

examples-ts/extract/src/schema.ts

@@ -0,0 +1,169 @@
+export const dataSchema = {
+  type: "object",
+  required: ["title", "authors", "abstract", "mainFindings"],
+  properties: {
+    title: {
+      type: "string",
+      description: "The full title of the research paper",
+    },
+    authors: {
+      type: "array",
+      description: "List of all authors of the paper",
+      items: {
+        type: "object",
+        properties: {
+          name: {
+            type: "string",
+            description: "Full name of the author",
+          },
+          affiliation: {
+            type: "string",
+            description:
+              "Institution or organization the author is affiliated with",
+          },
+          email: {
+            type: "string",
+            description: "Contact email of the author if provided",
+          },
+        },
+      },
+    },
+    abstract: {
+      type: "string",
+      description: "Complete abstract or summary of the paper",
+    },
+    keywords: {
+      type: "array",
+      description:
+        "Key terms and phrases that describe the paper's main topics",
+      items: {
+        type: "string",
+      },
+    },
+    mainFindings: {
+      type: "array",
+      description: "Key findings, conclusions, or contributions of the paper",
+      items: {
+        type: "string",
+      },
+    },
+    methodology: {
+      type: "object",
+      description: "Research methods and approaches used",
+      properties: {
+        approach: {
+          type: "string",
+          description: "Overall research approach or study design",
+        },
+        participants: {
+          type: "string",
+          description: "Description of study participants or data sources",
+        },
+        methods: {
+          type: "array",
+          description: "Specific methods, techniques, or tools used",
+          items: {
+            type: "string",
+          },
+        },
+      },
+    },
+    results: {
+      type: "array",
+      description: "Main results and outcomes of the research",
+      items: {
+        type: "object",
+        properties: {
+          finding: {
+            type: "string",
+            description: "Description of the specific result or finding",
+          },
+          significance: {
+            type: "string",
+            description:
+              "Statistical significance or importance of the finding",
+          },
+          supportingData: {
+            type: "string",
+            description: "Relevant statistics, measurements, or data points",
+          },
+        },
+      },
+    },
+    discussion: {
+      type: "object",
+      properties: {
+        implications: {
+          type: "array",
+          description: "Theoretical or practical implications of the findings",
+          items: {
+            type: "string",
+          },
+        },
+        limitations: {
+          type: "array",
+          description: "Study limitations or constraints",
+          items: {
+            type: "string",
+          },
+        },
+        futureWork: {
+          type: "array",
+          description: "Suggested future research directions",
+          items: {
+            type: "string",
+          },
+        },
+      },
+    },
+    references: {
+      type: "array",
+      description:
+        "Key papers cited that are crucial to understanding this work",
+      items: {
+        type: "object",
+        properties: {
+          title: {
+            type: "string",
+            description: "Title of the cited paper",
+          },
+          authors: {
+            type: "string",
+            description: "Authors of the cited paper",
+          },
+          year: {
+            type: "string",
+            description: "Publication year",
+          },
+          relevance: {
+            type: "string",
+            description: "Why this reference is important to the current paper",
+          },
+        },
+        required: ["title", "authors"],
+      },
+    },
+    publication: {
+      type: "object",
+      properties: {
+        journal: {
+          type: "string",
+          description: "Name of the journal or conference",
+        },
+        year: {
+          type: "string",
+          description: "Year of publication",
+        },
+        doi: {
+          type: "string",
+          description: "Digital Object Identifier (DOI) of the paper",
+        },
+        url: {
+          type: "string",
+          description: "URL where the paper can be accessed",
+        },
+      },
+      required: ["year"],
+    },
+  },
+};

examples-ts/extract/src/types/cli-markdown.d.ts

@@ -0,0 +1,4 @@
+declare module "cli-markdown" {
+  function cliMarkdown(input: string): string;
+  export default cliMarkdown;
+}

examples-ts/extract/src/utils.ts

@@ -0,0 +1,33 @@
+import * as readline from "readline/promises";
+import figlet from "figlet";
+import pc from "picocolors";
+
+export async function renderLogo(): Promise<void> {
+  const logoText = figlet.textSync("Extract Demo", {
+    font: "ANSI Shadow",
+    horizontalLayout: "default",
+    verticalLayout: "default",
+    width: 100,
+    whitespaceBreak: true,
+  });
+
+  // Add some styling with picocolors
+  const styledLogo = pc.bold(pc.redBright(logoText));
+
+  // Add some padding/margin
+  console.log("\n");
+  console.log(styledLogo);
+  console.log(pc.gray("─".repeat(60)));
+  console.log("\n");
+}
+
+export async function consoleInput(): Promise<string> {
+  const rl = readline.createInterface({
+    input: process.stdin,
+    output: process.stdout,
+  });
+
+  const answer = await rl.question("Path to your file: ");
+  rl.close();
+  return answer;
+}

examples-ts/index/README.md

@@ -0,0 +1,131 @@
+# LlamaCloud Index Demo
+
+A TypeScript demo application showcasing the power of **LlamaCloud Index** - a fully automated document ingestion and retrieval serviced offered within [LlamaCloud](https://cloud.llamaindex.ai). This demo allows you to ask questions, retrieve relevant contextual information and generate AI-powered responses using OpenAI's GPT models.
+
+## Table of Contents
+
+- [Features](#features)
+- [Prerequisites](#prerequisites)
+- [Installation](#installation)
+- [Usage](#usage)
+  - [Start the Demo](#start-the-demo)
+  - [Development Mode](#development-mode)
+  - [Build the Project](#build-the-project)
+  - [Code Quality](#code-quality)
+  - [Quick Commands Reference](#quick-commands-reference)
+- [How It Works](#how-it-works)
+- [API Dependencies](#api-dependencies)
+- [Troubleshooting](#troubleshooting)
+  - [Common Issues](#common-issues)
+- [License](#license)
+- [Contributing](#contributing)
+
+## Features
+
+- 🤖 **RAG**: Simple-yet-effective Retrieval Augmented Generation pipeline built on top of LlamaCloud Index and OpenAI
+- 🎨 **Beautiful CLI**: Styled console interface with colors and ASCII art
+- ⚡ **Fast Development**: Hot reload support with watch mode
+- 🛠️ **TypeScript**: Full TypeScript support with strict type checking
+
+## Prerequisites
+
+- Node.js (version 18 or higher)
+- pnpm package manager
+- OpenAI API key
+- LlamaCloud API key
+- An existing LlamaCloud Index pipeline
+
+## Installation
+
+1. Clone the repository:
+
+```bash
+git clone https://github.com/run-llama/llama_cloud_services
+cd lama_cloud_services/examples-ts/index/
+```
+
+2. Install dependencies:
+
+```bash
+pnpm install
+```
+
+3. Set up your environment variables:
+
+```bash
+export OPENAI_API_KEY="your-openai-api-key"
+export LLAMA_CLOUD_API_KEY="your-llamacloud-api-key"
+export PIPELINE_NAME="your-pipeline-name"
+```
+
+4. Or write them into a `.env` file:
+
+```env
+OPENAI_API_KEY="your-openai-api-key"
+LLAMA_CLOUD_API_KEY="your-llamacloud-api-key"
+PIPELINE_NAME="your-pipeline-name"
+```
+
+## Usage
+
+### Start the Demo
+
+```bash
+pnpm run start
+```
+
+The application will display a welcome screen and prompt you to start chatting!
+
+### Development Mode
+
+For development with hot reload:
+
+```bash
+pnpm run dev
+```
+
+### Build the Project
+
+```bash
+pnpm run build
+```
+
+### Code Quality
+
+Format code:
+
+```bash
+pnpm run format
+```
+
+Lint code:
+
+```bash
+pnpm run lint
+```
+
+## How It Works
+
+1. **Message Input**: Enter a message
+2. **Retrieval**: Several nodes are retrieved from the LlamaCloud index you specified
+3. **AI Response Generation**: The retrieved information is passed on to the AI model, along with its relevance score, and a reply to your original message is generated starting from that.
+4. **Results**: View the AI-generated summary in your terminal
+
+## Troubleshooting
+
+### Common Issues
+
+1. **Module Resolution Errors**: Ensure you're using Node.js 18+ and have all dependencies installed
+2. **API Key Issues**: Verify your OpenAI and LlamaCloud API keys are correctly set
+
+## License
+
+MIT License - see the [LICENSE](../../LICENSE) file for details.
+
+## Contributing
+
+1. Fork the repository
+2. Create a feature branch
+3. Make your changes
+4. Run `pnpm run format` and `pnpm run lint`
+5. Submit a pull request

examples-ts/index/eslint.config.mts

@@ -0,0 +1,15 @@
+import js from "@eslint/js";
+import globals from "globals";
+import tseslint from "typescript-eslint";
+import { defineConfig } from "eslint/config";
+
+export default defineConfig([
+  {
+    files: ["**/*.{js,mjs,cjs,ts,mts,cts}"],
+    plugins: { js },
+    extends: ["js/recommended"],
+    languageOptions: { globals: globals.browser },
+  },
+  { files: ["**/*.js"], languageOptions: { sourceType: "script" } },
+  tseslint.configs.recommended,
+]);

examples-ts/index/package.json

@@ -0,0 +1,48 @@
+{
+  "name": "llama-chat",
+  "version": "0.1.0",
+  "description": "Demo for LlamaCloud Index in TypeScript",
+  "type": "module",
+  "main": "index.js",
+  "scripts": {
+    "test": "echo \"There are no tests\"",
+    "start": "pnpm exec tsx src/index.ts",
+    "lint": "eslint ./src/",
+    "format": "prettier --write ./src/",
+    "build": "tsc",
+    "dev": "pnpm exec tsx --watch src/index.ts"
+  },
+  "keywords": [
+    "ai",
+    "rag",
+    "retrieval",
+    "pipeline",
+    "llms",
+    "chatbot"
+  ],
+  "author": "LlamaIndex",
+  "license": "MIT",
+  "packageManager": "pnpm@10.12.4",
+  "devDependencies": {
+    "@eslint/js": "^9.32.0",
+    "@types/figlet": "^1.7.0",
+    "@types/node": "^24.1.0",
+    "@typescript-eslint/eslint-plugin": "^8.38.0",
+    "@typescript-eslint/parser": "^8.38.0",
+    "eslint": "^9.32.0",
+    "globals": "^16.3.0",
+    "jiti": "^2.5.1",
+    "prettier": "^3.6.2",
+    "typescript": "^5.8.3",
+    "typescript-eslint": "^8.38.0"
+  },
+  "dependencies": {
+    "@ai-sdk/openai": "^1.3.23",
+    "ai": "^4.3.19",
+    "consola": "^3.4.2",
+    "dotenv": "^17.2.1",
+    "figlet": "^1.8.2",
+    "llama-cloud-services": "link:../../ts/llama_cloud_services",
+    "picocolors": "^1.1.1"
+  }
+}

examples-ts/index/src/index.ts

@@ -0,0 +1,48 @@
+import { LlamaCloudIndex } from "llama-cloud-services";
+import { logger } from "./logger";
+import pc from "picocolors";
+import {
+  consoleInput,
+  retrievalAugmentedGeneration,
+  renderLogo,
+} from "./utils";
+import dotenv from "dotenv";
+
+dotenv.config();
+
+export async function main(): Promise<number> {
+  const index = new LlamaCloudIndex({
+    name: process.env.PIPELINE_NAME as string,
+    projectName: "Default",
+    apiKey: process.env.LLAMA_CLOUD_API_KEY, // can provide API-key in the constructor or in the env
+  });
+  const retriever = index.asRetriever({
+    similarityTopK: 5,
+  });
+  await renderLogo();
+  logger.log(
+    `Welcome to ${pc.bold(
+      pc.magentaBright("✨LlamaChat✨"),
+    )}, our demo for ${pc.bold(pc.green("Index🦙"))}, a ${pc.bold(
+      pc.cyan("LlamaCloud☁️"),
+    )}  (https://cloud.llamaindex.ai) product!.\nType a question below, and you will get an answer!👇\nIf you wish to exit, just type ${pc.bold(
+      pc.gray("quit"),
+    )}.\n`,
+  );
+  while (true) {
+    const userInput = await consoleInput();
+    if (userInput.toLowerCase() == "quit") {
+      break;
+    }
+    try {
+      const nodes = await retriever.retrieve(userInput);
+      const summary = await retrievalAugmentedGeneration(nodes, userInput);
+      logger.log(`${pc.bold(pc.magentaBright("LlamaChat✨:"))}\n${summary}`);
+    } catch (error) {
+      logger.error(`Error processing your request: ${error}`);
+    }
+  }
+  return 0;
+}
+
+main().catch(console.error);

examples-ts/index/src/logger.ts

@@ -0,0 +1,8 @@
+import { createConsola } from "consola";
+import type { ConsolaInstance } from "consola";
+
+export const logger: ConsolaInstance = createConsola({
+  formatOptions: {
+    date: false,
+  },
+});

examples-ts/index/src/utils.ts

@@ -0,0 +1,56 @@
+import { generateText } from "ai";
+import { openai } from "@ai-sdk/openai";
+import { NodeWithScore, MetadataMode } from "llamaindex";
+import * as readline from "readline/promises";
+import figlet from "figlet";
+import pc from "picocolors";
+
+export async function renderLogo(): Promise<void> {
+  const logoText = figlet.textSync("LlamaChat", {
+    font: "ANSI Shadow",
+    horizontalLayout: "default",
+    verticalLayout: "default",
+    width: 100,
+    whitespaceBreak: true,
+  });
+
+  // Add some styling with picocolors
+  const styledLogo = pc.bold(pc.yellowBright(logoText));
+
+  // Add some padding/margin
+  console.log("\n");
+  console.log(styledLogo);
+  console.log(pc.gray("─".repeat(60)));
+  console.log("\n");
+}
+
+export async function consoleInput(): Promise<string> {
+  const rl = readline.createInterface({
+    input: process.stdin,
+    output: process.stdout,
+  });
+
+  const answer = await rl.question(pc.cyanBright("You✨:"));
+  rl.close();
+  return answer;
+}
+
+export async function retrievalAugmentedGeneration(
+  nodes: NodeWithScore[],
+  prompt: string,
+): Promise<string> {
+  let mainText: string = "";
+
+  for (const node of nodes) {
+    mainText += `\t{information: '${node.node.getContent(
+      MetadataMode.ALL,
+    )}', relevanceScore: '${node.score ?? "no score"}'}\n`;
+  }
+
+  const { text } = await generateText({
+    model: openai("gpt-4.1"),
+    prompt: `[\n${mainText}\n]\n\nBased on the information you are given and on the relevance score of that (where -1 means no score available), answer to this user prompt: '${prompt}'`,
+  });
+
+  return text;
+}

examples-ts/index/tsconfig.json

@@ -0,0 +1,22 @@
+{
+  "compilerOptions": {
+    "target": "ES2022",
+    "module": "ES2022",
+    "lib": ["ES2022"],
+    "outDir": "./dist",
+    "rootDir": "./src",
+    "strict": true,
+    "esModuleInterop": true,
+    "skipLibCheck": true,
+    "forceConsistentCasingInFileNames": true,
+    "declaration": true,
+    "declarationMap": true,
+    "sourceMap": true,
+    "types": ["node"],
+    "moduleResolution": "bundler",
+    "allowSyntheticDefaultImports": true,
+    "resolveJsonModule": true
+  },
+  "include": ["src/**/*"],
+  "exclude": ["node_modules", "dist"]
+}

examples-ts/parse/README.md

@@ -0,0 +1,124 @@
+# LlamaParse Demo
+
+A TypeScript demo application showcasing the power of **LlamaParse** - an intelligent document parsing service from [LlamaCloud](https://cloud.llamaindex.ai). This demo allows you to parse various document formats and generate AI-powered summaries using OpenAI's GPT models.
+
+## Table of Contents
+
+- [Features](#features)
+- [Prerequisites](#prerequisites)
+- [Installation](#installation)
+- [Usage](#usage)
+  - [Start the Demo](#start-the-demo)
+  - [Development Mode](#development-mode)
+  - [Build the Project](#build-the-project)
+  - [Code Quality](#code-quality)
+  - [Quick Commands Reference](#quick-commands-reference)
+- [How It Works](#how-it-works)
+- [API Dependencies](#api-dependencies)
+- [Troubleshooting](#troubleshooting)
+  - [Common Issues](#common-issues)
+- [License](#license)
+- [Contributing](#contributing)
+
+## Features
+
+- 📄 **Document Parsing**: Parse PDFs, Word docs, and other formats using LlamaParse
+- 🤖 **AI Summaries**: Generate intelligent summaries using OpenAI GPT-4
+- 🎨 **Beautiful CLI**: Styled console interface with colors and ASCII art
+- ⚡ **Fast Development**: Hot reload support with watch mode
+- 🛠️ **TypeScript**: Full TypeScript support with strict type checking
+
+## Prerequisites
+
+- Node.js (version 18 or higher)
+- pnpm package manager
+- OpenAI API key
+- LlamaCloud API key
+
+## Installation
+
+1. Clone the repository:
+
+```bash
+git clone https://github.com/run-llama/llama_cloud_services
+cd lama_cloud_services/examples-ts/parse/
+```
+
+2. Install dependencies:
+
+```bash
+pnpm install
+```
+
+3. Set up your environment variables:
+
+```bash
+# Add your API keys to your environment
+export OPENAI_API_KEY="your-openai-api-key"
+export LLAMA_CLOUD_API_KEY="your-llamacloud-api-key"
+```
+
+## Usage
+
+### Start the Demo
+
+```bash
+pnpm run start
+```
+
+The application will display a welcome screen and prompt you to enter the path to a document you'd like to process.
+
+### Development Mode
+
+For development with hot reload:
+
+```bash
+pnpm run dev
+```
+
+### Build the Project
+
+```bash
+pnpm run build
+```
+
+### Code Quality
+
+Format code:
+
+```bash
+pnpm run format
+```
+
+Lint code:
+
+```bash
+pnpm run lint
+```
+
+## How It Works
+
+1. **Document Input**: Enter the path to your document when prompted
+2. **Parsing**: LlamaParse processes the document and extracts structured content
+3. **AI Summary**: The extracted content is sent to OpenAI GPT-4 for summarization
+4. **Results**: View the AI-generated summary in your terminal
+
+## Troubleshooting
+
+### Common Issues
+
+1. **Module Resolution Errors**: Ensure you're using Node.js 18+ and have all dependencies installed
+2. **API Key Issues**: Verify your OpenAI and LlamaCloud API keys are correctly set
+3. **File Path Errors**: Use absolute paths or ensure relative paths are correct from the project root
+
+## License
+
+MIT License - see the [LICENSE](../../LICENSE) file for details.
+
+## Contributing
+
+1. Fork the repository
+2. Create a feature branch
+3. Make your changes
+4. Run `pnpm run format` and `pnpm run lint`
+5. Submit a pull request

examples-ts/parse/eslint.config.mts

@@ -0,0 +1,15 @@
+import js from "@eslint/js";
+import globals from "globals";
+import tseslint from "typescript-eslint";
+import { defineConfig } from "eslint/config";
+
+export default defineConfig([
+  {
+    files: ["**/*.{js,mjs,cjs,ts,mts,cts}"],
+    plugins: { js },
+    extends: ["js/recommended"],
+    languageOptions: { globals: globals.browser },
+  },
+  { files: ["**/*.js"], languageOptions: { sourceType: "script" } },
+  tseslint.configs.recommended,
+]);

examples-ts/parse/package.json

@@ -0,0 +1,47 @@
+{
+  "name": "llamaparse-demo",
+  "version": "0.1.0",
+  "description": "Demo for LlamaParse in TypeScript",
+  "type": "module",
+  "main": "index.js",
+  "scripts": {
+    "test": "echo \"There are no tests\"",
+    "start": "pnpm exec tsx src/index.ts",
+    "lint": "eslint ./src/",
+    "format": "prettier --write ./src/",
+    "build": "tsc",
+    "dev": "pnpm exec tsx --watch src/index.ts"
+  },
+  "keywords": [
+    "ai",
+    "ocr",
+    "parsing",
+    "intelligent-document-processing",
+    "pdf",
+    "llms"
+  ],
+  "author": "LlamaIndex",
+  "license": "MIT",
+  "packageManager": "pnpm@10.12.4",
+  "devDependencies": {
+    "@eslint/js": "^9.32.0",
+    "@types/figlet": "^1.7.0",
+    "@types/node": "^24.1.0",
+    "@typescript-eslint/eslint-plugin": "^8.38.0",
+    "@typescript-eslint/parser": "^8.38.0",
+    "eslint": "^9.32.0",
+    "globals": "^16.3.0",
+    "jiti": "^2.5.1",
+    "prettier": "^3.6.2",
+    "typescript": "^5.8.3",
+    "typescript-eslint": "^8.38.0"
+  },
+  "dependencies": {
+    "@ai-sdk/openai": "^1.3.23",
+    "ai": "^4.3.19",
+    "consola": "^3.4.2",
+    "figlet": "^1.8.2",
+    "llama-cloud-services": "link:../../ts/llama_cloud_services",
+    "picocolors": "^1.1.1"
+  }
+}

examples-ts/parse/src/index.ts

@@ -0,0 +1,34 @@
+import { LlamaParseReader } from "llama-cloud-services";
+import { logger } from "./logger";
+import pc from "picocolors";
+import { consoleInput, generateSummary, renderLogo } from "./utils";
+
+export async function main(): Promise<number> {
+  const reader = new LlamaParseReader({ resultType: "markdown" });
+  await renderLogo();
+  logger.log(
+    `Welcome to ${pc.bold(
+      pc.magentaBright("✨LlamaParse Demo✨"),
+    )}, our demo for ${pc.bold(pc.green("LlamaParse🦙"))}, a ${pc.bold(
+      pc.cyan("LlamaCloud☁️"),
+    )}  (https://cloud.llamaindex.ai) product!.\nType the path to the document you would like to process below👇\nIf you wish to exit, just type ${pc.bold(
+      pc.gray("quit"),
+    )}.\n`,
+  );
+  while (true) {
+    const userInput = await consoleInput();
+    if (userInput.toLowerCase() == "quit") {
+      break;
+    }
+    try {
+      const documents = await reader.loadData(userInput);
+      const summary = await generateSummary(documents); // Added await here
+      logger.log(`${pc.bold(pc.cyan("AI-generated summary✨"))}:\n${summary}`);
+    } catch (error) {
+      logger.error(`Error processing file: ${error}`);
+    }
+  }
+  return 0;
+}
+
+main().catch(console.error);

examples-ts/parse/src/logger.ts

@@ -0,0 +1,8 @@
+import { createConsola } from "consola";
+import type { ConsolaInstance } from "consola";
+
+export const logger: ConsolaInstance = createConsola({
+  formatOptions: {
+    date: false,
+  },
+});

examples-ts/parse/src/utils.ts

@@ -0,0 +1,51 @@
+import { generateText } from "ai";
+import { openai } from "@ai-sdk/openai";
+import { Document } from "llamaindex";
+import * as readline from "readline/promises";
+import figlet from "figlet";
+import pc from "picocolors";
+
+export async function renderLogo(): Promise<void> {
+  const logoText = figlet.textSync("LlamaParse Demo", {
+    font: "ANSI Shadow",
+    horizontalLayout: "default",
+    verticalLayout: "default",
+    width: 100,
+    whitespaceBreak: true,
+  });
+
+  // Add some styling with picocolors
+  const styledLogo = pc.bold(pc.magentaBright(logoText));
+
+  // Add some padding/margin
+  console.log("\n");
+  console.log(styledLogo);
+  console.log(pc.gray("─".repeat(60)));
+  console.log("\n");
+}
+
+export async function consoleInput(): Promise<string> {
+  const rl = readline.createInterface({
+    input: process.stdin,
+    output: process.stdout,
+  });
+
+  const answer = await rl.question("Path to your file: ");
+  rl.close();
+  return answer;
+}
+
+export async function generateSummary(documents: Document[]): Promise<string> {
+  let mainText: string = "";
+
+  for (const document of documents) {
+    mainText += `${document.text}\n\n---\n\n`;
+  }
+
+  const { text } = await generateText({
+    model: openai("gpt-4.1"),
+    prompt: `</chat>\n\t<text>${mainText}</text>\n\t<instructions>Could you please generate a summary of the given text?</instructions>\n</chat>`,
+  });
+
+  return text;
+}

examples-ts/parse/tsconfig.json

@@ -0,0 +1,22 @@
+{
+  "compilerOptions": {
+    "target": "ES2022",
+    "module": "ES2022",
+    "lib": ["ES2022"],
+    "outDir": "./dist",
+    "rootDir": "./src",
+    "strict": true,
+    "esModuleInterop": true,
+    "skipLibCheck": true,
+    "forceConsistentCasingInFileNames": true,
+    "declaration": true,
+    "declarationMap": true,
+    "sourceMap": true,
+    "types": ["node"],
+    "moduleResolution": "bundler",
+    "allowSyntheticDefaultImports": true,
+    "resolveJsonModule": true
+  },
+  "include": ["src/**/*"],
+  "exclude": ["node_modules", "dist"]
+}

examples/README.md

@@ -0,0 +1,9 @@
+# LlamaCloud Services Examples - Python
+
+In this folder you will find several python notebooks that contain examples regarding:
+
+- [LlamaParse](./parse/)
+- [LlamaExtract](./extract/)
+- [LlamaCloudIndex](./index/)
+
+Follow the instructions in each notebook to get started!

examples/extract/data/automotive_sector_analysis/modeling_assumptions.txt

@@ -0,0 +1,10 @@
+# Financial Modeling Assumptions
+Discount Rate: 8%
+Terminal Growth Rate: 2%
+Tax Rate: 25%
+Revenue Growth (Years 1-5): 10% per annum
+Revenue Growth (Years 6-10): 5% per annum
+Capital Expenditures as % of Revenue: 7%
+Working Capital Assumption: 3% of Revenue
+Depreciation Rate: 10% per annum
+Cost of Capital Assumption: 8%

examples/extract/data/insider_transactions/.gitignore

@@ -0,0 +1 @@
+sec_form_4_dump.json

examples/extract/extract_data_with_citations.ipynb

@@ -0,0 +1,440 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Extract Data from Financial Reports - with Citations and Reasoning\n",
+    "\n",
+    "Given complex files like financial reports, contracts, invoices etc, Llama Extract allows you to make use of an LLM to extract the information relevant to you, in a structured format.\n",
+    "\n",
+    "In this example, we'll be using [LlamaExtract](https://docs.cloud.llamaindex.ai/llamaextract/getting_started?utm_campaign=extract&utm_medium=recipe) to extract structured data from an SEC filing (specifically, the filing by Nvidia for fiscal year 2025).\n",
+    "\n",
+    "On top of simple data extraction, we'll ask our extraction agent to provide citations and reasoning for each extracted field. This allows us to:\n",
+    "- Confirm  the accuracy of the extracted field\n",
+    "- Understand the reasoning behind why the LLM extracted a given piece of information\n",
+    "- This last point allows us an opportunity to adjust the system prompt or field descriptions and improve on results where needed.\n",
+    "\n",
+    "\n",
+    "The example we go through below is also replicable within Llama Cloud as well, where you will also be able to pick between a number of pre-defined schemas, instead of building your own."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "!pip install llama-cloud-services"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Connect to Llama Cloud\n",
+    "\n",
+    "To get started, make sure you provide your [Llama Cloud](https://cloud.llamaindex.ai?utm_campaign=extract&utm_medium=recipe) API key."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Enter your Llama Cloud API Key: ··········\n"
+     ]
+    }
+   ],
+   "source": [
+    "import os\n",
+    "from getpass import getpass\n",
+    "\n",
+    "if \"LLAMA_CLOUD_API_KEY\" not in os.environ:\n",
+    "    os.environ[\"LLAMA_CLOUD_API_KEY\"] = getpass(\"Enter your Llama Cloud API Key: \")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Extract Data with Llama Extract Agent"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "No project_id provided, fetching default project.\n"
+     ]
+    }
+   ],
+   "source": [
+    "from llama_cloud_services import LlamaExtract\n",
+    "\n",
+    "# Optionally, provide your project id, if not, it will use the 'Default' project\n",
+    "llama_extract = LlamaExtract()"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Provide Your Custom Schema\n",
+    "\n",
+    "When using LlamaExtract via the API, you provide your own schema that describes what you want extracted from files and data provided to your agent. Here, we are essentially building an SEC filings extraction agent."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from pydantic import BaseModel, Field\n",
+    "from enum import Enum\n",
+    "\n",
+    "\n",
+    "class FilingType(str, Enum):\n",
+    "    ten_k = \"10 K\"\n",
+    "    ten_q = \"10-Q\"\n",
+    "    ten_ka = \"10-K/A\"\n",
+    "    ten_qa = \"10-Q/A\"\n",
+    "\n",
+    "\n",
+    "class FinancialReport(BaseModel):\n",
+    "    company_name: str = Field(description=\"The name of the company\")\n",
+    "    description: str = Field(\n",
+    "        description=\"Short description of the filing and what it contains\"\n",
+    "    )\n",
+    "    filing_type: FilingType = Field(description=\"Type of SEC filing\")\n",
+    "    filing_date: str = Field(description=\"Date when filing was submitted to SEC\")\n",
+    "    fiscal_year: int = Field(description=\"Fiscal year\")\n",
+    "    unit: str = Field(\n",
+    "        description=\"Unit of financial figures (thousands, millions, etc.)\"\n",
+    "    )\n",
+    "    revenue: int = Field(description=\"Total revenue for period\")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Set Up Citations and Reasoning\n",
+    "\n",
+    "Optionally, we can set the `ExtractConfig` to extract citations for each field the agent extracts. These cications will cite the specific pages and sections of the file from which a given field was extractedd.\n",
+    "\n",
+    "By setting `use_reasoning` to True, we als ask the agent to do an additional reasoning step, explaining why a given field was extracted."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from llama_cloud.types import ExtractConfig, ExtractMode\n",
+    "\n",
+    "config = ExtractConfig(\n",
+    "    use_reasoning=True, cite_sources=True, extraction_mode=ExtractMode.MULTIMODAL\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "/usr/local/lib/python3.11/dist-packages/llama_cloud_services/extract/extract.py:127: ExperimentalWarning: `use_reasoning` is an experimental feature. Results will be available in the `extraction_metadata` field for the extraction run.\n",
+      "  warnings.warn(\n",
+      "/usr/local/lib/python3.11/dist-packages/llama_cloud_services/extract/extract.py:133: ExperimentalWarning: `cite_sources` is an experimental feature. This may greatly increase the size of the response, and slow down the extraction. Results will be available in the `extraction_metadata` field for the extraction run.\n",
+      "  warnings.warn(\n"
+     ]
+    }
+   ],
+   "source": [
+    "agent = llama_extract.create_agent(\n",
+    "    name=\"filing-parser\", data_schema=FinancialReport, config=config\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Demo Time - Download a PDF and Extract Data with Citations"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "PDF downloaded successfully.\n"
+     ]
+    }
+   ],
+   "source": [
+    "import requests\n",
+    "\n",
+    "url = \"https://raw.githubusercontent.com/run-llama/llama_cloud_services/refs/heads/main/examples/extract/data/sec_filings/nvda_10k.pdf\"\n",
+    "\n",
+    "response = requests.get(url)\n",
+    "\n",
+    "if response.status_code == 200:\n",
+    "    with open(\"/content/nvda_10k.pdf\", \"wb\") as f:\n",
+    "        f.write(response.content)\n",
+    "    print(\"PDF downloaded successfully.\")\n",
+    "else:\n",
+    "    print(f\"Failed to download. Status code: {response.status_code}\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Uploading files: 100%|██████████| 1/1 [00:00<00:00,  1.83it/s]\n",
+      "Creating extraction jobs: 100%|██████████| 1/1 [00:00<00:00,  4.38it/s]\n",
+      "Extracting files: 100%|██████████| 1/1 [02:03<00:00, 123.40s/it]\n"
+     ]
+    }
+   ],
+   "source": [
+    "filing_info = agent.extract(\"/content/nvda_10k.pdf\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "{'company_name': 'NVIDIA Corporation',\n",
+       " 'description': \"The filing provides a detailed overview of NVIDIA's business as a full-stack computing infrastructure company, discusses various technologies including digital avatars and autonomous vehicles, outlines numerous risk factors affecting operations such as supply chain issues and geopolitical tensions, and describes employee stock purchase plans and related compliance requirements.\",\n",
+       " 'filing_type': '10 K',\n",
+       " 'filing_date': 'February 26, 2025',\n",
+       " 'fiscal_year': 2025,\n",
+       " 'unit': 'millions',\n",
+       " 'revenue': 130497}"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "filing_info.data"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### Inspect Citations and Reasoning"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "{'field_metadata': {'company_name': {'reasoning': 'VERBATIM EXTRACTION',\n",
+       "   'citation': [{'page': 1, 'matching_text': 'NVIDIA CORPORATION'},\n",
+       "    {'page': 2, 'matching_text': 'NVIDIA Corporation'},\n",
+       "    {'page': 3,\n",
+       "     'matching_text': 'All references to \"NVIDIA,\" \"we,\" \"us,\" \"our,\" or the \"Company\" mean NVIDIA Corporation and its subsidiaries.'},\n",
+       "    {'page': 35,\n",
+       "     'matching_text': 'Comparison of 5 Year Cumulative Total Return* Among NVIDIA Corporation'},\n",
+       "    {'page': 49,\n",
+       "     'matching_text': 'To the Board of Directors and Shareholders of NVIDIA Corporation'},\n",
+       "    {'page': 90, 'matching_text': 'NVIDIA Corporation'},\n",
+       "    {'page': 119,\n",
+       "     'matching_text': '*\"Company\"* means NVIDIA Corporation, a Delaware corporation.'},\n",
+       "    {'page': 126,\n",
+       "     'matching_text': 'Annual Report on Form 10-K of NVIDIA Corporation'}]},\n",
+       "  'filing_type': {'reasoning': \"VERBATIM EXTRACTION from multiple sources confirming the filing type as '10 K'.\",\n",
+       "   'citation': [{'page': 1, 'matching_text': 'FORM 10-K'},\n",
+       "    {'page': 2, 'matching_text': 'Item 16. | Form 10-K Summary'},\n",
+       "    {'page': 3,\n",
+       "     'matching_text': 'This Annual Report on Form 10-K contains forward-looking statements...'},\n",
+       "    {'page': 13, 'matching_text': 'this Annual Report on Form 10-K'},\n",
+       "    {'page': 15, 'matching_text': 'this Annual Report on Form 10-K'},\n",
+       "    {'page': 32,\n",
+       "     'matching_text': 'Annual Report on Form 10-K, which information is hereby incorporated by reference.'},\n",
+       "    {'page': 36, 'matching_text': 'this Annual Report on Form 10-K'},\n",
+       "    {'page': 43,\n",
+       "     'matching_text': 'Annual Report on Form 10-K for additional information'},\n",
+       "    {'page': 45, 'matching_text': 'Annual Report on Form 10-K'},\n",
+       "    {'page': 46, 'matching_text': 'this Annual Report on Form 10-K'},\n",
+       "    {'page': 62, 'matching_text': 'Annual Report on Form 10-K'},\n",
+       "    {'page': 83,\n",
+       "     'matching_text': 'Restated Certificate of Incorporation | 10-K'},\n",
+       "    {'page': 84, 'matching_text': 'Item 16. Form 10-K Summary'},\n",
+       "    {'page': 126, 'matching_text': 'which appears in this Form 10-K'},\n",
+       "    {'page': 127, 'matching_text': 'Annual Report on Form 10-K'},\n",
+       "    {'page': 128, 'matching_text': 'Annual Report on Form 10-K'},\n",
+       "    {'page': 129, 'matching_text': \"The Company's Annual Report on Form 10-K\"},\n",
+       "    {'page': 130,\n",
+       "     'matching_text': \"The Company's Annual Report on Form 10-K for the year ended January 26, 2025\"}]},\n",
+       "  'fiscal_year': {'reasoning': 'The fiscal year ended January 26, 2025, indicates the fiscal year is 2025. Additionally, multiple references throughout the text confirm the fiscal year 2025 in various contexts.',\n",
+       "   'citation': [{'page': 1,\n",
+       "     'matching_text': 'For the fiscal year ended January 26, 2025'},\n",
+       "    {'page': 6,\n",
+       "     'matching_text': 'In fiscal year 2025, we launched the NVIDIA Blackwell architecture'},\n",
+       "    {'page': 12, 'matching_text': 'fiscal year 2025'},\n",
+       "    {'page': 17,\n",
+       "     'matching_text': 'our gross margins in the second quarter of fiscal year 2025 were negatively impacted'},\n",
+       "    {'page': 20,\n",
+       "     'matching_text': 'we generated 53% of our revenue in fiscal year 2025 from sales outside the United States.'},\n",
+       "    {'page': 23,\n",
+       "     'matching_text': 'For fiscal year 2025, an indirect customer which primarily purchases our products through system integrators...'},\n",
+       "    {'page': 33,\n",
+       "     'matching_text': 'In fiscal year 2025, we repurchased 310 million shares of our common stock for $34.0 billion.'},\n",
+       "    {'page': 37,\n",
+       "     'matching_text': 'Our Data Center revenue in China grew in fiscal year 2025.'},\n",
+       "    {'page': 44,\n",
+       "     'matching_text': 'Cash provided by operating activities increased in fiscal year 2025 compared to fiscal year 2024'},\n",
+       "    {'page': 57,\n",
+       "     'matching_text': 'Fiscal years 2025, 2024 and 2023 were all 52-week years.'},\n",
+       "    {'page': 65,\n",
+       "     'matching_text': 'Beginning in the second quarter of fiscal year 2025'},\n",
+       "    {'page': 69, 'matching_text': 'In the fourth quarter of fiscal year 2025'},\n",
+       "    {'page': 78,\n",
+       "     'matching_text': 'Depreciation and amortization expense attributable to our Compute and Networking segment for fiscal years 2025'},\n",
+       "    {'page': 129, 'matching_text': 'for the year ended January 26, 2025'}]},\n",
+       "  'description': {'reasoning': 'The extracted data combines multiple descriptions from the source text, ensuring no duplication while maintaining the order and context of the information. Each section of the filing is summarized to reflect the key points without losing the essence of the original text.',\n",
+       "   'citation': [{'page': 4,\n",
+       "     'matching_text': 'NVIDIA is now a full-stack computing infrastructure company with data-center-scale offerings that are reshaping industry.'},\n",
+       "    {'page': 8,\n",
+       "     'matching_text': 'a suite of technologies that help developers bring digital avatars to life with generative Al...autonomous vehicles, or AV, and electric vehicles, or EV, is revolutionizing the transportation industry...Our worldwide sales and marketing strategy is key to achieving our objective of providing markets with our high-performance and efficient computing platforms and software.'},\n",
+       "    {'page': 14, 'matching_text': 'Risk Factors Summary'},\n",
+       "    {'page': 16,\n",
+       "     'matching_text': 'Risks Related to Demand, Supply, and Manufacturing\\n\\nLong manufacturing lead times and uncertain supply and component availability...'},\n",
+       "    {'page': 18,\n",
+       "     'matching_text': 'cryptocurrency mining, on demand for our products. Volatility in the cryptocurrency market, including new compute technologies...'},\n",
+       "    {'page': 21,\n",
+       "     'matching_text': 'supply-chain attacks or other business disruptions. We cannot guarantee that third parties and infrastructure in our supply chain...'},\n",
+       "    {'page': 22,\n",
+       "     'matching_text': 'We are monitoring the impact of the geopolitical conflict in and around Israel on our operations... Climate change may have a long-term impact on our business.'},\n",
+       "    {'page': 25,\n",
+       "     'matching_text': 'We are subject to complex laws, rules, regulations, and political and other actions, including restrictions on the export of our products, which may adversely impact our business.'},\n",
+       "    {'page': 28,\n",
+       "     'matching_text': 'Our competitive position has been harmed by the existing export controls, and our competitive position and future results may be further harmed'},\n",
+       "    {'page': 29,\n",
+       "     'matching_text': 'restrictions imposed by the Chinese government on the duration of gaming activities and access to games may adversely affect our Gaming revenue'},\n",
+       "    {'page': 29,\n",
+       "     'matching_text': 'our business depends on our ability to receive consistent and reliable supply from our overseas partners, especially in Taiwan and South Korea'},\n",
+       "    {'page': 29,\n",
+       "     'matching_text': 'Increased scrutiny from shareholders, regulators and others regarding our corporate sustainability practices could result in additional costs'},\n",
+       "    {'page': 29,\n",
+       "     'matching_text': 'Concerns relating to the responsible use of new and evolving technologies, such as Al, in our products and services may result in reputational or financial harm'},\n",
+       "    {'page': 31,\n",
+       "     'matching_text': 'Data protection laws around the world are quickly changing and may be interpreted and applied in an increasingly stringent fashion...'}]},\n",
+       "  'filing_date': {'reasoning': 'The filing date is consistently mentioned as February 26, 2025 across multiple entries, making it the most reliable date for the filing.',\n",
+       "   'citation': [{'page': 51, 'matching_text': 'February 26, 2025'},\n",
+       "    {'page': 86, 'matching_text': 'on February 26, 2025.'},\n",
+       "    {'page': 87, 'matching_text': 'February 26, 2025'},\n",
+       "    {'page': 126, 'matching_text': 'our report dated February 26, 2025'},\n",
+       "    {'page': 127, 'matching_text': 'Date: February 26, 2025'},\n",
+       "    {'page': 128, 'matching_text': 'Date: February 26, 2025'},\n",
+       "    {'page': 129, 'matching_text': 'Date: February 26, 2025'},\n",
+       "    {'page': 130, 'matching_text': 'Date: February 26, 2025'}]},\n",
+       "  'unit': {'reasoning': \"The unit of financial figures is explicitly mentioned multiple times in the text as 'millions', including in table headers and notes. This is confirmed by various citations from pages 38, 42, 43, 52, 53, 54, 56, 65, 71, 72, 73, 75, 77, 79, 80, and 82.\",\n",
+       "   'citation': [{'page': 38,\n",
+       "     'matching_text': '($ in millions, except per share data)'},\n",
+       "    {'page': 42, 'matching_text': '($ in millions)'},\n",
+       "    {'page': 43, 'matching_text': '($ in millions)'},\n",
+       "    {'page': 52, 'matching_text': '(In millions, except per share data)'},\n",
+       "    {'page': 53,\n",
+       "     'matching_text': 'Consolidated Statements of Comprehensive Income (In millions)'},\n",
+       "    {'page': 54,\n",
+       "     'matching_text': 'Consolidated Balance Sheets (In millions, except par value)'},\n",
+       "    {'page': 55, 'matching_text': '(In millions, except per share data)'},\n",
+       "    {'page': 56,\n",
+       "     'matching_text': 'Consolidated Statements of Cash Flows (In millions)'},\n",
+       "    {'page': 65,\n",
+       "     'matching_text': 'Year Ended<br/>Jan 26, 2025<br/>(In millions, except per share data)'},\n",
+       "    {'page': 71, 'matching_text': '(In millions) | (In millions)'},\n",
+       "    {'page': 72, 'matching_text': '(In millions)'}]},\n",
+       "  'revenue': {'reasoning': 'The total revenue for fiscal year 2025 is extracted from multiple sources within the text, all confirming the same figure of $130,497 million. The revenue recognized for fiscal year 2025 is also noted as $4,607 million, which is a separate figure. However, the primary focus is on the total revenue figure, which is consistently cited.',\n",
+       "   'citation': [{'page': 38,\n",
+       "     'matching_text': 'Revenue for fiscal year 2025 was $130.5 billion'},\n",
+       "    {'page': 41,\n",
+       "     'matching_text': 'Total                | $ 130,497    | $ | 60,922'},\n",
+       "    {'page': 52, 'matching_text': 'Revenue | $ 130,497'},\n",
+       "    {'page': 78,\n",
+       "     'matching_text': 'Revenue | $ 116,193 | $ 14,304 | $ - | $ 130,497'},\n",
+       "    {'page': 79, 'matching_text': 'Total revenue | $ 130,497'},\n",
+       "    {'page': 80, 'matching_text': 'Total revenue              | $ 130,497'}]}},\n",
+       " 'usage': {'num_pages_extracted': 130,\n",
+       "  'num_document_tokens': 105932,\n",
+       "  'num_output_tokens': 31306}}"
+      ]
+     },
+     "execution_count": null,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "filing_info.extraction_metadata"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## What's Next?\n",
+    "\n",
+    "In this example, we built an Extraction Agent that is capable of citing it's sources from the document it's extracting data from, and reasoning about its reponse. To further customize and improve on the results, you can also try to customize the `system_prompt` in the `ExtractConfig`.\n",
+    "\n",
+    "#### Learn More\n",
+    "\n",
+    "- [LlamaExtract Documentation](https://docs.cloud.llamaindex.ai/llamaextract/getting_started)\n",
+    "- [Example Notebooks](https://github.com/run-llama/llama_cloud_services/tree/main/examples/extract)"
+   ]
+  }
+ ],
+ "metadata": {
+  "colab": {
+   "provenance": []
+  },
+  "kernelspec": {
+   "display_name": "Python 3",
+   "name": "python3"
+  },
+  "language_info": {
+   "name": "python"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 0
+}

examples/extract/solar_panel_e2e_comparison.ipynb

@@ -0,0 +1,450 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "id": "00f6713b-2a32-4f8f-80e5-9a7d9b6e3b90",
+   "metadata": {},
+   "source": [
+    "# Solar Panel Datasheet Comparison Workflow\n",
+    "\n",
+    "<a href=\"https://colab.research.google.com/github/run-llama/llama_cloud_services/blob/main/examples/extract/solar_panel_e2e_comparison.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>\n",
+    "\n",
+    "\n",
+    "This notebook demonstrates an end‑to‑end agentic workflow using LlamaExtract and the LlamaIndex event‑driven workflow framework. In this workflow, we:\n",
+    "\n",
+    "1. **Extract** structured technical specifications from a solar panel datasheet (e.g. a PDF downloaded from a vendor).\n",
+    "2. **Load** design requirements (provided as a text blob) for a lab‑grade solar panel.\n",
+    "3. **Generate** a detailed comparison report by triggering an event that injects both the extracted data and the requirements into an LLM prompt.\n",
+    "\n",
+    "The workflow is designed for renewable energy engineers who need to quickly validate that a solar panel meets specific design criteria.\n",
+    "\n",
+    "The following notebook uses the event‑driven syntax (with custom events, steps, and a workflow class) adapted from the technical datasheet and contract review examples."
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "36d8e34e-ed98-46ac-b744-1642f6e253d5",
+   "metadata": {},
+   "source": [
+    "## Setup and Load Data\n",
+    "\n",
+    "We download the [Honey M TSM-DE08M.08(II) datasheet](https://static.trinasolar.com/sites/default/files/EU_Datasheet_HoneyM_DE08M.08%28II%29_2021_A.pdf) as a PDF.\n",
+    "\n",
+    "**NOTE**: The design requirements are already stored in `data/solar_panel_e2e_comparison/design_reqs.txt`."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "1de7b1b3-c285-492c-8b2e-b37974b4fc63",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "--2025-04-01 14:47:56--  https://static.trinasolar.com/sites/default/files/EU_Datasheet_HoneyM_DE08M.08%28II%29_2021_A.pdf\n",
+      "Resolving static.trinasolar.com (static.trinasolar.com)... 47.246.23.232, 47.246.23.234, 47.246.23.227, ...\n",
+      "Connecting to static.trinasolar.com (static.trinasolar.com)|47.246.23.232|:443... connected.\n",
+      "WARNING: cannot verify static.trinasolar.com's certificate, issued by ‘CN=DigiCert Global G2 TLS RSA SHA256 2020 CA1,O=DigiCert Inc,C=US’:\n",
+      "  Unable to locally verify the issuer's authority.\n",
+      "HTTP request sent, awaiting response... 200 OK\n",
+      "Length: 1888183 (1.8M) [application/pdf]\n",
+      "Saving to: ‘data/solar_panel_e2e_comparison/datasheet.pdf’\n",
+      "\n",
+      "data/solar_panel_e2 100%[===================>]   1.80M  7.47MB/s    in 0.2s    \n",
+      "\n",
+      "2025-04-01 14:47:56 (7.47 MB/s) - ‘data/solar_panel_e2e_comparison/datasheet.pdf’ saved [1888183/1888183]\n",
+      "\n"
+     ]
+    }
+   ],
+   "source": [
+    "!wget https://static.trinasolar.com/sites/default/files/EU_Datasheet_HoneyM_DE08M.08%28II%29_2021_A.pdf -O data/solar_panel_e2e_comparison/datasheet.pdf --no-check-certificate"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "89d2f4c9-f785-424d-a409-3381796c457c",
+   "metadata": {},
+   "source": [
+    "## Define the Structured Extraction Schema\n",
+    "\n",
+    "We define a new, rich schema called `SolarPanelSchema` to capture key technical details from the datasheet. This schema includes:\n",
+    "\n",
+    "- **PowerRange:** Structured as minimum and maximum power output (in Watts).\n",
+    "- **SolarPanelSpec:** Includes module name, power output range, maximum efficiency, certifications, and a mapping of page citations.\n",
+    "\n",
+    "This schema replaces the earlier LM317 schema and will be used when creating our extraction agent."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "bfb40d48-36e0-4b1c-97a1-32a1704c582b",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from pydantic import BaseModel, Field\n",
+    "from typing import List\n",
+    "\n",
+    "\n",
+    "class PowerRange(BaseModel):\n",
+    "    min_power: float = Field(..., description=\"Minimum power output in Watts\")\n",
+    "    max_power: float = Field(..., description=\"Maximum power output in Watts\")\n",
+    "    unit: str = Field(\"W\", description=\"Power unit\")\n",
+    "\n",
+    "\n",
+    "class SolarPanelSpec(BaseModel):\n",
+    "    module_name: str = Field(..., description=\"Name or model of the solar panel module\")\n",
+    "    power_output: PowerRange = Field(..., description=\"Power output range\")\n",
+    "    maximum_efficiency: float = Field(\n",
+    "        ..., description=\"Maximum module efficiency in percentage\"\n",
+    "    )\n",
+    "    temperature_coefficient: float = Field(\n",
+    "        ..., description=\"Temperature coefficient in %/°C\"\n",
+    "    )\n",
+    "    certifications: List[str] = Field([], description=\"List of certifications\")\n",
+    "    page_citations: dict = Field(\n",
+    "        ..., description=\"Mapping of each extracted field to its page numbers\"\n",
+    "    )\n",
+    "\n",
+    "\n",
+    "class SolarPanelSchema(BaseModel):\n",
+    "    specs: List[SolarPanelSpec] = Field(\n",
+    "        ..., description=\"List of extracted solar panel specifications\"\n",
+    "    )"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "19dc309e-7cec-43c1-8f6c-72e14df58f8f",
+   "metadata": {},
+   "source": [
+    "## Initialize Extraction Agent\n",
+    "\n",
+    "Here we initialize our extraction agent that will be responsible for extracting the schema from the solar panel datasheet."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "c9d9f4a2-2e14-493d-8a7e-d01159d38b8f",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from dotenv import load_dotenv\n",
+    "from llama_cloud_services import LlamaExtract\n",
+    "from llama_cloud.core.api_error import ApiError\n",
+    "from llama_cloud import ExtractConfig\n",
+    "\n",
+    "# Initialize the LlamaExtract client\n",
+    "llama_extract = LlamaExtract(\n",
+    "    project_id=\"2fef999e-1073-40e6-aeb3-1f3c0e64d99b\",\n",
+    "    organization_id=\"43b88c8f-e488-46f6-9013-698e3d2e374a\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "ec0eb2a7-6e02-45da-a6af-227e2f7c81f2",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "try:\n",
+    "    existing_agent = llama_extract.get_agent(name=\"solar-panel-datasheet\")\n",
+    "    if existing_agent:\n",
+    "        llama_extract.delete_agent(existing_agent.id)\n",
+    "except ApiError as e:\n",
+    "    if e.status_code == 404:\n",
+    "        pass\n",
+    "    else:\n",
+    "        raise\n",
+    "\n",
+    "extract_config = ExtractConfig(\n",
+    "    extraction_mode=\"BALANCED\",\n",
+    ")\n",
+    "\n",
+    "agent = llama_extract.create_agent(\n",
+    "    name=\"solar-panel-datasheet\", data_schema=SolarPanelSchema, config=extract_config\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "b4d7bb60-0456-4a2d-8d48-14f9bb3e71d2",
+   "metadata": {},
+   "source": [
+    "## Workflow Overview\n",
+    "\n",
+    "The workflow consists of four main steps:\n",
+    "\n",
+    "1. **parse_datasheet:** Reads the solar panel datasheet (PDF) and converts its content into text (with page citations).\n",
+    "2. **load_requirements:** Loads the design requirements (as a text blob) that will be injected into the prompt.\n",
+    "3. **generate_comparison_report:** Constructs a prompt using the extracted datasheet content and design requirements and triggers the LLM to generate a comparison report.\n",
+    "4. **output_result:** Logs and returns the final report as the workflow’s result.\n",
+    "\n",
+    "Each step is implemented as an asynchronous function decorated with `@step`, and the workflow is built by subclassing `Workflow`."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "7c482e3a-66b4-4e1b-8d2d-9a9c6b3967f3",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from llama_index.core.workflow import (\n",
+    "    Event,\n",
+    "    StartEvent,\n",
+    "    StopEvent,\n",
+    "    Context,\n",
+    "    Workflow,\n",
+    "    step,\n",
+    ")\n",
+    "from llama_index.llms.openai import OpenAI\n",
+    "from llama_index.core.prompts import ChatPromptTemplate\n",
+    "from llama_cloud_services import LlamaExtract\n",
+    "from llama_cloud.core.api_error import ApiError\n",
+    "from pydantic import BaseModel, Field\n",
+    "from typing import List\n",
+    "\n",
+    "\n",
+    "# Define output schema for the comparison report (for reference)\n",
+    "class ComparisonReportOutput(BaseModel):\n",
+    "    component_name: str = Field(\n",
+    "        ..., description=\"The name of the component being evaluated.\"\n",
+    "    )\n",
+    "    meets_requirements: bool = Field(\n",
+    "        ...,\n",
+    "        description=\"Overall indicator of whether the component meets the design criteria.\",\n",
+    "    )\n",
+    "    summary: str = Field(..., description=\"A brief summary of the evaluation results.\")\n",
+    "    details: dict = Field(\n",
+    "        ..., description=\"Detailed comparisons for each key parameter.\"\n",
+    "    )\n",
+    "\n",
+    "\n",
+    "# Define custom events\n",
+    "\n",
+    "\n",
+    "class DatasheetParseEvent(Event):\n",
+    "    datasheet_content: dict\n",
+    "\n",
+    "\n",
+    "class RequirementsLoadEvent(Event):\n",
+    "    requirements_text: str\n",
+    "\n",
+    "\n",
+    "class ComparisonReportEvent(Event):\n",
+    "    report: ComparisonReportOutput\n",
+    "\n",
+    "\n",
+    "class LogEvent(Event):\n",
+    "    msg: str\n",
+    "    delta: bool = False\n",
+    "\n",
+    "\n",
+    "# For our demonstration, we assume that LlamaExtract is used to parse the datasheet into text.\n",
+    "# We'll also use OpenAI (via LlamaIndex) as our LLM for generating the report.\n",
+    "\n",
+    "llm = OpenAI(model=\"gpt-4o\")  # or your preferred model"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "67a0c391-c7f5-4b93-8d6b-9e31b2d7a817",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "class SolarPanelComparisonWorkflow(Workflow):\n",
+    "    \"\"\"\n",
+    "    Workflow to extract data from a solar panel datasheet and generate a comparison report\n",
+    "    against provided design requirements.\n",
+    "    \"\"\"\n",
+    "\n",
+    "    def __init__(self, agent: LlamaExtract, requirements_path: str, **kwargs):\n",
+    "        super().__init__(**kwargs)\n",
+    "        self.agent = agent\n",
+    "        # Load design requirements from file as a text blob\n",
+    "        with open(requirements_path, \"r\") as f:\n",
+    "            self.requirements_text = f.read()\n",
+    "\n",
+    "    @step\n",
+    "    async def parse_datasheet(\n",
+    "        self, ctx: Context, ev: StartEvent\n",
+    "    ) -> DatasheetParseEvent:\n",
+    "        # datasheet_path is provided in the StartEvent\n",
+    "        datasheet_path = (\n",
+    "            ev.datasheet_path\n",
+    "        )  # e.g., \"./data/solar_panel_comparison/datasheet.pdf\"\n",
+    "        extraction_result = await self.agent.aextract(datasheet_path)\n",
+    "        datasheet_dict = (\n",
+    "            extraction_result.data\n",
+    "        )  # assumed to be a string with page citations\n",
+    "        await ctx.set(\"datasheet_content\", datasheet_dict)\n",
+    "        ctx.write_event_to_stream(LogEvent(msg=\"Datasheet parsed successfully.\"))\n",
+    "        return DatasheetParseEvent(datasheet_content=datasheet_dict)\n",
+    "\n",
+    "    @step\n",
+    "    async def load_requirements(\n",
+    "        self, ctx: Context, ev: DatasheetParseEvent\n",
+    "    ) -> RequirementsLoadEvent:\n",
+    "        # Use the pre-loaded requirements text from __init__\n",
+    "        req_text = self.requirements_text\n",
+    "        ctx.write_event_to_stream(LogEvent(msg=\"Design requirements loaded.\"))\n",
+    "        return RequirementsLoadEvent(requirements_text=req_text)\n",
+    "\n",
+    "    @step\n",
+    "    async def generate_comparison_report(\n",
+    "        self, ctx: Context, ev: RequirementsLoadEvent\n",
+    "    ) -> StopEvent:\n",
+    "        # Build a prompt that injects both the extracted datasheet content and the design requirements\n",
+    "        datasheet_content = await ctx.get(\"datasheet_content\")\n",
+    "        prompt_str = \"\"\"\n",
+    "You are an expert renewable energy engineer.\n",
+    "\n",
+    "Compare the following solar panel datasheet information with the design requirements.\n",
+    "\n",
+    "Design Requirements:\n",
+    "{requirements_text}\n",
+    "\n",
+    "Extracted Datasheet Information:\n",
+    "{datasheet_content}\n",
+    "\n",
+    "Generate a detailed comparison report in JSON format with the following schema:\n",
+    "  - component_name: string\n",
+    "  - meets_requirements: boolean\n",
+    "  - summary: string\n",
+    "  - details: dictionary of comparisons for each parameter\n",
+    "\n",
+    "For each parameter (Maximum Power, Open-Circuit Voltage, Short-Circuit Current, Efficiency, Temperature Coefficient),\n",
+    "indicate PASS or FAIL and provide brief explanations and recommendations.\n",
+    "\"\"\"\n",
+    "\n",
+    "        # extract from contract\n",
+    "        prompt = ChatPromptTemplate.from_messages([(\"user\", prompt_str)])\n",
+    "\n",
+    "        # Call the LLM to generate the report using the prompt\n",
+    "        report_output = await llm.astructured_predict(\n",
+    "            ComparisonReportOutput,\n",
+    "            prompt,\n",
+    "            requirements_text=ev.requirements_text,\n",
+    "            datasheet_content=str(datasheet_content),\n",
+    "        )\n",
+    "        ctx.write_event_to_stream(LogEvent(msg=\"Comparison report generated.\"))\n",
+    "        return StopEvent(\n",
+    "            result={\"report\": report_output, \"datasheet_content\": datasheet_content}\n",
+    "        )"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "id": "d205f532-1a11-4a48-b5a8-87a7f85e9ce7",
+   "metadata": {},
+   "source": [
+    "## Running the Workflow\n",
+    "\n",
+    "Below, we instantiate and run the workflow. We inject the design requirements as a text blob (no custom code to load) and pass the path to the solar panel datasheet (the HoneyM datasheet from Trina).\n",
+    "\n",
+    "The design requirements are:\n",
+    "\n",
+    "```\n",
+    "Solar Panel Design Requirements:\n",
+    "- Power Output Range: ≥ 350 W\n",
+    "- Maximum Efficiency: ≥ 18%\n",
+    "- Certifications: Must include IEC61215 and UL1703\n",
+    "```\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "6b24fa61-a2f5-4ebb-84eb-1c9b48683b1b",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import nest_asyncio\n",
+    "\n",
+    "nest_asyncio.apply()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "be3ebad5-1f70-4671-a2ec-17bf9e4d788f",
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Path to design requirements file (e.g., a text file with design criteria for solar panels)\n",
+    "requirements_path = \"./data/solar_panel_e2e_comparison/design_reqs.txt\"\n",
+    "\n",
+    "# Instantiate the workflow\n",
+    "workflow = SolarPanelComparisonWorkflow(\n",
+    "    agent=agent, requirements_path=requirements_path, verbose=True, timeout=120\n",
+    ")\n",
+    "\n",
+    "# Run the workflow; pass the datasheet path in the StartEvent\n",
+    "result = await workflow.run(\n",
+    "    datasheet_path=\"./data/solar_panel_e2e_comparison/datasheet.pdf\"\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "e1e61f1e-8701-4acc-8f99-cc89d8aae535",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "\n",
+      "********Final Comparison Report:********\n",
+      "\n",
+      "{\n",
+      "    \"component_name\": \"TSM-DE08M.08(II)\",\n",
+      "    \"meets_requirements\": true,\n",
+      "    \"summary\": \"The solar panel TSM-DE08M.08(II) meets all the design requirements, making it a suitable choice for the intended application.\",\n",
+      "    \"details\": {\n",
+      "        \"Maximum Power Output\": \"PASS - The panel's power output ranges from 360 W to 385 W, exceeding the minimum requirement of 350 W.\",\n",
+      "        \"Open-Circuit Voltage\": \"PASS - The datasheet does not specify Voc, but the panel meets other critical requirements. Verification of Voc is recommended.\",\n",
+      "        \"Short-Circuit Current\": \"PASS - The datasheet does not specify Isc, but the panel meets other critical requirements. Verification of Isc is recommended.\",\n",
+      "        \"Efficiency\": \"PASS - The panel's efficiency is 21.0%, which is above the required 18%.\",\n",
+      "        \"Temperature Coefficient\": \"PASS - The temperature coefficient is -0.34%/°C, which is better than the maximum allowable -0.5%/°C.\"\n",
+      "    }\n",
+      "}\n"
+     ]
+    }
+   ],
+   "source": [
+    "print(\"\\n********Final Comparison Report:********\\n\")\n",
+    "print(result[\"report\"].model_dump_json(indent=4))\n",
+    "# print(\"\\n********Datasheet Content:********\\n\", result[\"datasheet_content\"])"
+   ]
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "llama_parse",
+   "language": "python",
+   "name": "llama_parse"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 5
+}

examples/parse/advanced_rag/dynamic_section_retrieval.ipynb

@@ -7,7 +7,7 @@
    "source": [
     "# Dynamic Section Retrieval with LlamaParse\n",
     "\n",
-    "<a href=\"https://colab.research.google.com/github/run-llama/llama_cloud_services-demo/blob/main/examples/parse/advanced_rag/dynamic_section_retrieval.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>\n",
+    "<a href=\"https://colab.research.google.com/github/run-llama/llama_cloud_services/blob/main/examples/parse/advanced_rag/dynamic_section_retrieval.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>\n",
     "\n",
     "This notebook showcases a concept called \"dynamic section retrieval\".\n",
     "\n",
@@ -19,7 +19,12 @@
     "\n",
     "![](dynamic_section_retrieval_img.png)\n",
     "\n",
-    "This helps provide a solution to the common chunking problem of retrieving chunks that are only subsets of the entire section you're meant to retrieve."
+    "This helps provide a solution to the common chunking problem of retrieving chunks that are only subsets of the entire section you're meant to retrieve.\n",
+    "\n",
+    "Status:\n",
+    "| Last Executed | Version | State      |\n",
+    "|---------------|---------|------------|\n",
+    "| Aug-19-2025   | 0.6.61  | Maintained |"
    ]
   },
   {
@@ -32,18 +37,6 @@
     "Install core packages and download relevant files. Here we load some popular ICLR 2024 papers."
    ]
   },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "71bd0714-324f-48b3-8a93-72c6c3a10b53",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "import nest_asyncio\n",
-    "\n",
-    "nest_asyncio.apply()"
-   ]
-  },
   {
    "cell_type": "code",
    "execution_count": null,
@@ -51,8 +44,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "!pip install llama-index\n",
-    "!pip install llama-index-core\n",
+    "!pip install \"llama-index>=0.13.0<0.14.0\" \"llama-index-vector-stores-chroma>=0.5.1<0.6.0\"\n",
     "!pip install llama-cloud-services"
    ]
   },
@@ -101,48 +93,7 @@
    "execution_count": null,
    "id": "80137d15-f22b-47eb-adce-ac295ced7e71",
    "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "mkdir: iclr_docs: File exists\n",
-      "--2024-11-10 16:18:56--  https://openreview.net/pdf?id=VTF8yNQM66\n",
-      "Resolving openreview.net (openreview.net)... 35.184.86.251\n",
-      "Connecting to openreview.net (openreview.net)|35.184.86.251|:443... connected.\n",
-      "HTTP request sent, awaiting response... 200 OK\n",
-      "Length: 2680380 (2.6M) [application/pdf]\n",
-      "Saving to: ‘iclr_docs/swebench.pdf’\n",
-      "\n",
-      "iclr_docs/swebench. 100%[===================>]   2.56M  7.22MB/s    in 0.4s    \n",
-      "\n",
-      "2024-11-10 16:18:57 (7.22 MB/s) - ‘iclr_docs/swebench.pdf’ saved [2680380/2680380]\n",
-      "\n",
-      "--2024-11-10 16:18:57--  https://openreview.net/pdf?id=hSyW5go0v8\n",
-      "Resolving openreview.net (openreview.net)... 35.184.86.251\n",
-      "Connecting to openreview.net (openreview.net)|35.184.86.251|:443... connected.\n",
-      "HTTP request sent, awaiting response... 200 OK\n",
-      "Length: 1244749 (1.2M) [application/pdf]\n",
-      "Saving to: ‘iclr_docs/selfrag.pdf’\n",
-      "\n",
-      "iclr_docs/selfrag.p 100%[===================>]   1.19M  4.21MB/s    in 0.3s    \n",
-      "\n",
-      "2024-11-10 16:18:58 (4.21 MB/s) - ‘iclr_docs/selfrag.pdf’ saved [1244749/1244749]\n",
-      "\n",
-      "--2024-11-10 16:18:58--  https://openreview.net/pdf?id=c5pwL0Soay\n",
-      "Resolving openreview.net (openreview.net)... 35.184.86.251\n",
-      "Connecting to openreview.net (openreview.net)|35.184.86.251|:443... connected.\n",
-      "HTTP request sent, awaiting response... 200 OK\n",
-      "Length: 4775879 (4.6M) [application/pdf]\n",
-      "Saving to: ‘iclr_docs/metra.pdf’\n",
-      "\n",
-      "iclr_docs/metra.pdf 100%[===================>]   4.55M  4.06MB/s    in 1.1s    \n",
-      "\n",
-      "2024-11-10 16:19:00 (4.06 MB/s) - ‘iclr_docs/metra.pdf’ saved [4775879/4775879]\n",
-      "\n"
-     ]
-    }
-   ],
+   "outputs": [],
    "source": [
     "!mkdir \"{data_dir}\"\n",
     "for url, paper in zip(urls, papers):\n",
@@ -168,8 +119,8 @@
     "from llama_index.llms.openai import OpenAI\n",
     "from llama_index.embeddings.openai import OpenAIEmbedding\n",
     "\n",
-    "embed_model = OpenAIEmbedding(model=\"text-embedding-3-large\")\n",
-    "llm = OpenAI(model=\"gpt-4o\")\n",
+    "embed_model = OpenAIEmbedding(model=\"text-embedding-3-large\", api_key=\"sk-...\")\n",
+    "llm = OpenAI(model=\"gpt-5-mini\", api_key=\"sk-...\")\n",
     "\n",
     "Settings.embed_model = embed_model\n",
     "Settings.llm = llm"
@@ -192,7 +143,15 @@
    "source": [
     "from llama_cloud_services import LlamaParse\n",
     "\n",
-    "parser = LlamaParse(result_type=\"markdown\")"
+    "parser = LlamaParse(\n",
+    "    parse_mode=\"parse_page_with_agent\",\n",
+    "    model=\"openai-gpt-4-1-mini\",\n",
+    "    high_res_ocr=True,\n",
+    "    adaptive_long_table=True,\n",
+    "    outlined_table_extraction=True,\n",
+    "    output_tables_as_HTML=True,\n",
+    "    api_key=\"llx-...\",\n",
+    ")"
    ]
   },
   {
@@ -201,30 +160,56 @@
    "id": "f9d6f0e8-323e-4786-a4a8-e393441ecd61",
    "metadata": {},
    "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Getting job results:   0%|          | 0/3 [00:00<?, ?it/s]"
+     ]
+    },
     {
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "Started parsing the file under job_id 827f328d-b72e-4b70-8b4b-47dbba859d69\n",
-      "Started parsing the file under job_id d3104cd5-731e-4def-bdbc-889e8731989c\n",
-      "Started parsing the file under job_id 6046274e-e522-46af-9185-3c036e9c3ad6\n"
+      "Started parsing the file under job_id d8f0df2d-5b55-4e4f-bbe9-81cf4b8a4782\n",
+      "Started parsing the file under job_id 6aef247f-f548-43f5-9ddb-cf8ba8373130\n",
+      "Started parsing the file under job_id 5c1c4baf-fa43-4ed4-b671-16c45f99461c\n",
+      "..."
+     ]
+    },
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Getting job results:  67%|██████▋   | 2/3 [01:40<00:46, 46.97s/it]"
+     ]
+    },
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "....."
+     ]
+    },
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "Getting job results: 100%|██████████| 3/3 [05:49<00:00, 116.59s/it]\n"
      ]
     }
    ],
    "source": [
     "from pathlib import Path\n",
     "\n",
-    "paper_dicts = {}\n",
-    "\n",
+    "paths_to_parse = []\n",
     "for paper_path in papers:\n",
     "    paper_base = Path(paper_path).stem\n",
     "    full_paper_path = str(Path(data_dir) / paper_path)\n",
-    "    md_json_objs = parser.get_json_result(full_paper_path)\n",
-    "    json_dicts = md_json_objs[0][\"pages\"]\n",
-    "    paper_dicts[paper_path] = {\n",
-    "        \"paper_path\": full_paper_path,\n",
-    "        \"json_dicts\": json_dicts,\n",
-    "    }"
+    "    paths_to_parse.append(full_paper_path)\n",
+    "\n",
+    "\n",
+    "results = await parser.aparse(paths_to_parse)"
    ]
   },
   {
@@ -234,44 +219,7 @@
    "source": [
     "#### Get Text Nodes\n",
     "\n",
-    "Convert the dictionary above into TextNode objects that we can put into a vector store."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "18c24174-05ce-417f-8dd2-79c3f375db03",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "from llama_index.core.schema import TextNode\n",
-    "from typing import Optional"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "8e331dfe-a627-4e23-8c57-70ab1d9342e4",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "# NOTE: these are utility functions to sort the dumped images by the page number\n",
-    "# (they are formatted like \"{uuid}-{page_num}.jpg\"\n",
-    "import re\n",
-    "\n",
-    "\n",
-    "def get_page_number(file_name):\n",
-    "    match = re.search(r\"-page-(\\d+)\\.jpg$\", str(file_name))\n",
-    "    if match:\n",
-    "        return int(match.group(1))\n",
-    "    return 0\n",
-    "\n",
-    "\n",
-    "def _get_sorted_image_files(image_dir):\n",
-    "    \"\"\"Get image files sorted by page.\"\"\"\n",
-    "    raw_files = [f for f in list(Path(image_dir).iterdir()) if f.is_file()]\n",
-    "    sorted_files = sorted(raw_files, key=get_page_number)\n",
-    "    return sorted_files"
+    "Using each result object, we can create a list of text nodes with metadata attached."
    ]
   },
   {
@@ -281,21 +229,20 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "from copy import deepcopy\n",
-    "from pathlib import Path\n",
+    "from llama_index.core.schema import TextNode\n",
     "\n",
     "\n",
     "# attach image metadata to the text nodes\n",
-    "def get_text_nodes(json_dicts, paper_path):\n",
+    "def get_text_nodes(result):\n",
     "    \"\"\"Split docs into nodes, by separator.\"\"\"\n",
     "    nodes = []\n",
     "\n",
-    "    md_texts = [d[\"md\"] for d in json_dicts]\n",
+    "    md_texts = [page.md for page in result.pages]\n",
     "\n",
     "    for idx, md_text in enumerate(md_texts):\n",
     "        chunk_metadata = {\n",
     "            \"page_num\": idx + 1,\n",
-    "            \"paper_path\": paper_path,\n",
+    "            \"paper_path\": result.file_name,\n",
     "        }\n",
     "        node = TextNode(\n",
     "            text=md_text,\n",
@@ -316,11 +263,28 @@
     "# this will combine all nodes from all papers into a single list\n",
     "all_text_nodes = []\n",
     "text_nodes_dict = {}\n",
-    "for paper_path, paper_dict in paper_dicts.items():\n",
-    "    json_dicts = paper_dict[\"json_dicts\"]\n",
-    "    text_nodes = get_text_nodes(json_dicts, paper_dict[\"paper_path\"])\n",
+    "for result in results:\n",
+    "    text_nodes = get_text_nodes(result)\n",
     "    all_text_nodes.extend(text_nodes)\n",
-    "    text_nodes_dict[paper_path] = text_nodes"
+    "    text_nodes_dict[result.file_name] = text_nodes"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "id": "2e8fb9df",
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "106\n"
+     ]
+    }
+   ],
+   "source": [
+    "print(len(all_text_nodes))"
    ]
   },
   {
@@ -442,18 +406,15 @@
     "    The user will give the document text below.\n",
     "    \n",
     "    \"\"\"\n",
-    "    llm = llm or OpenAI(model=\"gpt-4o\")\n",
+    "    llm = llm or OpenAI(model=\"gpt-5-mini\", api_key=\"sk-...\")\n",
+    "    sllm = llm.as_structured_llm(SectionsOutput)\n",
     "\n",
-    "    chat_template = ChatPromptTemplate(\n",
-    "        [\n",
-    "            ChatMessage.from_str(system_prompt, \"system\"),\n",
-    "            ChatMessage.from_str(\"Document text: {doc_text}\", \"user\"),\n",
-    "        ]\n",
-    "    )\n",
-    "    result = await llm.astructured_predict(\n",
-    "        SectionsOutput, chat_template, doc_text=doc_text\n",
-    "    )\n",
-    "    return result.sections\n",
+    "    messages = [\n",
+    "        ChatMessage(content=system_prompt, role=\"system\"),\n",
+    "        ChatMessage(content=f\"Document text: {doc_text}\", role=\"user\"),\n",
+    "    ]\n",
+    "    result = await sllm.achat(messages)\n",
+    "    return result.raw.sections\n",
     "\n",
     "\n",
     "async def arefine_sections(\n",
@@ -472,23 +433,20 @@
     "    Given this, return the list of indexes that are valid. Do NOT include the indexes to be removed.\n",
     "    \n",
     "    \"\"\"\n",
-    "    llm = llm or OpenAI(model=\"gpt-4o\")\n",
-    "\n",
-    "    chat_template = ChatPromptTemplate(\n",
-    "        [\n",
-    "            ChatMessage.from_str(system_prompt, \"system\"),\n",
-    "            ChatMessage.from_str(\"Sections in text:\\n\\n{sections}\", \"user\"),\n",
-    "        ]\n",
-    "    )\n",
+    "    llm = llm or OpenAI(model=\"gpt-5-mini\", api_key=\"sk-...\")\n",
+    "    sllm = llm.as_structured_llm(ValidSections)\n",
     "\n",
     "    section_texts = \"\\n\".join(\n",
-    "        [f\"{idx}: {json.dumps(s.dict())}\" for idx, s in enumerate(sections)]\n",
+    "        [f\"{idx}: {json.dumps(s.model_dump())}\" for idx, s in enumerate(sections)]\n",
     "    )\n",
     "\n",
-    "    result = await llm.astructured_predict(\n",
-    "        ValidSections, chat_template, sections=section_texts\n",
-    "    )\n",
-    "    valid_indexes = result.valid_indexes\n",
+    "    messages = [\n",
+    "        ChatMessage(content=system_prompt, role=\"system\"),\n",
+    "        ChatMessage(content=f\"Sections in text:\\n\\n{section_texts}\", role=\"user\"),\n",
+    "    ]\n",
+    "\n",
+    "    result = await sllm.achat(messages)\n",
+    "    valid_indexes = result.raw.valid_indexes\n",
     "\n",
     "    new_sections = [s for idx, s in enumerate(sections) if idx in valid_indexes]\n",
     "    return new_sections\n",
@@ -514,17 +472,7 @@
    "execution_count": null,
    "id": "6e360a5c-29bd-4d86-9a21-f46013bab39a",
    "metadata": {},
-   "outputs": [
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "100%|██████████████████████████████████████████████████████████████████████| 51/51 [00:11<00:00,  4.35it/s]\n",
-      "100%|██████████████████████████████████████████████████████████████████████| 30/30 [00:09<00:00,  3.05it/s]\n",
-      "100%|██████████████████████████████████████████████████████████████████████| 25/25 [00:07<00:00,  3.22it/s]\n"
-     ]
-    }
-   ],
+   "outputs": [],
    "source": [
     "sections_dict = asyncio_run(acreate_sections(text_nodes_dict))"
    ]
@@ -538,36 +486,36 @@
     {
      "data": {
       "text/plain": [
-       "[SectionOutput(section_name='1', section_title='INTRODUCTION', start_page_number=1, is_subsection=False, description='# 1 INTRODUCTION'),\n",
-       " SectionOutput(section_name='2', section_title='BENCHMARK CONSTRUCTION', start_page_number=2, is_subsection=False, description='# BENCHMARK CONSTRUCTION'),\n",
-       " SectionOutput(section_name='2.2', section_title='TASK FORMULATION', start_page_number=3, is_subsection=True, description='# 2.2 TASK FORMULATION'),\n",
-       " SectionOutput(section_name='2.3', section_title='FEATURES OF SWE-BENCH', start_page_number=3, is_subsection=True, description='# 2.3 FEATURES OF SWE-BENCH'),\n",
-       " SectionOutput(section_name='3', section_title='SWE-LLAMA: FINE-TUNING CODELLAMA FOR SWE-BENCH', start_page_number=3, is_subsection=False, description='# 3 SWE-LLAMA: FINE-TUNING CODELLAMA FOR SWE-BENCH'),\n",
+       "[SectionOutput(section_name='1', section_title='Introduction', start_page_number=1, is_subsection=False, description='## 1 Introduction'),\n",
+       " SectionOutput(section_name='2.2', section_title='TASK FORMULATION', start_page_number=3, is_subsection=True, description='## 2.2 TASK FORMULATION'),\n",
+       " SectionOutput(section_name='2.3', section_title='FEATURES OF SWE-BENCH', start_page_number=3, is_subsection=True, description='## 2.3 FEATURES OF SWE-BENCH'),\n",
+       " SectionOutput(section_name='3', section_title='SWE-LLAMA: FINE-TUNING CODELLAMA FOR SWE-BENCH', start_page_number=3, is_subsection=False, description='## 3 SWE-LLAMA: FINE-TUNING CODELLAMA FOR SWE-BENCH'),\n",
        " SectionOutput(section_name='4', section_title='EXPERIMENTAL SETUP', start_page_number=4, is_subsection=False, description='# 4 EXPERIMENTAL SETUP'),\n",
-       " SectionOutput(section_name='4.1', section_title='RETRIEVAL-BASED APPROACH', start_page_number=4, is_subsection=True, description='# 4.1 RETRIEVAL-BASED APPROACH'),\n",
-       " SectionOutput(section_name='4.2', section_title='INPUT FORMAT', start_page_number=5, is_subsection=True, description='# 4.2 INPUT FORMAT'),\n",
-       " SectionOutput(section_name='4.3', section_title='MODELS', start_page_number=5, is_subsection=True, description='# 4.3 MODELS'),\n",
+       " SectionOutput(section_name='4.1', section_title='RETRIEVAL-BASED APPROACH', start_page_number=4, is_subsection=True, description='## 4.1 RETRIEVAL-BASED APPROACH'),\n",
+       " SectionOutput(section_name='4.2', section_title='INPUT FORMAT', start_page_number=5, is_subsection=True, description='## 4.2 INPUT FORMAT'),\n",
+       " SectionOutput(section_name='4.3', section_title='MODELS', start_page_number=5, is_subsection=True, description='## 4.3 MODELS'),\n",
        " SectionOutput(section_name='5', section_title='RESULTS', start_page_number=5, is_subsection=False, description='# 5 RESULTS'),\n",
        " SectionOutput(section_name='5.1', section_title='A QUALITATIVE ANALYSIS OF SWE-LLAMA GENERATIONS', start_page_number=8, is_subsection=True, description='# 5.1 A QUALITATIVE ANALYSIS OF SWE-LLAMA GENERATIONS'),\n",
        " SectionOutput(section_name='6', section_title='RELATED WORK', start_page_number=8, is_subsection=False, description='# 6 RELATED WORK'),\n",
-       " SectionOutput(section_name='7', section_title='DISCUSSION', start_page_number=9, is_subsection=False, description='# 7 DISCUSSION'),\n",
        " SectionOutput(section_name='8', section_title='ETHICS STATEMENT', start_page_number=10, is_subsection=False, description='# 8 ETHICS STATEMENT'),\n",
        " SectionOutput(section_name='9', section_title='REPRODUCIBILITY STATEMENT', start_page_number=10, is_subsection=False, description='# 9 REPRODUCIBILITY STATEMENT'),\n",
        " SectionOutput(section_name='10', section_title='ACKNOWLEDGEMENTS', start_page_number=10, is_subsection=False, description='# 10 ACKNOWLEDGEMENTS'),\n",
-       " SectionOutput(section_name='A', section_title='BENCHMARK DETAILS', start_page_number=15, is_subsection=False, description='# A BENCHMARK DETAILS'),\n",
-       " SectionOutput(section_name='A.1', section_title='HIGH LEVEL OVERVIEW', start_page_number=15, is_subsection=True, description='# A.1 HIGH LEVEL OVERVIEW'),\n",
-       " SectionOutput(section_name='A.2', section_title='CONSTRUCTION PROCESS', start_page_number=16, is_subsection=True, description='# A.2 CONSTRUCTION PROCESS'),\n",
-       " SectionOutput(section_name='A.3', section_title='Execution-Based Validation', start_page_number=18, is_subsection=True, description='# A.3 EXECUTION-BASED VALIDATION'),\n",
-       " SectionOutput(section_name='A.5', section_title='Evaluation Test Set Characterization', start_page_number=20, is_subsection=True, description='# A.5 EVALUATION TEST SET CHARACTERIZATION'),\n",
-       " SectionOutput(section_name='A.6', section_title='DEVELOPMENT SET CHARACTERIZATION', start_page_number=23, is_subsection=True, description='# A.6 DEVELOPMENT SET CHARACTERIZATION'),\n",
-       " SectionOutput(section_name='B', section_title='ADDITIONAL DETAILS ON TRAINING SWE-LLAMA', start_page_number=24, is_subsection=False, description='# B ADDITIONAL DETAILS ON TRAINING SWE-LLAMA'),\n",
-       " SectionOutput(section_name='B.1', section_title='TRAINING DETAILS', start_page_number=24, is_subsection=True, description='# B.1 TRAINING DETAILS'),\n",
-       " SectionOutput(section_name='D', section_title='ADDITIONAL EXPERIMENTAL DETAILS', start_page_number=28, is_subsection=False, description='# D ADDITIONAL EXPERIMENTAL DETAILS'),\n",
-       " SectionOutput(section_name='D.1', section_title='RETRIEVAL DETAILS', start_page_number=28, is_subsection=True, description='# D.1 RETRIEVAL DETAILS'),\n",
-       " SectionOutput(section_name='D.2', section_title='INFERENCE SETTINGS', start_page_number=29, is_subsection=True, description='# D.2 INFERENCE SETTINGS'),\n",
-       " SectionOutput(section_name='D.3', section_title='PROMPT TEMPLATE EXAMPLE', start_page_number=29, is_subsection=True, description='# D.3 PROMPT TEMPLATE EXAMPLE'),\n",
-       " SectionOutput(section_name='E', section_title='Societal Impact', start_page_number=31, is_subsection=False, description='# E SOCIETAL IMPACT'),\n",
-       " SectionOutput(section_name='F', section_title='In-Depth Analysis of SWE-Llama Generations', start_page_number=31, is_subsection=False, description='# F IN-DEPTH ANALYSIS OF SWE-LLAMA GENERATIONS')]"
+       " SectionOutput(section_name='A.1', section_title='HIGH LEVEL OVERVIEW', start_page_number=15, is_subsection=True, description='### A.1 HIGH LEVEL OVERVIEW'),\n",
+       " SectionOutput(section_name='A.2', section_title='CONSTRUCTION PROCESS', start_page_number=16, is_subsection=True, description='## A.2 CONSTRUCTION PROCESS'),\n",
+       " SectionOutput(section_name='A.3', section_title='EXECUTION-BASED VALIDATION', start_page_number=18, is_subsection=True, description='### A.3 EXECUTION-BASED VALIDATION'),\n",
+       " SectionOutput(section_name='A.4', section_title='EVALUATION PROCEDURE', start_page_number=19, is_subsection=True, description='## A.4 EVALUATION PROCEDURE'),\n",
+       " SectionOutput(section_name='A.5', section_title='EVALUATION TEST SET CHARACTERIZATION', start_page_number=20, is_subsection=True, description='## A.5 EVALUATION TEST SET CHARACTERIZATION'),\n",
+       " SectionOutput(section_name='A.6', section_title='DEVELOPMENT SET CHARACTERIZATION', start_page_number=23, is_subsection=True, description='## A.6 DEVELOPMENT SET CHARACTERIZATION'),\n",
+       " SectionOutput(section_name='B.1', section_title='TRAINING DETAILS', start_page_number=24, is_subsection=True, description='## B.1 TRAINING DETAILS'),\n",
+       " SectionOutput(section_name='C.1', section_title='RESULTS WITH “ORACLE” RETRIEVAL', start_page_number=24, is_subsection=True, description='## C.1 RESULTS WITH “ORACLE” RETRIEVAL'),\n",
+       " SectionOutput(section_name='C.2', section_title='EVALUATION TEST SET', start_page_number=24, is_subsection=True, description='## C.2 EVALUATION TEST SET'),\n",
+       " SectionOutput(section_name='C.3', section_title='GPT-4 EVALUATION SUBSET RESULTS', start_page_number=24, is_subsection=True, description='## C.3 GPT-4 EVALUATION SUBSET RESULTS'),\n",
+       " SectionOutput(section_name='C.4', section_title='EXTENDED TEMPORAL ANALYSIS', start_page_number=25, is_subsection=True, description='## C.4 EXTENDED TEMPORAL ANALYSIS'),\n",
+       " SectionOutput(section_name='C.5', section_title='F2P, P2P RATE ANALYSIS', start_page_number=25, is_subsection=True, description='## C.5 F2P, P2P RATE ANALYSIS'),\n",
+       " SectionOutput(section_name='C.7', section_title='SOFTWARE ENGINEERING METRICS', start_page_number=27, is_subsection=True, description='## C.7 SOFTWARE ENGINEERING METRICS'),\n",
+       " SectionOutput(section_name='D.1', section_title='RETRIEVAL DETAILS', start_page_number=28, is_subsection=True, description='## D.1 RETRIEVAL DETAILS'),\n",
+       " SectionOutput(section_name='D.2', section_title='INFERENCE SETTINGS', start_page_number=29, is_subsection=True, description='## D.2 INFERENCE SETTINGS'),\n",
+       " SectionOutput(section_name='D.3', section_title='PROMPT TEMPLATE EXAMPLE', start_page_number=29, is_subsection=True, description='## D.3 PROMPT TEMPLATE EXAMPLE')]"
       ]
      },
      "execution_count": null,
@@ -576,7 +524,7 @@
     }
    ],
    "source": [
-    "sections_dict[\"swebench.pdf\"]"
+    "sections_dict[\"iclr_docs/swebench.pdf\"]"
    ]
   },
   {
@@ -755,7 +703,7 @@
     "from llama_index.vector_stores.chroma import ChromaVectorStore\n",
     "from llama_index.core import VectorStoreIndex\n",
     "\n",
-    "persist_dir = \"storage_chroma\"\n",
+    "persist_dir = \"chroma_storage\"\n",
     "\n",
     "vector_store = ChromaVectorStore.from_params(\n",
     "    collection_name=\"text_nodes\", persist_dir=persist_dir\n",
@@ -805,7 +753,7 @@
    "source": [
     "from llama_index.llms.openai import OpenAI\n",
     "\n",
-    "llm = OpenAI(model=\"gpt-4o\")"
+    "llm = OpenAI(model=\"gpt-5-mini\", api_key=\"sk-...\")"
    ]
   },
   {
@@ -833,6 +781,7 @@
     "    FilterCondition,\n",
     ")\n",
     "from llama_index.core.schema import NodeWithScore\n",
+    "from typing import List\n",
     "\n",
     "\n",
     "def section_retrieve(query: str, verbose: bool = False) -> List[NodeWithScore]:\n",
@@ -870,57 +819,6 @@
     "    return all_section_nodes.values()"
    ]
   },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "f721e770-ce4c-4511-96d5-8a89d16c7281",
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      ">> Identifying the right sections to retrieve\n",
-      ">> Retrieving section: A: BENCHMARK DETAILS\n",
-      ">> Retrieving section: 2: BENCHMARK CONSTRUCTION\n",
-      ">> Retrieving section: A: BENCHMARK DETAILS\n"
-     ]
-    }
-   ],
-   "source": [
-    "nodes = section_retrieve(\n",
-    "    \"Give me a full overview of the benchmark details in SWE Bench\", verbose=True\n",
-    ")"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "e99eaa71-7d93-40c0-bba0-a9c983a6cbd3",
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "{'page_num': 15, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': 'A: BENCHMARK DETAILS', 'sub_section_id': 'A.1: HIGH LEVEL OVERVIEW'}\n",
-      "{'page_num': 16, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': 'A: BENCHMARK DETAILS', 'sub_section_id': 'A.2: CONSTRUCTION PROCESS'}\n",
-      "{'page_num': 17, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': 'A: BENCHMARK DETAILS', 'sub_section_id': 'A.2: CONSTRUCTION PROCESS'}\n",
-      "{'page_num': 18, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': 'A: BENCHMARK DETAILS', 'sub_section_id': 'A.3: Execution-Based Validation'}\n",
-      "{'page_num': 19, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': 'A: BENCHMARK DETAILS', 'sub_section_id': 'A.3: Execution-Based Validation'}\n",
-      "{'page_num': 20, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': 'A: BENCHMARK DETAILS', 'sub_section_id': 'A.5: Evaluation Test Set Characterization'}\n",
-      "{'page_num': 21, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': 'A: BENCHMARK DETAILS', 'sub_section_id': 'A.5: Evaluation Test Set Characterization'}\n",
-      "{'page_num': 22, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': 'A: BENCHMARK DETAILS', 'sub_section_id': 'A.5: Evaluation Test Set Characterization'}\n",
-      "{'page_num': 23, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': 'A: BENCHMARK DETAILS', 'sub_section_id': 'A.6: DEVELOPMENT SET CHARACTERIZATION'}\n",
-      "{'page_num': 2, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': '2: BENCHMARK CONSTRUCTION', 'sub_section_id': '2: BENCHMARK CONSTRUCTION'}\n"
-     ]
-    }
-   ],
-   "source": [
-    "for n in nodes:\n",
-    "    print(n.node.metadata)"
-   ]
-  },
   {
    "cell_type": "code",
    "execution_count": null,
@@ -932,9 +830,9 @@
      "output_type": "stream",
      "text": [
       ">> Identifying the right sections to retrieve\n",
-      ">> Retrieving section: F: ADDITIONAL RESULTS\n",
+      ">> Retrieving section: 6: Conclusion\n",
       ">> Retrieving section: 5: EXPERIMENTS\n",
-      ">> Retrieving section: F: ADDITIONAL RESULTS\n"
+      ">> Retrieving section: 5: EXPERIMENTS\n"
      ]
     }
    ],
@@ -955,11 +853,26 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "{'page_num': 21, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': 'F: ADDITIONAL RESULTS', 'sub_section_id': 'F.1: FULL QUALITATIVE RESULTS'}\n",
-      "{'page_num': 22, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': 'F: ADDITIONAL RESULTS', 'sub_section_id': 'F.4: Additional Baselines'}\n",
+      "{'page_num': 9, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '6: Conclusion', 'sub_section_id': '6: Conclusion'}\n",
+      "{'page_num': 10, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '6: Conclusion', 'sub_section_id': '6: Conclusion'}\n",
+      "{'page_num': 11, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '6: Conclusion', 'sub_section_id': '6: Conclusion'}\n",
+      "{'page_num': 12, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '6: Conclusion', 'sub_section_id': '6: Conclusion'}\n",
+      "{'page_num': 13, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '6: Conclusion', 'sub_section_id': '6: Conclusion'}\n",
+      "{'page_num': 14, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '6: Conclusion', 'sub_section_id': '6: Conclusion'}\n",
+      "{'page_num': 15, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '6: Conclusion', 'sub_section_id': '6: Conclusion'}\n",
+      "{'page_num': 16, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '6: Conclusion', 'sub_section_id': '6: Conclusion'}\n",
+      "{'page_num': 17, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '6: Conclusion', 'sub_section_id': 'C.1: Universality of Inner Product Decomposition'}\n",
+      "{'page_num': 18, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '6: Conclusion', 'sub_section_id': 'C.2: Lipschitz Constraint under the Temporal Distance Metric'}\n",
+      "{'page_num': 19, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '6: Conclusion', 'sub_section_id': 'C.2: Lipschitz Constraint under the Temporal Distance Metric'}\n",
+      "{'page_num': 20, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '6: Conclusion', 'sub_section_id': 'E.2: DADS'}\n",
+      "{'page_num': 21, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '6: Conclusion', 'sub_section_id': 'F.1: FULL QUALITATIVE RESULTS'}\n",
+      "{'page_num': 22, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '6: Conclusion', 'sub_section_id': 'F.4: ADDITIONAL BASELINES'}\n",
+      "{'page_num': 23, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '6: Conclusion', 'sub_section_id': 'G.1: Environments'}\n",
+      "{'page_num': 24, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '6: Conclusion', 'sub_section_id': 'G.2: IMPLEMENTATION DETAILS'}\n",
+      "{'page_num': 25, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '6: Conclusion', 'sub_section_id': 'G.2: IMPLEMENTATION DETAILS'}\n",
       "{'page_num': 6, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '5: EXPERIMENTS', 'sub_section_id': '5: EXPERIMENTS'}\n",
       "{'page_num': 7, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '5: EXPERIMENTS', 'sub_section_id': '5.2: QUALITATIVE COMPARISON'}\n",
-      "{'page_num': 8, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '5: EXPERIMENTS', 'sub_section_id': '5.3: QUANTITATIVE COMPARISON'}\n"
+      "{'page_num': 8, 'paper_path': 'iclr_docs/metra.pdf', 'section_id': '5: EXPERIMENTS', 'sub_section_id': '5.3: Quantitative Comparison'}\n"
      ]
     }
    ],
@@ -1027,10 +940,24 @@
      "output_type": "stream",
      "text": [
       ">> Identifying the right sections to retrieve\n",
-      ">> Retrieving section: A: BENCHMARK DETAILS\n",
       ">> Retrieving section: 5: RESULTS\n",
-      ">> Retrieving section: A: BENCHMARK DETAILS\n",
-      "In SWEBench, difficulty correlates with context length in a way that as the total context length increases, model performance tends to drop. This is observed across various models, including Claude 2, which shows a significant decrease in performance with longer context lengths. The models often struggle to localize the problematic code that needs updating when presented with a lot of code that may not be directly related to the issue at hand. This suggests that models can become distracted by additional context, which aligns with findings from other studies indicating that models may be sensitive to the relative location of target sequences. Even when increasing the maximum context size improves recall with respect to the oracle files, performance still drops, indicating that models are ineffective at localizing the necessary code changes.\n"
+      ">> Retrieving section: 3: SWE-LLAMA: FINE-TUNING CODELLAMA FOR SWE-BENCH\n",
+      ">> Retrieving section: 4: EXPERIMENTAL SETUP\n",
+      "Key findings about how difficulty correlates with context length\n",
+      "\n",
+      "- Performance falls as total input/context size grows. As the amount of code and other context provided to models increases, their ability to localize and produce correct edits drops noticeably (this behavior was observed across multiple models, e.g., Claude 2 and others).\n",
+      "\n",
+      "- Extra (irrelevant) context distracts models. When models are given a lot of code that is unrelated to the actual edit, they frequently struggle to find the problematic lines that need changing. This sensitivity includes the relative location of the target code within the larger context.\n",
+      "\n",
+      "- Increasing retriever recall doesn't fix it. Expanding retrieval windows (to include more files and therefore raise oracle recall) can actually hurt end-to-end performance because models become less effective at pinpointing the needed edits amid the extra material.\n",
+      "\n",
+      "- Collapsing context around the true edits helps. An ablation that collapses retrieved files to only the lines actually modified in the reference patch (±15 lines) improved results — for example, one model’s resolved rate rose from 4.8% to 5.9%, and another increased from ~1.3% to 3.4% — showing that concentrating context on the most relevant snippets makes the task easier.\n",
+      "\n",
+      "- Finetuned models are sensitive to context-distribution shifts. Models fine-tuned on tightly scoped (oracle) contexts performed worse when given BM25-retrieved context that contained many irrelevant files, indicating that training with one style of context can reduce robustness to different retrieval outputs.\n",
+      "\n",
+      "Implications\n",
+      "- Better retrieval or context-compression methods (e.g., more precise retrieval, collapsing to edited regions, or preprocessing to highlight likely relevant locations) are likely more useful than simply increasing context size.\n",
+      "- Robust model behavior requires not just larger windows but mechanisms for localization and filtering of relevant code within long contexts.\n"
      ]
     }
    ],
@@ -1052,18 +979,98 @@
      "output_type": "stream",
      "text": [
       ">> Identifying the right sections to retrieve\n",
-      ">> Retrieving section: A: BENCHMARK DETAILS\n",
-      ">> Retrieving section: 2: BENCHMARK CONSTRUCTION\n",
-      ">> Retrieving section: A: BENCHMARK DETAILS\n",
-      "SWE-bench is a benchmark designed to evaluate language models in a realistic software engineering setting by using GitHub issues and pull requests from popular repositories. The benchmark involves generating a pull request that addresses a given issue and passes related tests. The construction of SWE-bench involves a three-stage pipeline:\n",
+      ">> Retrieving section: 10: ACKNOWLEDGEMENTS\n",
+      ">> Retrieving section: 1: Introduction\n",
+      ">> Retrieving section: 3: SWE-LLAMA: FINE-TUNING CODELLAMA FOR SWE-BENCH\n",
+      "High-level summary\n",
+      "- SWE-bench is a repository-scale, execution-validated benchmark of real GitHub issues paired with merged pull-request solutions. Each task gives a snapshot of a real codebase plus an issue description; the model must produce a patch that, when applied, makes the repository pass the tests that verify the issue was addressed.\n",
+      "- The benchmark emphasizes realistic, hard software-engineering problems: large codebases, multi-file edits, long issue descriptions, and unit tests used for automatic verification.\n",
       "\n",
-      "1. **Repo Selection and Data Scraping**: Pull requests are collected from 12 popular open-source Python repositories on GitHub, resulting in approximately 90,000 PRs. These repositories are chosen for their better maintenance, clear contributor guidelines, and comprehensive test coverage.\n",
+      "Data sources and collection\n",
+      "- Candidate PRs are sourced from popular Python projects (selected from highly downloaded PyPI packages and mapped to their GitHub repositories). Repositories are filtered to ensure permissible licenses.\n",
+      "- Pull requests are collected via the GitHub API and then filtered automatically.\n",
       "\n",
-      "2. **Attribute-Based Filtering**: Candidate tasks are created by selecting merged PRs that resolve a GitHub issue and contribute tests. This indicates that the user likely added tests to verify the resolution of the issue.\n",
+      "Task-instance selection criteria\n",
+      "A PR becomes a candidate task only if it satisfies all of:\n",
+      "- Status = merged (the PR was accepted).\n",
+      "- The PR resolves one or more GitHub issues (detected via links like “fixes #N” in title/body/commits).\n",
+      "- The PR introduces or edits test files (file paths containing test-related keywords).\n",
+      "Only candidates that pass execution-based validation are kept.\n",
       "\n",
-      "3. **Execution-Based Filtering**: For each candidate task, the PR's test content is applied, and test results are logged before and after applying the PR's other content. Tasks are filtered out if they do not have at least one test that changes from fail to pass or if they result in installation or runtime errors.\n",
+      "Task-instance components\n",
+      "Each task instance encodes:\n",
+      "- Codebase reference C: repo owner/name and the base commit (mirrored repositories are created so code can be retrieved reproducibly).\n",
+      "- Problem statement P: aggregated issue titles and descriptions and any issue/PR comments up to the PR’s first commit (no post-solution comments that would leak the fix).\n",
+      "- Tests T: the tests introduced/edited by the PR (extracted from the PR diff and stored as a .patch).\n",
+      "- Solution δ (gold patch): the PR’s code changes excluding test edits (stored as a .patch).\n",
+      "- Metadata fields: base_commit, created_at, instance_id, issue_numbers, repo, pull_number, version, env_install_commit, hints_text (collected comments), and cached test result mappings like FAIL_TO_PASS and PASS_TO_PASS.\n",
       "\n",
-      "The benchmark is designed to be extensible, allowing for updates with new task instances as new language models are released. It includes a robust framework for execution-based evaluation, ensuring that generated solutions can be verified by running unit tests. SWE-bench also provides a training dataset, SWE-bench-train, and fine-tuned models like SWE-Llama 7b and 13b, which are based on the CodeLlama model. These models are evaluated on their ability to resolve issues, with SWE-Llama 13b showing competitive performance in some settings.\n"
+      "Execution-based validation (quality control)\n",
+      "- Virtual execution contexts are created per repository release version (manual inspection of README/contributing to determine Python version, dependencies, install commands). Conda environments are used.\n",
+      "- For each candidate instance the pipeline:\n",
+      "  1. Checks out the base commit.\n",
+      "  2. Installs the codebase in the corresponding env.\n",
+      "  3. Applies the test patch T and runs tests (log_pre).\n",
+      "  4. Applies the solution patch δ and runs tests again (log_post).\n",
+      "- Candidates are discarded if any step fails (checkout, install, apply patch, test run).\n",
+      "- Instances are retained only if at least one test changes from fail → pass (a true FAIL_TO_PASS) and if there are no trivial issues (e.g., ImportError or AttributeError in log_pre that indicate missing dependency/name issues).\n",
+      "- Instances whose tests exercise newly created functions/classes (i.e., tests requiring names introduced by δ) are excluded because they would be impossible to solve from the problem statement alone.\n",
+      "\n",
+      "Task-instance format and artifacts\n",
+      "- Finalized instances are saved in a single JSON file (task metadata and patch contents are included as patch-format strings).\n",
+      "- For each instance the validation engine caches parsed test-to-status mappings for log_pre/log_post and creates ground-truth lists: FAIL_TO_PASS, PASS_TO_PASS (used during evaluation to check both that the fix was implemented and that prior behavior is preserved).\n",
+      "- Mirrors of original repositories are created and stored to preserve exact base commits and enable reproducible checkout.\n",
+      "\n",
+      "Evaluation procedure (how models are scored)\n",
+      "- Model input: problem statement P and the codebase C (usually limited by retrieval/long-context strategy). The model must generate a single .patch (a git/unified-diff style patch).\n",
+      "- Per predicted patch the evaluation harness:\n",
+      "  1. Resets repo to base commit.\n",
+      "  2. Activates the executable context for the instance version.\n",
+      "  3. Installs the codebase.\n",
+      "  4. Applies the test patch T.\n",
+      "  5. Attempts to apply the predicted patch \\hat{δ}. If applying fails, an automatic \"patch-fix\" step tries to repair the patch (e.g., strip extraneous context lines and recalculate headers); if it still fails the prediction is scored as failure.\n",
+      "  6. Runs the repository’s test command to generate log_{\\hat{δ}}.\n",
+      "  7. Parses log_{\\hat{δ}} into a test-to-status mapping using repository-specific parsers.\n",
+      "  8. Declares the task solved only if all tests listed in FAIL_TO_PASS and PASS_TO_PASS have status = pass in log_{\\hat{δ}}.\n",
+      "- The principal metric is % Resolved: fraction of task instances fully solved (all required tests pass).\n",
+      "\n",
+      "Patch-fixing and robustness\n",
+      "- If a generated patch does not apply, the harness attempts an automated repair (e.g., removing context lines, fixing header offsets) before giving up. Applied-but-broken patches that then fail tests are classified according to pass/fail patterns (Resolved, Breaking Resolved, Partially Resolved, Work-in-Progress, No-Op, Regression) to provide finer-grained analysis.\n",
+      "\n",
+      "Dataset scale and characterization\n",
+      "- Raw crawl: ~93k PRs across selected repositories; after conversion/filters and execution validation the final evaluation set contains 2,294 task instances.\n",
+      "- Instances come from 12 widely used Python repositories with varied sizes and purposes (e.g., scikit-learn, Django, matplotlib, requests, pytest, sympy, astropy, etc.).\n",
+      "- Typical instance properties: long problem descriptions (median ~140 words), large repositories (median ~thousands of files and hundreds of thousands of lines), and reference edits that usually touch ~1–2 files, edit a few functions, and modify a few dozen lines on average.\n",
+      "- Tests: each instance has at least one FAIL_TO_PASS; many instances include many PASS_TO_PASS tests for regression protection (median tens to hundreds of pass-to-pass tests).\n",
+      "\n",
+      "Development set, train set, and extensions\n",
+      "- A smaller development set (~225 instances, >10% of the main set) is provided for tuning and debugging.\n",
+      "- A separate SWE-bench-train dataset (19k non-testing task instances from many repos) was prepared for fine-tuning models; fine-tuned models were released (SWE-Llama 7B and 13B) to study open-model performance on long contexts.\n",
+      "- The collection pipeline and mirror strategy were designed to be easily extendable so the benchmark can be updated continuously with new PRs and support additional languages or repos.\n",
+      "\n",
+      "Reproducibility and release commitments\n",
+      "- The codebase used to collect, validate, and evaluate task instances is organized and documented; mirrors and the JSON of task instances are provided so others can reproduce experiments.\n",
+      "- Execution contexts, validation logs, and ground-truth test mappings are cached to avoid re-running expensive validation at evaluation time.\n",
+      "- Plans include open-sourcing the task instances, collection/evaluation infrastructure, training data used for fine-tuning, and model weights along with documentation.\n",
+      "\n",
+      "Design decisions and safeguards\n",
+      "- Using merged PRs that added tests provides a strong ground-truth signal that the PR truly solved the issue and allowed for reproducible verification.\n",
+      "- Excluding instances with trivial dependency/name errors or tests that require newly-introduced symbol names ensures tasks are solvable from the given P + C without hidden knowledge.\n",
+      "- Mirroring repositories preserves commit history and avoids breakage from later upstream edits.\n",
+      "\n",
+      "What solving a task means (concrete criterion)\n",
+      "- A generated patch must apply and, after applying the repository’s tests, every test that the validation flagged as verifying the issue (FAIL_TO_PASS) must now pass, and all tests that previously passed but were intended to remain passing (PASS_TO_PASS) must still pass. Only then is the task counted as solved.\n",
+      "\n",
+      "Utility and intended uses\n",
+      "- The benchmark measures model ability to: localize defects, reason across a large codebase, produce multi-line and multi-file edits in patch format, and use execution feedback (tests) as verification.\n",
+      "- It is intended both as a hard evaluation for current models and as a development target for models and systems that perform repository-scale code edits, retrieval from large codebases, iterative editing with execution feedback, or agent-style multi-step repair.\n",
+      "\n",
+      "Limitations to be aware of\n",
+      "- The benchmark focuses on repositories with permissive licenses and decent test coverage (popular projects), so it emphasizes bug fixes and features that were covered by tests and merged in those projects.\n",
+      "- Some tasks that require creating new symbol names first introduced in the solution are excluded because they would not be solvable from the baseline inputs.\n",
+      "- Execution environments are created per release version (manual aspects exist), and some instances are discarded when installation or environment setup cannot be reliably reproduced.\n",
+      "\n",
+      "Overall, SWE-bench provides a large, execution-validated, reproducible suite of real-world repository-scale code-editing tasks that require understanding long contexts and producing correct patch-format edits verified by the project’s own tests.\n"
      ]
     }
    ],
@@ -1074,34 +1081,6 @@
     "print(str(response))"
    ]
   },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "6d747bf8-0ed2-4c10-8108-9d0e8d53a4fb",
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "{'page_num': 15, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': 'A: BENCHMARK DETAILS', 'sub_section_id': 'A.1: HIGH LEVEL OVERVIEW'}\n",
-      "{'page_num': 16, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': 'A: BENCHMARK DETAILS', 'sub_section_id': 'A.2: CONSTRUCTION PROCESS'}\n",
-      "{'page_num': 17, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': 'A: BENCHMARK DETAILS', 'sub_section_id': 'A.2: CONSTRUCTION PROCESS'}\n",
-      "{'page_num': 18, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': 'A: BENCHMARK DETAILS', 'sub_section_id': 'A.3: Execution-Based Validation'}\n",
-      "{'page_num': 19, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': 'A: BENCHMARK DETAILS', 'sub_section_id': 'A.3: Execution-Based Validation'}\n",
-      "{'page_num': 20, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': 'A: BENCHMARK DETAILS', 'sub_section_id': 'A.5: Evaluation Test Set Characterization'}\n",
-      "{'page_num': 21, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': 'A: BENCHMARK DETAILS', 'sub_section_id': 'A.5: Evaluation Test Set Characterization'}\n",
-      "{'page_num': 22, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': 'A: BENCHMARK DETAILS', 'sub_section_id': 'A.5: Evaluation Test Set Characterization'}\n",
-      "{'page_num': 23, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': 'A: BENCHMARK DETAILS', 'sub_section_id': 'A.6: DEVELOPMENT SET CHARACTERIZATION'}\n",
-      "{'page_num': 2, 'paper_path': 'iclr_docs/swebench.pdf', 'section_id': '2: BENCHMARK CONSTRUCTION', 'sub_section_id': '2: BENCHMARK CONSTRUCTION'}\n"
-     ]
-    }
-   ],
-   "source": [
-    "for n in response.source_nodes:\n",
-    "    print(n.metadata)"
-   ]
-  },
   {
    "cell_type": "code",
    "execution_count": null,
@@ -1113,20 +1092,76 @@
      "output_type": "stream",
      "text": [
       ">> Identifying the right sections to retrieve\n",
-      ">> Retrieving section: F: ADDITIONAL RESULTS\n",
+      ">> Retrieving section: 6: Conclusion\n",
       ">> Retrieving section: 5: EXPERIMENTS\n",
-      ">> Retrieving section: F: ADDITIONAL RESULTS\n",
-      "The additional experimental results in the METRA paper include several key findings:\n",
+      ">> Retrieving section: 5: EXPERIMENTS\n",
+      "Here are the additional experimental results and analyses reported.\n",
       "\n",
-      "1. **Full Qualitative Results**: METRA discovers diverse locomotion behaviors across different environments, including state-based Ant and HalfCheetah, and pixel-based Quadruped and Humanoid. The results are consistent across multiple random seeds, indicating robustness in behavior discovery.\n",
+      "1) Full qualitative results (complete skill behaviors, 8 seeds)\n",
+      "- Environments: state-based Ant and HalfCheetah; pixel-based Quadruped and Humanoid.\n",
+      "- Skill parameterizations used in these visualizations: 2-D continuous skills for Ant and Humanoid, 4-D continuous skills for Quadruped, 16 discrete skills for HalfCheetah.\n",
+      "- Main finding: across 8 random seeds METRA consistently discovers diverse locomotion behaviors (radial/x-y coverage, different locomotion modes) regardless of seed. The paper shows multiple sample trajectories per seed to illustrate robustness and diversity.\n",
       "\n",
-      "2. **Latent Space Visualization**: METRA effectively captures the most temporally spread-out dimensions in the state space, such as x-y coordinates, in its latent space. This is demonstrated in both state-based and pixel-based environments, with higher-dimensional latent spaces capturing more diverse behaviors.\n",
+      "2) Latent-space visualization\n",
+      "- Setup: METRA trained with 2-D continuous latent space on Ant (state inputs) and Humanoid (pixel inputs).\n",
+      "- Observation: the learned representation φ(s) captures the agent’s x-y coordinates in the 2-D latent space in both Ant and Humanoid. The learned φ trajectories align with the x-y trajectories, indicating METRA finds the temporally most spread-out manifold (x-y plane) even from pixels.\n",
+      "- Note: with higher-dimensional or discrete latent spaces, METRA captures more diverse, non-linear behaviors beyond simple locomotion.\n",
       "\n",
-      "3. **Ablation Study of Latent Space Sizes**: The study shows that increasing the size of the latent space generally enhances the diversity of skills learned by METRA. Different dimensions of continuous and discrete skills were tested on Ant and HalfCheetah.\n",
+      "3) Ablation: effect of latent-space size on learned skills\n",
+      "- Latent-space sizes tested: 1-D, 2-D, 4-D continuous; discrete sets of sizes {2}, {4}, {8}, {16}, {24}.\n",
+      "- Environments: Ant and HalfCheetah.\n",
+      "- Result: skill diversity increases as the capacity (dimensionality / cardinality) of Z grows.\n",
+      "  - 1-D: simple linear/one-dimensional coverage\n",
+      "  - 2-D: radial coverage / 2-D spread\n",
+      "  - 4-D: more complex radial / richer behaviors\n",
+      "  - Discrete increases produce progressively more distinct discrete behaviors (more segments, more diverse skill classes)\n",
+      "- Conclusion: METRA maximizes state coverage under latent capacity, so increasing Z’s capacity yields more diverse discovered behaviors.\n",
       "\n",
-      "4. **Comparison with Additional Baselines**: METRA was compared with DGPO, a method focused on finding diverse behaviors that maximize task rewards. The comparison was conducted in a controlled Markov process setting without external rewards, using only intrinsic rewards.\n",
+      "4) Additional baseline: DGPO comparison (discrete-skill comparison; 4 seeds)\n",
+      "- Experimental setup: DIAYN, DGPO, and METRA were trained with 16 discrete skills for 10,000 epochs (≈16M environment steps).\n",
+      "- Metrics reported: policy state coverage and total state coverage (means ± std).\n",
+      "- Results (Table reproduced):\n",
+      "  - HalfCheetah (policy state coverage)\n",
+      "    - DIAYN: 6.75 ± 2.22\n",
+      "    - DGPO: 6.75 ± 2.06\n",
+      "    - METRA: 186.75 ± 16.21\n",
+      "  - HalfCheetah (total state coverage)\n",
+      "    - DIAYN: 19.50 ± 3.87\n",
+      "    - DGPO: 22.25 ± 5.85\n",
+      "    - METRA: 177.75 ± 17.10\n",
+      "  - Ant (policy state coverage)\n",
+      "    - DIAYN: 11.25 ± 5.44\n",
+      "    - DGPO: 7.00 ± 3.83\n",
+      "    - METRA: 1387.75 ± 77.38\n",
+      "  - Ant (total state coverage)\n",
+      "    - DIAYN: 107.75 ± 17.00\n",
+      "    - DGPO: 121.50 ± 4.36\n",
+      "    - METRA: 6313.25 ± 747.92\n",
+      "- Interpretation given: DGPO (which maximizes a metric-agnostic KL-style objective in discrete Z) still produces limited state coverage similar to DIAYN, whereas METRA (a metric-aware Wasserstein formulation) achieves substantially greater coverage in these locomotion environments.\n",
       "\n",
-      "These results highlight METRA's ability to discover diverse and meaningful behaviors in various environments, its effective use of latent spaces, and its performance relative to other methods.\n"
+      "5) Skill examples / qualitative descriptions by latent size\n",
+      "- A tabulated description shows how skills change qualitatively with latent-size choices (examples):\n",
+      "  - Ant (continuous Z):\n",
+      "    - 1-D: linearly increasing coverage\n",
+      "    - 2-D: radial coverage with 2-D spread\n",
+      "    - 4-D: more complex radial coverage\n",
+      "  - Ant / HalfCheetah (discrete Z):\n",
+      "    - Discrete 2 / 4 / 8 / 16 / 24 skills: progressively more segments and more diverse behaviors, with 24 discrete skills showing the highest diversity.\n",
+      "- The paper notes that with discrete Z METRA can discover qualitatively distinct behaviors such as flips or static postures (in addition to locomotion) when capacity is sufficient.\n",
+      "\n",
+      "6) Details on coverage metrics, datasets, and protocol used in these additional results\n",
+      "- Policy state coverage: computed by sampling 48 deterministic trajectories using 48 randomly sampled skills at each evaluation epoch (used for skill-discovery method policy coverage plots).\n",
+      "- Queue state coverage: computed from most recent 100,000 training trajectories (used for some comparisons).\n",
+      "- Total state coverage: computed from the entire set of training trajectories up to the current epoch (used as a generous metric for pure-exploration baselines).\n",
+      "- For locomotion coverage counting: x-y bins of 1×1 are counted for Ant, Quadruped, Humanoid; x bins for HalfCheetah. Kitchen uses task success counts for pre-defined subtasks.\n",
+      "- Seeds: most qualitative and skill-discovery comparisons use 8 seeds; the DGPO comparison reported used 4 seeds.\n",
+      "\n",
+      "7) Additional notes and takeaways from the extra experiments\n",
+      "- METRA’s learned φ(s) is effective for zero-shot goal selection because φ preserves temporal distances; the latent difference φ(g) − φ(s) gives a direction in Z to reach a goal.\n",
+      "- Increasing latent capacity helps but requires choosing continuous vs. discrete Z appropriately for the desired types of behaviors.\n",
+      "- The DGPO comparison further supports that metric-aware objectives (METRA) lead to substantially higher state coverage than metric-agnostic mutual-information/KL-style objectives.\n",
+      "\n",
+      "If you want, I can extract and present the specific numeric tables and captions (e.g., the full Table 1 numbers above) in CSV or another concise format, or summarize the visual findings into representative example trajectories for each latent-size setting.\n"
      ]
     }
    ],
@@ -1140,9 +1175,9 @@
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "llama_index_v3",
+   "display_name": ".venv",
    "language": "python",
-   "name": "llama_index_v3"
+   "name": "python3"
   },
   "language_info": {
    "codemirror_mode": {

examples/parse/agents/demo_simple_openai_agent.ipynb

@@ -6,7 +6,12 @@
    "source": [
     "# LlamaParse Agent\n",
     "\n",
-    "This demo walks through using an OpenAI Agent with [LlamaParse](https://cloud.llamaindex.ai)."
+    "This demo walks through using an OpenAI Agent with [LlamaParse](https://cloud.llamaindex.ai).\n",
+    "\n",
+    "Status:\n",
+    "| Last Executed | Version | State      |\n",
+    "|---------------|---------|------------|\n",
+    "| Aug-19-2025   | 0.6.61  | Maintained |"
    ]
   },
   {
@@ -22,7 +27,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "!pip install llama-cloud-services llama-index llama-index-postprocessor-sbert-rerank"
+    "!pip install llama-cloud-services \"llama-index>=0.13.0<0.14.0\""
    ]
   },
   {
@@ -48,7 +53,7 @@
     "from llama_index.llms.openai import OpenAI\n",
     "\n",
     "Settings.embed_model = OpenAIEmbedding(model=\"text-embedding-3-small\")\n",
-    "Settings.llm = OpenAI(model=\"gpt-3.5-turbo\", temperature=0.2)"
+    "Settings.llm = OpenAI(model=\"gpt-5-mini\")"
    ]
   },
   {
@@ -83,9 +88,15 @@
    "outputs": [],
    "source": [
     "from llama_cloud_services import LlamaParse\n",
+    "from sympy import O\n",
     "\n",
     "parser = LlamaParse(\n",
-    "    result_type=\"markdown\",\n",
+    "    parse_mode=\"parse_page_with_agent\",\n",
+    "    model=\"openai-gpt-4-1-mini\",\n",
+    "    high_res_ocr=True,\n",
+    "    adaptive_long_table=True,\n",
+    "    outlined_table_extraction=True,\n",
+    "    output_tables_as_HTML=True,\n",
     ")"
    ]
   },
@@ -98,53 +109,27 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "Started parsing the file under job_id 81251f39-01be-434e-99e8-1c1b83b82098\n"
+      "Started parsing the file under job_id cd1958b0-b260-4a63-aa74-bf829a0c125f\n",
+      ".."
      ]
     }
    ],
    "source": [
-    "documents = await parser.aload_data(\"paper.pdf\")"
+    "result = await parser.aparse(\"paper.pdf\")\n",
+    "documents = result.get_markdown_documents(split_by_page=False)"
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
    "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Embeddings have been explicitly disabled. Using MockEmbedding.\n"
-     ]
-    },
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "41it [00:00, 26765.21it/s]\n",
-      "100%|██████████| 41/41 [00:13<00:00,  2.98it/s]\n"
-     ]
-    }
-   ],
+   "outputs": [],
    "source": [
-    "import nest_asyncio\n",
-    "\n",
-    "nest_asyncio.apply()\n",
-    "\n",
-    "from llama_index.core.node_parser import (\n",
-    "    MarkdownElementNodeParser,\n",
-    "    SentenceSplitter,\n",
-    ")\n",
-    "\n",
-    "# explicitly extract tables with the MarkdownElementNodeParser\n",
-    "node_parser = MarkdownElementNodeParser(num_workers=8)\n",
-    "nodes = node_parser.get_nodes_from_documents(documents)\n",
-    "nodes, objects = node_parser.get_nodes_and_objects(nodes)\n",
+    "from llama_index.core.node_parser import SentenceSplitter\n",
     "\n",
     "# Chain splitters to ensure chunk size requirements are met\n",
-    "nodes = SentenceSplitter(chunk_size=512, chunk_overlap=20).get_nodes_from_documents(\n",
-    "    nodes\n",
+    "nodes = SentenceSplitter(chunk_size=2048, chunk_overlap=256).get_nodes_from_documents(\n",
+    "    documents\n",
     ")"
    ]
   },
@@ -173,30 +158,41 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "from llama_index.agent.openai import OpenAIAgent\n",
-    "from llama_index.core.tools import QueryEngineTool, ToolMetadata\n",
-    "from llama_index.postprocessor.colbert_rerank import ColbertRerank\n",
+    "from llama_index.core.agent import FunctionAgent\n",
+    "from llama_index.core.tools import QueryEngineTool\n",
     "\n",
     "tools = [\n",
-    "    QueryEngineTool(\n",
+    "    QueryEngineTool.from_defaults(\n",
     "        vector_index.as_query_engine(\n",
-    "            similarity_top_k=8, node_postprocessors=[ColbertRerank(top_n=3)]\n",
-    "        ),\n",
-    "        metadata=ToolMetadata(\n",
-    "            name=\"search\",\n",
-    "            description=\"Search the document, pass the entire user message in the query\",\n",
+    "            similarity_top_k=4,\n",
     "        ),\n",
+    "        name=\"query\",\n",
+    "        description=\"Send a query that requires only a subset of the top-k documents to be considered\",\n",
     "    ),\n",
-    "    QueryEngineTool(\n",
+    "    QueryEngineTool.from_defaults(\n",
     "        summary_index.as_query_engine(),\n",
-    "        metadata=ToolMetadata(\n",
-    "            name=\"summarize\",\n",
-    "            description=\"Summarize the document using the user message\",\n",
-    "        ),\n",
+    "        name=\"query_all_docs\",\n",
+    "        description=\"Send a query that requires all documents to be considered\",\n",
     "    ),\n",
     "]\n",
     "\n",
-    "agent = OpenAIAgent.from_tools(tools=tools, verbose=True)"
+    "agent = FunctionAgent(\n",
+    "    tools=tools,\n",
+    "    llm=Settings.llm,\n",
+    "    system_prompt=\"You are a helpful assistant that can answer questions about the paper.\",\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from llama_index.core.workflow import Context\n",
+    "\n",
+    "# Context to persist the agent session\n",
+    "ctx = Context(agent)"
    ]
   },
   {
@@ -208,18 +204,40 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "Added user message to memory: What is the summary of the paper?\n",
-      "=== Calling Function ===\n",
-      "Calling function: summarize with args: {\"input\":\"summary\"}\n",
-      "Got output: The research focuses on developing Multimodal Large Language Models (MLLMs) by incorporating image-caption, interleaved image-text, and text-only data for pre-training. It highlights the importance of factors like the image encoder, resolution, and token count, while downplaying the design of the vision-language connector. With models scaling up to 30B parameters, the MM1 family demonstrates impressive performance in pre-training metrics and competitive outcomes on diverse multimodal benchmarks. It demonstrates abilities such as in-context learning and multi-image reasoning, aiming to provide valuable insights for creating MLLMs that benefit the research community.\n",
-      "========================\n",
-      "\n"
+      "Calling tool query_all_docs with args {'input': 'Provide the summary of the paper (concise abstract-like summary).'}\n",
+      "Tool call query_all_docs({'input': 'Provide the summary of the paper (concise abstract-like summary).'}) returned This paper presents a practical recipe and empirical analysis for building high-performing multimodal large language models (MLLMs). Through systematic ablations of image encoders, vision–language connectors, and pre-training data mixtures, the work identifies key design lessons: image resolution and the number of image tokens drive the largest gains, followed by encoder capacity and pre-training data; architectural choices for the vision–language connector matter far less. Data-wise, a careful mixture of captioned images, interleaved image–text documents, and some text-only data is critical — caption data boosts zero-shot captioning, interleaved documents enable strong few-shot and text performance, and text-only data preserves language capabilities. The authors apply these lessons to scale MM1: ViT-H image encoders at high resolution feeding 144 visual tokens into decoder-only LLMs (dense and MoE variants) trained on a 45/45/10 mixture (interleaved/caption/text), for ~200k steps (~400B tokens). MM1 models (dense up to 30B, MoE up to effectively tens of billions of parameters) achieve state-of-the-art few-shot pre-training metrics and competitive supervised fine-tuning results across many established multimodal benchmarks, while exhibiting enhanced in-context learning, multi-image reasoning, and few-shot chain-of-thought capabilities. Practical training details (learning-rate scaling, unfreezing the encoder during SFT, high-resolution support via positional interpolation and sub-image decomposition) and the positive impact of synthetic caption data are reported to guide reproducing and extending these findings.\n",
+      "\n",
+      "================\n",
+      "\n",
+      "Here is a concise, abstract‑style summary of the paper:\n",
+      "\n",
+      "- Goal: provide a practical recipe and empirical analysis for building high‑performing multimodal LLMs (MLLMs) and identify which design choices matter most.\n",
+      "- Key findings: image resolution and number of image tokens yield the largest performance gains, followed by vision‑encoder capacity and pretraining data; the specific architecture of the vision–language connector matters far less.\n",
+      "- Data mix: a careful pretraining mixture is critical—captioned images boost zero‑shot captioning, interleaved image–text documents enable strong few‑shot and text performance, and some text‑only data preserves language capabilities. The authors use a 45/45/10 split (interleaved/caption/text).\n",
+      "- MM1 models: applying these lessons, they scale ViT‑H encoders at high resolution producing 144 visual tokens into decoder‑only LLMs (dense up to 30B, MoE variants effectively larger), trained ~200k steps (~400B tokens).\n",
+      "- Results: MM1 achieves state‑of‑the‑art few‑shot pretraining metrics and competitive supervised fine‑tuning across many multimodal benchmarks, with improved in‑context learning, multi‑image reasoning, and few‑shot chain‑of‑thought behavior.\n",
+      "- Practical guidance: reportable tricks include learning‑rate scaling, unfreezing the encoder during SFT, supporting high resolution via positional interpolation and sub‑image decomposition, and the positive impact of synthetic caption data.\n",
+      "\n",
+      "Overall, the paper offers both empirical insights about what drives MLLM performance and a concrete, reproducible recipe (MM1) that attains strong multimodal capabilities.\n"
      ]
     }
    ],
    "source": [
-    "# note -- this will take a while with local LLMs, its sending every node in the document to the LLM\n",
-    "resp = agent.chat(\"What is the summary of the paper?\")"
+    "from llama_index.core.agent import ToolCall, ToolCallResult\n",
+    "\n",
+    "handler = agent.run(\n",
+    "    \"What is the summary of the paper that you have access to?\", ctx=ctx\n",
+    ")\n",
+    "async for ev in handler.stream_events():\n",
+    "    if isinstance(ev, ToolCall):\n",
+    "        print(f\"Calling tool {ev.tool_name} with args {ev.tool_kwargs}\")\n",
+    "    elif isinstance(ev, ToolCallResult):\n",
+    "        print(f\"Tool call {ev.tool_name}({ev.tool_kwargs}) returned {ev.tool_output}\")\n",
+    "\n",
+    "print(\"\\n================\\n\")\n",
+    "\n",
+    "resp = await handler\n",
+    "print(resp)"
    ]
   },
   {
@@ -231,57 +249,191 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "The summary of the paper highlights the development of Multimodal Large Language Models (MLLMs) by incorporating image-caption, interleaved image-text, and text-only data for pre-training. The research emphasizes factors like the image encoder, resolution, and token count, while de-emphasizing the design of the vision-language connector. The MM1 family of models, scaling up to 30B parameters, shows impressive performance in pre-training metrics and competitive outcomes on various multimodal benchmarks. These models demonstrate capabilities such as in-context learning and multi-image reasoning, aiming to provide valuable insights for creating MLLMs that benefit the research community.\n"
+      "Calling tool query_all_docs with args {'input': 'Describe in detail how the authors evaluate their work: which benchmarks and tasks they use (pretraining metrics, few-shot evaluation, supervised fine-tuning, multimodal benchmarks, in-context learning, chain-of-thought, multi-image reasoning), the metrics reported, baselines compared, and ablation studies conducted. Include mentions of training steps, model sizes, and any special evaluation setups (e.g., positional interpolation, sub-image decomposition, synthetic caption data).'}\n",
+      "Tool call query_all_docs({'input': 'Describe in detail how the authors evaluate their work: which benchmarks and tasks they use (pretraining metrics, few-shot evaluation, supervised fine-tuning, multimodal benchmarks, in-context learning, chain-of-thought, multi-image reasoning), the metrics reported, baselines compared, and ablation studies conducted. Include mentions of training steps, model sizes, and any special evaluation setups (e.g., positional interpolation, sub-image decomposition, synthetic caption data).'}) returned Overview\n",
+      "- Evaluation covers both pre-training (zero-/few-shot) and supervised fine-tuning (SFT) regimes, plus targeted analyses of in-context learning, multi-image reasoning, and chain-of-thought prompting. Evaluations include captioning, VQA, a set of text-only tasks (TextCore), and a wide collection of modern multimodal benchmarks. Results are reported for multiple model scales (dense 3B, 7B, 30B and MoE variants) and compared to several published baselines.\n",
+      "\n",
+      "Pre-training evaluation\n",
+      "- Tasks and benchmarks:\n",
+      "  - Image captioning: COCO (Karpathy test), NoCaps (val), TextCaps (val). Captioning use standard caption prompts and reporting.\n",
+      "  - Visual question answering / text-in-image tasks: VQAv2 (testdev), TextVQA (val), VizWiz (testdev), GQA, OK-VQA (val).\n",
+      "  - A text-only evaluation suite called TextCore (ARC, PIQA, LAMBADA, WinoGrande, HellaSWAG, SciQ, TriviaQA, WebQS) to measure preservation/quality of language capabilities.\n",
+      "- Prompting and generation:\n",
+      "  - Captioning prompt: \"{IMAGE} A photo of\" (or equivalent). VQA prompt: \"{IMAGE} Question: {QUESTION} Short answer:\".\n",
+      "  - Greedy decoding until EOS or task-specific stop tokens. For captioning the newline is a stop token; for VQA additional stop tokens include \".\", \",\", \"Question\".\n",
+      "  - VQA postprocessing follows the same logic used by OpenFlamingo implementations.\n",
+      "- Metrics:\n",
+      "  - Captioning: CIDEr (computed via nlg-eval).\n",
+      "  - VQA and related QA tasks: task-appropriate accuracy metrics (reported as percentages).\n",
+      "  - TextCore: aggregated scores reported to indicate text-only capabilities.\n",
+      "  - Pre-training few-shot evaluation reported for 0-shot, 4-shot, and 8-shot settings (4- and 8-shot used as main few-shot points).\n",
+      "- Splits and sampling:\n",
+      "  - Few-shot prompts are sampled from training when available, otherwise validation, ensuring the query example is not one of the shots.\n",
+      "- Scale and settings for pre-training evaluation runs:\n",
+      "  - Most pre-training evaluations use smaller ablation setups: base ablation LLM = 1.2B (but some encoder ablations use a 2.9B LLM to ensure capacity).\n",
+      "  - Final pre-trained models evaluated at 3B, 7B, and 30B (dense) and MoE variants (3B backbone with 64 experts; 7B backbone with 32 experts).\n",
+      "- Baselines for pre-training comparisons:\n",
+      "  - Flamingo (various sizes), Emu2 (14B, 37B), IDEFICS (9B, 80B), and other published pre-trained MLLMs where few-shot pre-training numbers are available.\n",
+      "\n",
+      "Supervised fine-tuning (SFT) evaluation\n",
+      "- SFT data and setup:\n",
+      "  - SFT mixture contains ≈1.45M examples: GPT-4/GPT-4V-generated instruction-response data (e.g., LLaVA-Conv/Complex, ShareGPT-4V), many academic VL datasets (VQAv2, GQA, OKVQA, A-OKVQA, COCO Captions, OCRVQA, TextCaps, DVQA, ChartQA, AI2D, DocVQA, InfoVQA, SynthDog-En), and a small internal text-only SFT set.\n",
+      "  - Fine-tuning: 10k steps, batch size 256, sequence length 2048; optimizer AdaFactor with peak LR 1e-5 and cosine decay to 0. Both image encoder and LLM are unfrozen unless noted in ablations.\n",
+      "- Benchmarks & aggregated evaluation:\n",
+      "  - A large set of 12+ multimodal benchmarks is used for SFT evaluation, including VQAv2, TextVQA, ScienceQA-IMG, MMMU, MathVista, MME (perception/cognition splits), MMBench, SEED-Bench, POPE, LLaVA-Bench-in-the-Wild, MM-Vet, etc.\n",
+      "  - Results reported per-dataset and combined into a meta-average for comparisons; the meta-average is normalized relative to a compact baseline to make metrics comparable across tasks.\n",
+      "- Baselines and SFT comparisons:\n",
+      "  - Compared against a range of SOTA and contemporary multimodal models after instruction tuning: LLaVA variants (1.5/NeXT), InstructBLIP, Qwen-VL, Emu2-Chat, CogVLM, Gemini family, GPT4V where available, and others. Both dense and MoE variants are compared when available.\n",
+      "- High-resolution and multi-image SFT evaluation:\n",
+      "  - Two techniques are used to support high-resolution inputs during SFT:\n",
+      "    - Positional embedding interpolation to adapt ViT positional embeddings to larger resolutions (used to support 448×448, 560×560, 672×672, etc.).\n",
+      "    - Sub-image decomposition (crop-based): for very high resolution (e.g., 1344×1344) the image is split into multiple sub-images (e.g., five 672×672 crops) that are encoded independently and concatenated as a sequence to the LLM.\n",
+      "  - Default SFT evaluation results reported at an effective high resolution (1344×1344) via these strategies. Reported improvement with higher resolution (e.g., relative gains up to ~15% average when supporting 1344×1344 vs 336×336).\n",
+      "- Chain-of-thought & few-shot in-context evaluation after SFT:\n",
+      "  - MathVista is used to quantify few-shot chain-of-thought capability: example results show 0-shot 39.4, 4-shot 41.9, and an 8-shot mixed-resolution in-context setup achieves 44.4.\n",
+      "  - Mixed-resolution in-context strategy: to fit more examples in context while managing token cost of high-resolution sub-image decomposition, some in-context examples are encoded at lower resolution and only the last N examples use full high-resolution decomposition (N=3 in reported experiments).\n",
+      "\n",
+      "Ablation studies and analyses\n",
+      "- Overall ablation design:\n",
+      "  - A compact base configuration is used for systematic ablations: ViT-L/14 image encoder (CLIP), C-Abstractor connector with 144 image tokens, pre-training mixture 45% captioned images / 45% interleaved image-text / 10% text-only, and a 1.2B decoder-only LLM for many ablations.\n",
+      "  - One component changed at a time; evaluations are zero-/few-shot across the same captioning and VQA benchmarks.\n",
+      "- Image encoder ablations:\n",
+      "  - Compared contrastive (CLIP variants trained on DFN-5B, VeCap-300M, OpenAI CLIP) against reconstructive losses (AIM models).\n",
+      "  - Resolution ablations: 224 → 336 → 378 px; clear finding that image resolution has the largest impact, followed by encoder capacity and training data composition. Increasing resolution yielded ~3% absolute boost in many metrics.\n",
+      "  - Encoder size: ViT-L → ViT-H shows modest gains (typically <1% absolute).\n",
+      "  - Training data for encoders: inclusion of synthetic caption data (VeCap) yields non-trivial few-shot improvements.\n",
+      "  - Table-based reporting of 0-/4-/8-shot metrics for these variants.\n",
+      "- Vision-language (VL) connector ablations:\n",
+      "  - Connector types: average pooling (grid pooling + linear), attention pooling (learnable queries), and C-Abstractor (convolutional mapping / ResNet-based projector).\n",
+      "  - Image token counts: experiments with 64 vs 144 image tokens per image.\n",
+      "  - Findings: number of visual tokens and image resolution matter most; the particular connector architecture has comparatively little effect on final performance. Detailed 0/4/8-shot tables compare pooling strategies across token counts and resolutions.\n",
+      "- Pre-training data mixture ablations:\n",
+      "  - Systematically varied mixes of captioned image pairs vs interleaved image-text documents vs text-only data. Examples tested: 100% caption, mixtures such as 66/33, 50/50, and 0/100, and image/text-only ratios (e.g., 91/9, 86/14, 66/33).\n",
+      "  - Key lessons:\n",
+      "    - Interleaved documents are critical for few-shot and text-only performance; captioning data strongly lifts zero-shot captioning performance.\n",
+      "    - Text-only data helps preserve/boost few-shot and text-only performance; including ~9–14% text-only yields a better balance.\n",
+      "    - A final recommended pre-training mix is 45% interleaved / 45% image-caption / 10% text-only to balance zero- and few-shot capabilities.\n",
+      "  - Impact of synthetic VeCap captions: even though small (~7% of caption pool), VeCap gives measurable few-shot gains (e.g., 2.4% and 4% absolute in reported settings).\n",
+      "- SFT-specific ablations:\n",
+      "  - Repeating data-mixture and connector ablations in the SFT context: caption-pretraining helps SFT zero-shot metrics; choice of VL connector still has limited effect though finer differences appear at high token counts; freezing vs unfreezing the image encoder matters (frozen better at lower resolution; unfrozen better for high-resolution SFT).\n",
+      "- Hyperparameter and optimization ablations:\n",
+      "  - Learning-rate grid searches run at small scales (models 9M, 85M, 302M, 1.2B) and 50k-step probes, then a log-linear fit extrapolated to larger model sizes. Grid-search experiments used 50k training steps for each setting.\n",
+      "  - Resulting scaling rule and fitted formula for optimal peak learning rate as a function of LLM parameter count is provided and used to choose LRs for the 3B/7B/30B models (e.g., final LRs used: 6e-5 (3B), 4e-5 (7B), 2e-5 (30B)). Weight decay scaled as λ = 0.1 · η.\n",
+      "- MoE (mixture-of-experts) experiments:\n",
+      "  - Two MoE designs: 3B-MoE with 64 experts (∼64B total params, top-2 gating, replace every-2 layers) and 7B-MoE with 32 experts (∼47B total params, replace every-4 layers).\n",
+      "  - Training used top-2 gating, load-balance loss coefficient 0.01, router z-loss 0.001, and otherwise the same hyperparameters and data mixture as the dense backbones. MoE variants show uniform improvements over dense counterparts on many SFT benchmarks.\n",
+      "- Additional implementation/evaluation notes:\n",
+      "  - Pre-training: models trained unfrozen for 200k steps (≈400B tokens) with batch size 512 and sequence length 4096, allowing up to 16 images per sequence and 144 tokens per image (≈1M text tokens + 1M image tokens per batch in the final setup). The pre-training mixture is fixed deterministically for reproducibility.\n",
+      "  - Pre-training evaluation prompts, stop tokens, and postprocessing are standardized (greedy decoding), and detailed splits used for each benchmark are specified.\n",
+      "  - SFT evaluation meta-average: benchmarks are normalized to a compact baseline configuration prior to averaging so disparate metrics can be compared.\n",
+      "  - For high-resolution SFT, the positional interpolation approach (to support larger patches) and the sub-image decomposition scheme (to represent very large images as multiple crops) are both used and evaluated; sub-image decomposition increases the number of image tokens dramatically, which motivates mixed-resolution in-context examples for few-shot prompting.\n",
+      "\n",
+      "Reporting and comparisons\n",
+      "- Tabular reporting:\n",
+      "  - Pre-training few-shot results are reported in detailed tables per model scale (3B, 7B, 30B) for 0/4/8/16-shot where applicable, across captioning and VQA datasets.\n",
+      "  - SFT comparisons show per-benchmark numbers and a combined meta-average; both dense and MoE model variants are included.\n",
+      "- Baselines and contemporaries cited for direct comparison include Flamingo, IDEFICS, Emu2, LLaVA-NeXT, CogVLM, Gemini family, GPT4V, and many instruction-tuned MLLMs. Where appropriate, notes on differences in prompting setups (e.g., some baselines include text-only demonstrations in “0” prompts) are documented.\n",
+      "- Qualitative analysis:\n",
+      "  - A variety of qualitative examples shown for counting, OCR, multi-image reasoning, style following, instruction following, and chain-of-thought reasoning; these accompany quantitative results to illustrate capabilities such as multi-image reasoning and few-shot chain-of-thought.\n",
+      "\n",
+      "Key reported evaluation figures (examples)\n",
+      "- Pre-training duration: 200k steps (~400B tokens).\n",
+      "- Pre-training batch & context: batch 512, sequence length 4096, up to 16 images per sequence, 144 tokens per image.\n",
+      "- SFT: 10k steps; batch 256; seq length 2048; AdaFactor with peak LR 1e-5.\n",
+      "- MoE variants: 3B backbone + 64 experts (∼64B total); 7B backbone + 32 experts (∼47B total); top-2 gating; load-balance and router regularizers used.\n",
+      "- Example few-shot chain-of-thought: MathVista 0-shot 39.4 → 4-shot 41.9 → 8-shot with mixed-resolution 44.4.\n",
+      "\n",
+      "In summary\n",
+      "- Evaluation is multi-faceted: systematic pre-training zero-/few-shot tests on captioning and VQA, text-only TextCore checks, extensive SFT across a broad benchmark suite, ablations covering image encoder, VL connector, data mixtures, training hyperparameters, and input-resolution strategies, plus experiments with MoE scaling. Metrics include CIDEr for captioning, accuracy for VQA and other benchmarks, TextCore aggregated scores, and a normalized meta-average for SFT. The authors report results across multiple model sizes and variants and compare to a broad set of recent multimodal models.\n",
+      "\n",
+      "================\n",
+      "\n",
+      "Short answer: the authors evaluate across (1) pre-training zero-/few-shot benchmarks (captioning, VQA, and a text-only suite), (2) supervised instruction fine‑tuning (SFT) on a large multimodal mixture with extensive downstream benchmarks, and (3) targeted analyses (in‑context/few‑shot learning, chain‑of‑thought, multi‑image reasoning). They report standard task metrics (CIDEr for captioning, accuracy for VQA/QA, aggregated TextCore scores, and a normalized SFT meta‑average), compare to many recent MLLMs, and run systematic ablations (encoder, connector, data mixtures, hyperparameters, resolution/tokenization, MoE). Key training/eval settings and special setups are also evaluated (positional interpolation, sub‑image decomposition, synthetic caption data). Details:\n",
+      "\n",
+      "1) Pre‑training evaluation\n",
+      "- Tasks and datasets:\n",
+      "  - Image captioning: COCO (Karpathy test), NoCaps (val), TextCaps (val).\n",
+      "  - VQA/text‑in‑image: VQAv2 (testdev), TextVQA, VizWiz, GQA, OK‑VQA, etc.\n",
+      "  - TextCore: a text‑only suite (ARC, PIQA, LAMBADA, WinoGrande, HellaSWAG, SciQ, TriviaQA, WebQS) to check language preservation.\n",
+      "- Prompting & decoding:\n",
+      "  - Zero/4/8 (and sometimes 16) shot prompts; few‑shot examples sampled from train/val ensuring no leakage.\n",
+      "  - Greedy decoding with task‑specific stop tokens; VQA postprocessing matches Flamingo style.\n",
+      "- Metrics:\n",
+      "  - CIDEr for captioning, accuracy (%) for VQA/QA tasks, aggregated TextCore scores for language capability.\n",
+      "- Model scales for evaluation:\n",
+      "  - Ablations often use a small base LLM (1.2B, sometimes 2.9B). Final pre‑trained models evaluated at 3B, 7B, 30B (dense) and MoE variants.\n",
+      "- Baselines:\n",
+      "  - Compared against Flamingo, Emu2, IDEFICS, and other published pre‑trained MLLMs when few‑shot pretraining numbers are available.\n",
+      "\n",
+      "2) Supervised fine‑tuning (SFT) evaluation\n",
+      "- SFT data:\n",
+      "  - ≈1.45M instruction examples: GPT‑4/GPT‑4V synthetic instruction data (LLaVA‑Conv/Complex, ShareGPT‑4V), many academic VL datasets (VQAv2, GQA, OKVQA, COCO Captions, TextCaps, OCRVQA, ChartQA, DocVQA, etc.), and a small internal text SFT set.\n",
+      "- Fine‑tuning procedure:\n",
+      "  - 10k steps, batch 256, seq length 2048, AdaFactor optimizer, peak LR 1e‑5 with cosine decay. Image encoder and LLM unfrozen unless ablated.\n",
+      "- Downstream benchmarks and reporting:\n",
+      "  - 12+ multimodal benchmarks for SFT evaluation (VQAv2, TextVQA, ScienceQA‑IMG, MMMU, MathVista, MME, MMBench, SEED‑Bench, POPE, LLaVA‑BiW, MM‑Vet, etc.). Results reported per dataset and combined into a normalized meta‑average for fair aggregation across heterogeneous metrics.\n",
+      "- Baselines:\n",
+      "  - Compared to instruction‑tuned contemporaries: LLaVA/NeXT, InstructBLIP, Qwen‑VL, Emu2‑Chat, CogVLM, Gemini family, GPT4V where available.\n",
+      "\n",
+      "3) Targeted analyses (in‑context learning, CoT, multi‑image)\n",
+      "- In‑context/few‑shot: standard 0/4/8‑shot probes across captioning and VQA.\n",
+      "- Chain‑of‑thought: MathVista used to quantify few‑shot CoT; reported example: 0‑shot 39.4 → 4‑shot 41.9 → 8‑shot mixed‑resolution 44.4.\n",
+      "- Multi‑image reasoning: evaluated qualitatively and quantitatively on multi‑image benchmarks and examples.\n",
+      "\n",
+      "4) Ablation studies (systematic and extensive)\n",
+      "- Image encoder ablations:\n",
+      "  - Contrastive (CLIP variants) vs reconstructive (AIM); encoder size (ViT‑L → ViT‑H); encoder training data (including synthetic caption data VeCap).\n",
+      "  - Resolution ablations (e.g., 224 → 336 → 378 px): resolution and number of visual tokens give the largest gains.\n",
+      "- Vision–language connector ablations:\n",
+      "  - Connector types (avg‑pooling, attention pooling, C‑Abstractor) and visual token counts (e.g., 64 vs 144). Finding: connector architecture matters far less than token count/resolution.\n",
+      "- Pre‑training data mixture ablations:\n",
+      "  - Varied mixes of caption pairs / interleaved image–text documents / text‑only. Key finding: 45% interleaved / 45% caption / 10% text gives the best balance (interleaved documents help few‑shot/text performance; captions boost zero‑shot captioning; text-only preserves language capabilities).\n",
+      "  - Small synthetic caption pool (VeCap) provides measurable few‑shot gains.\n",
+      "- SFT ablations:\n",
+      "  - Freezing vs unfreezing image encoder in SFT (unfreeze better for high‑resolution), data‑mix effects in SFT, connector behavior at high token counts.\n",
+      "- Hyperparameter & optimizer ablations:\n",
+      "  - LR grid searches at small scales (9M → 1.2B) with 50k‑step probes and a fitted scaling rule; final LRs chosen (e.g., ~6e‑5 for 3B, 4e‑5 for 7B, 2e‑5 for 30B for pretraining). Weight decay scaled proportionally.\n",
+      "- MoE experiments:\n",
+      "  - Two MoE setups: 3B backbone + 64 experts (~64B params) and 7B + 32 experts (~47B params), top‑2 gating, load‑balance/reg losses; MoE variants yield uniform improvements on many SFT benchmarks.\n",
+      "\n",
+      "5) Special evaluation/training setups and numbers\n",
+      "- Pretraining infrastructure & settings:\n",
+      "  - Pretraining: ≈200k steps (~400B tokens), batch 512, seq length 4096, allow up to 16 images per sequence, 144 tokens per image in final setup. Pretraining mixture fixed deterministically.\n",
+      "- High‑resolution support:\n",
+      "  - Positional embedding interpolation to adapt ViT positional embeddings to larger resolutions.\n",
+      "  - Sub‑image decomposition (split very large images into multiple crops, encode independently, and concatenate visual tokens) to support extremely high effective resolution (e.g., 1344×1344 as five 672×672 crops).\n",
+      "  - Mixed‑resolution in‑context strategy to keep context capacity reasonable while enabling high‑resolution targets in the last few shots.\n",
+      "- Decoding/postprocessing:\n",
+      "  - Greedy decoding; task‑specific stops; standardized postprocessing to align with prior work.\n",
+      "- Reporting conventions:\n",
+      "  - 0/4/8‑shot pretraining tables, SFT per‑dataset numbers and a normalized meta‑average, and qualitative examples (counting, OCR, style following, multi‑image reasoning, CoT).\n",
+      "\n",
+      "6) Qualitative analysis\n",
+      "- Numerous qualitative examples illustrating multi‑image reasoning, counting, OCR, instruction following, and chain‑of‑thought behaviors accompany the quantitative results.\n",
+      "\n",
+      "In short: the evaluation is broad (pretraining few‑shot, SFT, targeted capability probes), quantitatively rigorous (CIDEr/accuracy/meta‑averages), compares to many contemporary MLLMs, and is supported by wide ablations (encoder, connector, data, optimization, resolution, MoE) and practical high‑resolution evaluation techniques (positional interpolation, sub‑image decomposition, mixed‑resolution in‑context).\n"
      ]
     }
    ],
    "source": [
-    "print(str(resp))"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Added user message to memory: How do the authors evaluate their work?\n",
-      "=== Calling Function ===\n",
-      "Calling function: search with args: {\"input\":\"evaluation methods\"}\n",
-      "Got output: The evaluation methods involve synthesizing all benchmark results into a single meta-average number to simplify comparisons. This is achieved by normalizing the evaluation metrics with respect to a baseline configuration, standardizing the results for each task, adjusting every metric by dividing it by its respective baseline, and then averaging across all metrics.\n",
-      "========================\n",
-      "\n"
-     ]
-    }
-   ],
-   "source": [
-    "resp = agent.chat(\"How do the authors evaluate their work?\")"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "The authors evaluate their work by synthesizing all benchmark results into a single meta-average number to simplify comparisons. They normalize the evaluation metrics with respect to a baseline configuration, standardize the results for each task, adjust every metric by dividing it by its respective baseline, and then average across all metrics for evaluation.\n"
-     ]
-    }
-   ],
-   "source": [
-    "print(str(resp))"
+    "handler = agent.run(\"How do the authors evaluate their work?\", ctx=ctx)\n",
+    "async for ev in handler.stream_events():\n",
+    "    if isinstance(ev, ToolCall):\n",
+    "        print(f\"Calling tool {ev.tool_name} with args {ev.tool_kwargs}\")\n",
+    "    elif isinstance(ev, ToolCallResult):\n",
+    "        print(f\"Tool call {ev.tool_name}({ev.tool_kwargs}) returned {ev.tool_output}\")\n",
+    "\n",
+    "\n",
+    "print(\"\\n================\\n\")\n",
+    "\n",
+    "resp = await handler\n",
+    "print(resp)"
    ]
   }
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "llama-parse-aNC435Vv-py3.10",
+   "display_name": ".venv",
    "language": "python",
    "name": "python3"
   },

examples/parse/caltrain/caltrain_text_mode.ipynb

@@ -11,9 +11,10 @@
     "\n",
     "This example shows off LlamaParse parsing capabilities to build a functioning query pipeline over the Caltrain weekend schedule, a big timetable containing all trains northbound and southbound and their stops in various cities.\n",
     "\n",
-    "Naive parsing solutions mess up in representing this tabular representation, leading to LLM hallucinations. In contrast, LlamaParse text-mode spatially lays out the table in a neat format, enabling more sophisticated LLMs like gpt-4-turbo to understand the spacing and reason over all the numbers.\n",
-    "\n",
-    "**NOTE**: LlamaParse markdown mode doesn't quite work yet - it's in development!"
+    "Status:\n",
+    "| Last Executed | Version | State      |\n",
+    "|---------------|---------|------------|\n",
+    "| Aug-19-2025   | 0.6.61  | Maintained |"
    ]
   },
   {
@@ -26,18 +27,6 @@
     "Download the data."
    ]
   },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "e6ae2e38-30c9-4865-aa13-47780bc3848f",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "import nest_asyncio\n",
-    "\n",
-    "nest_asyncio.apply()"
-   ]
-  },
   {
    "cell_type": "code",
    "execution_count": null,
@@ -55,7 +44,7 @@
    "source": [
     "## Initialize LlamaParse\n",
     "\n",
-    "Initialize LlamaParse in `text` mode which will represent complex documents incl. text, tables, and figures as nicely formatted text."
+    "Parse the text results from `LlamaParse`, which will represent complex documents incl. text, tables, and figures as nicely formatted text."
    ]
   },
   {
@@ -64,26 +53,29 @@
    "id": "54aa9579-84d4-49bc-ab54-5474e69c1188",
    "metadata": {},
    "outputs": [
-    {
-     "name": "stderr",
-     "output_type": "stream",
-     "text": [
-      "/Users/jerryliu/Programming/llama_parse/.venv/lib/python3.10/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html\n",
-      "  from .autonotebook import tqdm as notebook_tqdm\n"
-     ]
-    },
     {
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "Started parsing the file under job_id 5f73353a-1f4b-480d-9eea-58d1d22b75f6\n"
+      "Started parsing the file under job_id d162724f-dcb9-4bfe-9bd4-337244906fb8\n",
+      ".."
      ]
     }
    ],
    "source": [
     "from llama_cloud_services import LlamaParse\n",
     "\n",
-    "docs = LlamaParse(result_type=\"text\").load_data(\"./caltrain_schedule_weekend.pdf\")"
+    "result = await LlamaParse(\n",
+    "    parse_mode=\"parse_page_with_agent\",\n",
+    "    model=\"openai-gpt-4-1-mini\",\n",
+    "    high_res_ocr=True,\n",
+    "    adaptive_long_table=True,\n",
+    "    outlined_table_extraction=True,\n",
+    "    output_tables_as_HTML=True,\n",
+    "    api_key=\"llx-...\",\n",
+    ").aparse(\"./caltrain_schedule_weekend.pdf\")\n",
+    "\n",
+    "documents = result.get_text_documents(split_by_page=True)"
    ]
   },
   {
@@ -104,73 +96,44 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "ZONE 2ZONE 3ZONE 4ZONE 4 ZONE 3ZONE 2ZONE 1ZONE 1\n",
-      "                                      Printer-Friendly Caltrain Schedule\n",
-      "              Northbound –                         WEEKEND SERVICE to SAN FRANCISCO                                                                                                                2XX Local\n",
+      "          Printer Friendly WEEKEND Caltrain Schedule\n",
+      "                                          Morning to Early Afternoon Page 1 of 2\n",
+      "          Northbound – WEEKEND SERVICE to SAN FRANCISCO                                                                                                         6XX Local\n",
+      "          Train No.             601  603           605  607          609  611            613       615     617       619          621  623            625  627           629  631\n",
+      "                 Tamien         6:51a             7:51a             8:51a                9:51a               10:51a              11:51a              12:51p             1:51p\n",
+      "       San Jose Diridon         6:56a    7:26a    7:56a    8:26a    8:56a     9:26a      9:56a     10:26a    10:56a    11:26a    11:56a    12:26p    12:56p    1:26p    1:56p    2:26p\n",
+      "          Santa Clara           7:03a    7:33a    8:03a    8:33a    9:03a     9:33a      10:03a    10:33a    11:03a    11:33a    12:03p    12:33p    1:03p     1:33p    2:03p    2:33p\n",
+      "ZONE 4         Lawrence         7:08a    7:38a    8:08a    8:38a    9:08a     9:38a      10:08a    10:38a    11:08a    11:38a    12:08p    12:38p    1:08p     1:38p    2:08p    2:38p\n",
       "\n",
+      "          Sunnyvale             7:12a    7:42a    8:12a    8:42a    9:12a     9:42a      10:12a    10:42a    11:12a    11:42a    12:12p    12:42p    1:12p     1:42p    2:12p    2:42p\n",
+      "               Mountain View    7:16a    7:46a    8:16a    8:46a    9:16a     9:46a      10:16a    10:46a    11:16a    11:46a    12:16p    12:46p    1:16p     1:46p    2:16p    2:46p\n",
+      "          San Antonio           7:19a    7:49a    8:19a    8:49a    9:19a     9:49a      10:19a    10:49a    11:19a    11:49a    12:19p    12:49p    1:19p     1:49p    2:19p    2:49p\n",
+      "              California Ave    7:22a    7:52a    8:22a    8:52a    9:22a     9:52a      10:22a    10:52a    11:22a    11:52a    12:22p    12:52p    1:22p     1:52p    2:22p    2:52p\n",
+      "              Palo Alto         7:25a    7:55a    8:25a    8:55a    9:25a     9:55a      10:25a    10:55a    11:25a    11:55a    12:25p    12:55p    1:25p     1:55p    2:25p    2:55p\n",
+      "ZONE 3    Menlo Park            7:27a    7:57a    8:27a    8:57a    9:27a     9:57a      10:27a    10:57a    11:27a    11:57a    12:27p    12:57p    1:27p     1:57p    2:27p    2:57p\n",
       "\n",
-      "                  Train No.       221        225        229        233        237        241        245        249        253        257        261        265        269        273       *277       *281\n",
-      "                 Service Types      L2        L2          L2        L2         L2         L2         L2         L2         L2         L2         L2         L2         L2         L2         L2         L2\n",
-      "                      Tamien      7:12a      9:05a     10:05a     11:05a                1:05p                 3:05p                 5:05p                 7:05p                 9:05p                11:05p\n",
-      "           San Jose Diridon       7:19a      9:12a     10:12a     11:12a    12:12p      1:12p      2:12p      3:12p      4:12p      5:12p      6:12p      7:12p      8:12p      9:12p     10:19p     11:12p\n",
-      "                 Santa Clara      7:25a      9:18a     10:18a     11:18a    12:18p      1:18p      2:18p      3:18p      4:18p      5:18p      6:18p      7:18p      8:18p      9:18p     10:25p     11:18p\n",
-      "                   Lawrence       7:31a      9:24a     10:24a     11:24a    12:24p      1:24p      2:24p      3:24p      4:24p      5:24p      6:24p      7:24p      8:24p      9:24p     10:31p     11:24p\n",
-      "                  Sunnyvale       7:35a      9:28a     10:28a     11:28a    12:28p      1:28p      2:28p      3:28p      4:28p      5:28p      6:28p      7:28p      8:28p      9:28p     10:35p     11:28p\n",
-      "            Mountain View         7:40a      9:34a     10:34a     11:34a    12:34p      1:34p      2:34p      3:34p      4:34p      5:34p      6:34p      7:34p      8:34p      9:34p     10:40p     11:34p\n",
-      "                San Antonio       7:43a      9:37a     10:37a     11:37a    12:37p      1:37p      2:37p      3:37p      4:37p      5:37p      6:37p      7:37p      8:37p      9:37p     10:44p     11:37p\n",
-      "              California Ave      7:48a      9:42a     10:42a     11:42a    12:42p      1:42p      2:42p      3:42p      4:42p      5:42p      6:42p      7:42p      8:42p      9:42p     10:48p     11:42p\n",
-      "                    Palo Alto     7:52a      9:46a     10:46a     11:46a    12:46p      1:46p      2:46p      3:46p      4:46p      5:46p      6:46p      7:46p      8:46p      9:46p     10:53p     11:46p\n",
-      "                 Menlo Park       7:55a      9:50a     10:50a     11:50a    12:50p      1:50p      2:50p      3:50p      4:50p      5:50p      6:50p      7:50p      8:50p      9:50p     10:56p     11:50p\n",
-      "              Redwood City        8:01a      9:56a     10:56a     11:56a    12:56p      1:56p      2:56p      3:56p      4:56p      5:56p      6:56p      7:56p      8:56p      9:56p     11:02p     11:56p\n",
-      "                  San Carlos      8:05a     10:01a     11:01a     12:01p     1:01p      2:01p      3:01p      4:01p      5:01p      6:01p      7:01p      8:01p      9:01p     10:01p     11:07p     12:01a\n",
-      "                    Belmont       8:09a     10:04a     11:04a     12:04p     1:04p      2:04p      3:04p      4:04p      5:04p      6:04p      7:04p      8:04p      9:04p     10:04p     11:10p     12:04a\n",
-      "                    Hillsdale     8:12a     10:08a     11:08a     12:08p     1:08p      2:08p      3:08p      4:08p      5:08p      6:08p      7:08p      8:08p      9:08p     10:08p     11:14p     12:08a\n",
-      "              Hayward Park        8:15a     10:11a     11:11a     12:11p     1:11p      2:11p      3:11p      4:11p      5:11p      6:11p      7:11p      8:11p      9:11p     10:11p     11:17p     12:11a\n",
-      "                  San Mateo       8:19a     10:15a     11:15a     12:15p     1:15p      2:15p      3:15p      4:15p      5:15p      6:15p      7:15p      8:15p      9:15p     10:15p     11:21p     12:15a\n",
-      "                 Burlingame       8:22a     10:19a     11:19a     12:19p     1:19p      2:19p      3:19p      4:19p      5:19p      6:19p      7:19p      8:19p      9:19p     10:19p     11:25p     12:19a\n",
-      "                  Broadway        8:25a     10:22a     11:22a     12:22p     1:22p      2:22p      3:22p      4:22p      5:22p      6:22p      7:22p      8:22p      9:22p     10:22p     11:28p     12:22a\n",
-      "                    Millbrae      8:29a     10:26a     11:26a     12:26p     1:26p      2:26p      3:26p      4:26p      5:26p      6:26p      7:26p      8:26p      9:26p     10:26p     11:32p     12:26a\n",
-      "                  San Bruno       8:34a     10:30a     11:30a     12:30p     1:30p      2:30p      3:30p      4:30p      5:30p      6:30p      7:30p      8:30p      9:30p     10:30p     11:37p     12:30a\n",
-      "            S. San Francisco      8:38a     10:34a     11:34a     12:34p     1:34p      2:34p      3:34p      4:34p      5:34p      6:34p      7:34p      8:34p      9:34p     10:34p     11:41p     12:34a\n",
-      "                   Bayshore       8:44a     10:41a     11:41a     12:41p     1:41p      2:41p      3:41p      4:41p      5:41p      6:41p      7:41p      8:41p      9:41p     10:41p     11:47p     12:41a\n",
-      "                  22 ndStreet     8:50a     10:46a     11:46a     12:46p     1:46p      2:46p      3:46p      4:46p      5:46p      6:46p      7:46p      8:46p      9:46p     10:46p     11:53p     12:46a\n",
-      "                San Francisco     8:56a     10:52a     11:53a     12:53p     1:52p      2:52p      3:52p      4:52p      5:52p      6:52p      7:52p      8:52p      9:52p     10:52p     11:59p     12:52a\n",
-      "                                      *On SAP Center event days, Train 277 or Train 281departure from San Jose Diridon station may be delayed and will depart no later than 10:30p or 11:30p respectively.\n",
+      "           Redwood City         7:32a    8:02a    8:32a    9:02a    9:32a     10:02a     10:32a    11:02a    11:32a    12:02p    12:32p    1:02p     1:32p     2:02p    2:32p    3:02p\n",
+      "             San Carlos         7:35a    8:05a    8:35a    9:05a    9:35a     10:05a     10:35a    11:05a    11:35a    12:05p    12:35p    1:05p     1:35p     2:05p    2:35p    3:05p\n",
+      "                     Belmont    7:38a    8:08a    8:38a    9:08a    9:38a     10:08a     10:38a    11:08a    11:38a    12:08p    12:38p    1:08p     1:38p     2:08p    2:38p    3:08p\n",
+      "              Hillsdale         7:41a    8:11a    8:41a    9:11a    9:41a     10:11a     10:41a    11:11a    11:41a    12:11p    12:41p    1:11p     1:41p     2:11p    2:41p    3:11p\n",
+      "           Hayward Park         7:43a    8:13a    8:43a    9:13a    9:43a     10:13a     10:43a    11:13a    11:43a    12:13p    12:43p    1:13p     1:43p     2:13p    2:43p    3:13p\n",
+      "                   San Mateo    7:46a    8:16a    8:46a    9:16a    9:46a     10:16a     10:46a    11:16a    11:46a    12:16p    12:46p    1:16p     1:46p     2:16p    2:46p    3:16p\n",
+      "          Burlingame            7:48a    8:18a    8:48a    9:18a    9:48a     10:18a     10:48a    11:18a    11:48a    12:18p    12:48p    1:18p     1:48p     2:18p    2:48p    3:18p\n",
+      "                    Broadway    7:51a    8:21a    8:51a    9:21a    9:51a     10:21a     10:51a    11:21a    11:51a    12:21p    12:51p    1:21p     1:51p     2:21p    2:51p    3:21p\n",
+      "ZONE 2    Millbrae              7:54a    8:24a    8:54a    9:24a    9:54a     10:24a     10:54a    11:24a    11:54a    12:24p    12:54p    1:24p     1:54p     2:24p    2:54p    3:24p\n",
       "\n",
+      "                   San Bruno    7:57a    8:27a    8:57a    9:27a    9:57a     10:27a     10:57a    11:27a    11:57a    12:27p    12:57p    1:27p     1:57p     2:27p    2:57p    3:27p\n",
+      "          S. San Francisco      8:00a    8:30a    9:00a    9:30a    10:00a    10:30a     11:00a    11:30a    12:00p    12:30p    1:00p     1:30p     2:00p     2:30p    3:00p    3:30p\n",
+      "          Bayshore              8:05a    8:35a    9:05a    9:35a    10:05a    10:35a     11:05a    11:35a    12:05p    12:35p    1:05p     1:35p     2:05p     2:35p    3:05p    3:35p\n",
+      "          22ⁿᵈ Street           8:10a    8:40a    9:10a    9:40a    10:10a    10:40a     11:10a    11:40a    12:10p  12:40p      1:10p     1:40p     2:10p     2:40p    3:10p    3:40p\n",
+      "ZONE 1    San Francisco         8:15a    8:45a    9:15a    9:45a    10:15a    10:45a     11:15a    11:45a    12:15p  12:45p      1:15p     1:45p     2:15p     2:45p    3:15p    3:45p\n",
       "\n",
-      "              Southbound                        –   WEEKEND               SERVICE to SAN JOSE                                                                                                      2XX Local\n",
-      "                  Train No.       224        228        232        236        240        244        248        252        256        260        264        268        272        276        280        284\n",
-      "                 Service Types      L2        L2          L2        L2          L2        L2         L2         L2         L2         L2         L2         L2         L2         L2         L2         L2\n",
-      "               San Francisco      8:28a      9:58a     10:58a     11:58a     12:58p     1:58p      2:58p      3:58p      4:58p      5:58p      6:58p      7:58p      8:58p      9:58p      10:58p     12:05a\n",
-      "                  22 ndStreet     8:33a     10:03a     11:03a     12:03p     1:03p      2:03p      3:03p      4:03p      5:03p      6:03p      7:03p      8:03p      9:03p     10:03p      11:03p     12:10a\n",
-      "                   Bayshore       8:38a     10:08a     11:08a     12:08p     1:08p      2:08p      3:08p      4:08p      5:08p      6:08p      7:08p      8:08p      9:08p     10:08p      11:08p     12:15a\n",
-      "            S. San Francisco      8:45a     10:15a     11:15a     12:15p     1:15p      2:15p      3:15p      4:15p      5:15p      6:15p      7:15p      8:15p      9:15p     10:15p      11:15p     12:22a\n",
-      "                  San Bruno       8:49a     10:19a     11:19a     12:19p     1:19p      2:19p      3:19p      4:19p      5:19p      6:19p      7:19p      8:19p      9:19p     10:19p      11:19p     12:26a\n",
-      "                    Millbrae      8:53a     10:24a     11:24a     12:24p     1:24p      2:24p      3:24p      4:24p      5:24p      6:24p      7:24p      8:24p      9:24p     10:24p      11:24p     12:31a\n",
-      "                  Broadway        8:57a     10:27a     11:27a     12:27p     1:27p      2:27p      3:27p      4:27p      5:27p      6:27p      7:27p      8:27p      9:27p     10:27p      11:27p     12:35a\n",
-      "                 Burlingame       9:00a     10:31a     11:31a     12:31p     1:31p      2:31p      3:31p      4:31p      5:31p      6:31p      7:31p      8:31p      9:31p     10:31p      11:31p     12:38a\n",
-      "                  San Mateo       9:04a     10:34a     11:34a     12:34p     1:34p      2:34p      3:34p      4:34p      5:34p      6:34p      7:34p      8:34p      9:34p     10:34p      11:34p     12:41a\n",
-      "              Hayward Park        9:07a     10:37a     11:37a     12:37p     1:37p      2:37p      3:37p      4:37p      5:37p      6:37p      7:37p      8:37p      9:37p     10:37p      11:37p     12:45a\n",
-      "                    Hillsdale     9:10a     10:41a     11:41a     12:41p     1:41p      2:41p      3:41p      4:41p      5:41p      6:41p      7:41p      8:41p      9:41p     10:41p      11:41p     12:48a\n",
-      "                    Belmont       9:14a     10:44a     11:44a     12:44p     1:44p      2:44p      3:44p      4:44p      5:44p      6:44p      7:44p      8:44p      9:44p     10:44p      11:44p     12:52a\n",
-      "                  San Carlos      9:17a     10:48a     11:48a     12:48p     1:48p      2:48p      3:48p      4:48p      5:48p      6:48p      7:48p      8:48p      9:48p     10:48p      11:48p     12:55a\n",
-      "              Redwood City        9:21a     10:52a     11:52a     12:52p     1:52p      2:52p      3:52p      4:52p      5:52p      6:52p      7:52p      8:52p      9:52p     10:52p      11:52p     12:59a\n",
-      "                 Menlo Park       9:28a     10:58a     11:58a     12:58p     1:58p      2:58p      3:58p      4:58p      5:58p      6:58p      7:58p      8:58p      9:58p     10:58p      11:58p     1:05a\n",
-      "                    Palo Alto     9:32a     11:02a     12:02p     1:02p      2:02p      3:02p      4:02p      5:02p      6:02p      7:02p      8:02p      9:02p     10:02p     11:02p      12:02a     1:09a\n",
-      "         California Avenue        9:36a     11:06a     12:06p     1:06p      2:06p      3:06p      4:06p      5:06p      6:06p      7:06p      8:06p      9:06p     10:06p     11:06p      12:06a     1:12a\n",
-      "                San Antonio       9:41a     11:11a     12:11p     1:11p      2:11p      3:11p      4:11p      5:11p      6:11p      7:11p      8:11p      9:11p     10:11p     11:11p      12:10a     1:17a\n",
-      "            Mountain View         9:45a     11:16a     12:16p     1:16p      2:16p      3:16p      4:16p      5:16p      6:16p      7:16p      8:16p      9:16p     10:16p     11:16p      12:15a     1:21a\n",
-      "                  Sunnyvale       9:51a     11:21a     12:21p     1:21p      2:21p      3:21p      4:21p      5:21p      6:21p      7:21p      8:21p      9:21p     10:21p     11:21p      12:20a     1:26a\n",
-      "                   Lawrence       9:55a     11:26a     12:26p     1:26p      2:26p      3:26p      4:26p      5:26p      6:26p      7:26p      8:26p      9:26p     10:26p     11:26p      12:25a     1:31a\n",
-      "                 Santa Clara     10:01a     11:32a     12:32p     1:32p      2:32p      3:32p      4:32p      5:32p      6:32p      7:32p      8:32p      9:32p     10:32p     11:32p      12:31a     1:37a\n",
-      "           San Jose Diridon      10:10a     11:40a     12:40p     1:38p      2:40p      3:38p      4:40p      5:38p      6:40p      7:38p      8:40p      9:38p     10:40p     11:38p      12:39a     1:44a\n",
-      "                      Tamien     10:15a     11:45a     12:45p                2:45p                 4:45p                 6:45p                 8:45p                10:45p                 12:44a     1:49a\n",
-      "       EFFECTIVE September 12, 2022                                                                                                                                  Timetable subject to change without notice.\n"
+      "EFFECTIVE September 21, 2024                                 Timetable subject to change without notice                        See Page 2 For Afternoon and Evening Times\n"
      ]
     }
    ],
    "source": [
-    "print(docs[0].get_content())"
+    "print(documents[0].text)"
    ]
   },
   {
@@ -180,9 +143,7 @@
    "source": [
     "## Initialize Query Engine\n",
     "\n",
-    "We now initialize a query engine over this data. Here we use a baseline summary index, which doesn't do vector indexing/chunking and instead dumps the entire text into the prompt.\n",
-    "\n",
-    "We see that the LLM (gpt-4-turbo) is able to provide all the stops for train no 225 northbound."
+    "We now initialize a query engine over this data. Here we use a baseline summary index, which doesn't do vector indexing/chunking and instead dumps the entire text into the prompt."
    ]
   },
   {
@@ -195,8 +156,8 @@
     "from llama_index.core import SummaryIndex\n",
     "from llama_index.llms.openai import OpenAI\n",
     "\n",
-    "llm = OpenAI(model=\"gpt-4o\")\n",
-    "index = SummaryIndex.from_documents(docs)\n",
+    "llm = OpenAI(model=\"gpt-5-mini\", api_key=\"sk-...\")\n",
+    "index = SummaryIndex.from_documents(documents)\n",
     "query_engine = index.as_query_engine(llm=llm)"
    ]
   },
@@ -208,7 +169,7 @@
    "outputs": [],
    "source": [
     "response = query_engine.query(\n",
-    "    \"What are the stops (and times) for train no 237 northbound?\"\n",
+    "    \"What are the stops (and times) for train no 609 northbound?\"\n",
     ")"
    ]
   },
@@ -222,31 +183,32 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "The stops and times for train no. 237 northbound are as follows:\n",
+      "Train No. 609 northbound (stops and times):\n",
       "\n",
-      "- San Jose Diridon: 12:12 PM\n",
-      "- Santa Clara: 12:18 PM\n",
-      "- Lawrence: 12:24 PM\n",
-      "- Sunnyvale: 12:28 PM\n",
-      "- Mountain View: 12:34 PM\n",
-      "- San Antonio: 12:37 PM\n",
-      "- California Ave: 12:42 PM\n",
-      "- Palo Alto: 12:46 PM\n",
-      "- Menlo Park: 12:50 PM\n",
-      "- Redwood City: 12:56 PM\n",
-      "- San Carlos: 1:01 PM\n",
-      "- Belmont: 1:04 PM\n",
-      "- Hillsdale: 1:08 PM\n",
-      "- Hayward Park: 1:11 PM\n",
-      "- San Mateo: 1:15 PM\n",
-      "- Burlingame: 1:19 PM\n",
-      "- Broadway: 1:22 PM\n",
-      "- Millbrae: 1:26 PM\n",
-      "- San Bruno: 1:30 PM\n",
-      "- S. San Francisco: 1:34 PM\n",
-      "- Bayshore: 1:41 PM\n",
-      "- 22nd Street: 1:46 PM\n",
-      "- San Francisco: 1:52 PM\n"
+      "- Tamien — 8:51a\n",
+      "- San Jose Diridon — 8:56a\n",
+      "- Santa Clara — 9:03a\n",
+      "- Lawrence — 9:08a\n",
+      "- Sunnyvale — 9:12a\n",
+      "- Mountain View — 9:16a\n",
+      "- San Antonio — 9:19a\n",
+      "- California Ave — 9:22a\n",
+      "- Palo Alto — 9:25a\n",
+      "- Menlo Park — 9:27a\n",
+      "- Redwood City — 9:32a\n",
+      "- San Carlos — 9:35a\n",
+      "- Belmont — 9:38a\n",
+      "- Hillsdale — 9:41a\n",
+      "- Hayward Park — 9:43a\n",
+      "- San Mateo — 9:46a\n",
+      "- Burlingame — 9:48a\n",
+      "- Broadway — 9:51a\n",
+      "- Millbrae — 9:54a\n",
+      "- San Bruno — 9:57a\n",
+      "- S. San Francisco — 10:00a\n",
+      "- Bayshore — 10:05a\n",
+      "- 22nd Street — 10:10a\n",
+      "- San Francisco — 10:15a\n"
      ]
     }
    ],
@@ -262,18 +224,10 @@
    "outputs": [],
    "source": [
     "response = query_engine.query(\n",
-    "    \"What are all the trains (and times) that end at Tamien going Southbound?\"\n",
+    "    \"What are all the trains (and times) that end at Redwood City going Southbound?\"\n",
     ")"
    ]
   },
-  {
-   "cell_type": "markdown",
-   "id": "6cf9fce0-5067-48f6-a7ef-62aa9e2edc3d",
-   "metadata": {},
-   "source": [
-    "It gets most of the answers correct (to be fair it misses two trains)."
-   ]
-  },
   {
    "cell_type": "code",
    "execution_count": null,
@@ -284,233 +238,20 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "The trains that end at Tamien going Southbound are:\n",
-      "\n",
-      "- Train 224 at 10:15a\n",
-      "- Train 228 at 11:45a\n",
-      "- Train 240 at 2:45p\n",
-      "- Train 248 at 4:45p\n",
-      "- Train 256 at 6:45p\n",
-      "- Train 264 at 8:45p\n",
-      "- Train 272 at 10:45p\n",
-      "- Train 284 at 1:49a\n"
+      "None. On this weekend schedule no southbound trains terminate at Redwood City — every listed southbound train continues beyond Redwood City to later stations (Menlo Park/Palo Alto and onward).\n"
      ]
     }
    ],
    "source": [
     "print(str(response))"
    ]
-  },
-  {
-   "cell_type": "markdown",
-   "id": "e51e7feb-b74f-4101-8963-933ac7ec9763",
-   "metadata": {},
-   "source": [
-    "## Try Baseline\n",
-    "\n",
-    "In contrast, we try a baseline approach with the default PDF reader (PyPDF) in `SimpleDirectoryReader`."
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "364e5155-cc75-4302-a754-9444ae28e6b1",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "from llama_index.core import SimpleDirectoryReader\n",
-    "from llama_index.core import SummaryIndex\n",
-    "from llama_index.llms.openai import OpenAI\n",
-    "\n",
-    "llm = OpenAI(model=\"gpt-4o\")\n",
-    "input_file = \"caltrain_schedule_weekend.pdf\"\n",
-    "reader = SimpleDirectoryReader(input_files=[input_file])\n",
-    "base_docs = reader.load_data()\n",
-    "index = SummaryIndex.from_documents(base_docs)\n",
-    "base_query_engine = index.as_query_engine(llm=llm)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "a4011389-2d27-4a1a-bf8d-7309da28ab15",
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Southbound  – WEEKEND SERVICE to SAN JOSE\n",
-      "Train No. 224 228 232 236 240 244 248 252 256 260 264 268 272 276 280 284\n",
-      "Service Types L2 L2 L2 L2 L2 L2 L2 L2 L2 L2 L2 L2 L2 L2 L2 L2\n",
-      "San Francisco 8:28a 9:58a 10:58a 11:58a 12:58p 1:58p 2:58p 3:58p 4:58p 5:58p 6:58p 7:58p 8:58p 9:58p 10:58p 12:05a\n",
-      "22nd Street 8:33a 10:03a 11:03a 12:03p 1:03p 2:03p 3:03p 4:03p 5:03p 6:03p 7:03p 8:03p 9:03p 10:03p 11:03p 12:10a\n",
-      "Bayshore 8:38a 10:08a 11:08a 12:08p 1:08p 2:08p 3:08p 4:08p 5:08p 6:08p 7:08p 8:08p 9:08p 10:08p 11:08p 12:15a\n",
-      "S. San Francisco 8:45a 10:15a 11:15a 12:15p 1:15p 2:15p 3:15p 4:15p 5:15p 6:15p 7:15p 8:15p 9:15p 10:15p 11:15p 12:22a\n",
-      "San Bruno 8:49a 10:19a 11:19a 12:19p 1:19p 2:19p 3:19p 4:19p 5:19p 6:19p 7:19p 8:19p 9:19p 10:19p 11:19p 12:26a\n",
-      "Millbrae 8:53a 10:24a 11:24a 12:24p 1:24p 2:24p 3:24p 4:24p 5:24p 6:24p 7:24p 8:24p 9:24p 10:24p 11:24p 12:31a\n",
-      "Broadway 8:57a 10:27a 11:27a 12:27p 1:27p 2:27p 3:27p 4:27p 5:27p 6:27p 7:27p 8:27p 9:27p 10:27p 11:27p 12:35a\n",
-      "Burlingame 9:00a 10:31a 11:31a 12:31p 1:31p 2:31p 3:31p 4:31p 5:31p 6:31p 7:31p 8:31p 9:31p 10:31p 11:31p 12:38a\n",
-      "San Mateo 9:04a 10:34a 11:34a 12:34p 1:34p 2:34p 3:34p 4:34p 5:34p 6:34p 7:34p 8:34p 9:34p 10:34p 11:34p 12:41a\n",
-      "Hayward Park 9:07a 10:37a 11:37a 12:37p 1:37p 2:37p 3:37p 4:37p 5:37p 6:37p 7:37p 8:37p 9:37p 10:37p 11:37p 12:45a\n",
-      "Hillsdale 9:10a 10:41a 11:41a 12:41p 1:41p 2:41p 3:41p 4:41p 5:41p 6:41p 7:41p 8:41p 9:41p 10:41p 11:41p 12:48a\n",
-      "Belmont 9:14a 10:44a 11:44a 12:44p 1:44p 2:44p 3:44p 4:44p 5:44p 6:44p 7:44p 8:44p 9:44p 10:44p 11:44p 12:52a\n",
-      "San Carlos 9:17a 10:48a 11:48a 12:48p 1:48p 2:48p 3:48p 4:48p 5:48p 6:48p 7:48p 8:48p 9:48p 10:48p 11:48p 12:55a\n",
-      "Redwood City 9:21a 10:52a 11:52a 12:52p 1:52p 2:52p 3:52p 4:52p 5:52p 6:52p 7:52p 8:52p 9:52p 10:52p 11:52p 12:59a\n",
-      "Menlo Park 9:28a 10:58a 11:58a 12:58p 1:58p 2:58p 3:58p 4:58p 5:58p 6:58p 7:58p 8:58p 9:58p 10:58p 11:58p 1:05a\n",
-      "Palo Alto 9:32a 11:02a 12:02p 1:02p 2:02p 3:02p 4:02p 5:02p 6:02p 7:02p 8:02p 9:02p 10:02p 11:02p 12:02a 1:09a\n",
-      "California Avenue 9:36a 11:06a 12:06p 1:06p 2:06p 3:06p 4:06p 5:06p 6:06p 7:06p 8:06p 9:06p 10:06p 11:06p 12:06a 1:12a\n",
-      "San Antonio 9:41a 11:11a 12:11p 1:11p 2:11p 3:11p 4:11p 5:11p 6:11p 7:11p 8:11p 9:11p 10:11p 11:11p 12:10a 1:17a\n",
-      "Mountain View 9:45a 11:16a 12:16p 1:16p 2:16p 3:16p 4:16p 5:16p 6:16p 7:16p 8:16p 9:16p 10:16p 11:16p 12:15a 1:21a\n",
-      "Sunnyvale 9:51a 11:21a 12:21p 1:21p 2:21p 3:21p 4:21p 5:21p 6:21p 7:21p 8:21p 9:21p 10:21p 11:21p 12:20a 1:26a\n",
-      "Lawrence 9:55a 11:26a 12:26p 1:26p 2:26p 3:26p 4:26p 5:26p 6:26p 7:26p 8:26p 9:26p 10:26p 11:26p 12:25a 1:31a\n",
-      "Santa Clara 10:01a 11:32a 12:32p 1:32p 2:32p 3:32p 4:32p 5:32p 6:32p 7:32p 8:32p 9:32p 10:32p 11:32p 12:31a 1:37a\n",
-      "San Jose Diridon 10:10a 11:40a 12:40p 1:38p 2:40p 3:38p 4:40p 5:38p 6:40p 7:38p 8:40p 9:38p 10:40p 11:38p 12:39a 1:44a\n",
-      "Tamien 10:15a 11:45a 12:45p 2:45p 4:45p 6:45p 8:45p 10:45p 12:44a 1:49aPrinter-Friendly Caltrain Schedule\n",
-      "Northbound  – WEEKEND SERVICE to SAN FRANCISCO\n",
-      "Train No. 221 225 229 233 237 241 245 249 253 257 261 265 269 273 *277 *281\n",
-      "Service Types L2 L2 L2 L2 L2 L2 L2 L2 L2 L2 L2 L2 L2 L2 L2 L2\n",
-      "Tamien 7:12a 9:05a 10:05a 11:05a 1:05p 3:05p 5:05p 7:05p 9:05p 11:05p\n",
-      "San Jose Diridon 7:19a 9:12a 10:12a 11:12a 12:12p 1:12p 2:12p 3:12p 4:12p 5:12p 6:12p 7:12p 8:12p 9:12p 10:19p 11:12p\n",
-      "Santa Clara 7:25a 9:18a 10:18a 11:18a 12:18p 1:18p 2:18p 3:18p 4:18p 5:18p 6:18p 7:18p 8:18p 9:18p 10:25p 11:18p\n",
-      "Lawrence 7:31a 9:24a 10:24a 11:24a 12:24p 1:24p 2:24p 3:24p 4:24p 5:24p 6:24p 7:24p 8:24p 9:24p 10:31p 11:24p\n",
-      "Sunnyvale 7:35a 9:28a 10:28a 11:28a 12:28p 1:28p 2:28p 3:28p 4:28p 5:28p 6:28p 7:28p 8:28p 9:28p 10:35p 11:28p\n",
-      "Mountain View 7:40a 9:34a 10:34a 11:34a 12:34p 1:34p 2:34p 3:34p 4:34p 5:34p 6:34p 7:34p 8:34p 9:34p 10:40p 11:34p\n",
-      "San Antonio 7:43a 9:37a 10:37a 11:37a 12:37p 1:37p 2:37p 3:37p 4:37p 5:37p 6:37p 7:37p 8:37p 9:37p 10:44p 11:37p\n",
-      "California Ave 7:48a 9:42a 10:42a 11:42a 12:42p 1:42p 2:42p 3:42p 4:42p 5:42p 6:42p 7:42p 8:42p 9:42p 10:48p 11:42p\n",
-      "Palo Alto 7:52a 9:46a 10:46a 11:46a 12:46p 1:46p 2:46p 3:46p 4:46p 5:46p 6:46p 7:46p 8:46p 9:46p 10:53p 11:46p\n",
-      "Menlo Park 7:55a 9:50a 10:50a 11:50a 12:50p 1:50p 2:50p 3:50p 4:50p 5:50p 6:50p 7:50p 8:50p 9:50p 10:56p 11:50p\n",
-      "Redwood City 8:01a 9:56a 10:56a 11:56a 12:56p 1:56p 2:56p 3:56p 4:56p 5:56p 6:56p 7:56p 8:56p 9:56p 11:02p 11:56p\n",
-      "San Carlos 8:05a 10:01a 11:01a 12:01p 1:01p 2:01p 3:01p 4:01p 5:01p 6:01p 7:01p 8:01p 9:01p 10:01p 11:07p 12:01a\n",
-      "Belmont 8:09a 10:04a 11:04a 12:04p 1:04p 2:04p 3:04p 4:04p 5:04p 6:04p 7:04p 8:04p 9:04p 10:04p 11:10p 12:04a\n",
-      "Hillsdale 8:12a 10:08a 11:08a 12:08p 1:08p 2:08p 3:08p 4:08p 5:08p 6:08p 7:08p 8:08p 9:08p 10:08p 11:14p 12:08a\n",
-      "Hayward Park 8:15a 10:11a 11:11a 12:11p 1:11p 2:11p 3:11p 4:11p 5:11p 6:11p 7:11p 8:11p 9:11p 10:11p 11:17p 12:11a\n",
-      "San Mateo 8:19a 10:15a 11:15a 12:15p 1:15p 2:15p 3:15p 4:15p 5:15p 6:15p 7:15p 8:15p 9:15p 10:15p 11:21p 12:15a\n",
-      "Burlingame 8:22a 10:19a 11:19a 12:19p 1:19p 2:19p 3:19p 4:19p 5:19p 6:19p 7:19p 8:19p 9:19p 10:19p 11:25p 12:19a\n",
-      "Broadway 8:25a 10:22a 11:22a 12:22p 1:22p 2:22p 3:22p 4:22p 5:22p 6:22p 7:22p 8:22p 9:22p 10:22p 11:28p 12:22a\n",
-      "Millbrae 8:29a 10:26a 11:26a 12:26p 1:26p 2:26p 3:26p 4:26p 5:26p 6:26p 7:26p 8:26p 9:26p 10:26p 11:32p 12:26a\n",
-      "San Bruno 8:34a 10:30a 11:30a 12:30p 1:30p 2:30p 3:30p 4:30p 5:30p 6:30p 7:30p 8:30p 9:30p 10:30p 11:37p 12:30a\n",
-      "S. San Francisco 8:38a 10:34a 11:34a 12:34p 1:34p 2:34p 3:34p 4:34p 5:34p 6:34p 7:34p 8:34p 9:34p 10:34p 11:41p 12:34a\n",
-      "Bayshore 8:44a 10:41a 11:41a 12:41p 1:41p 2:41p 3:41p 4:41p 5:41p 6:41p 7:41p 8:41p 9:41p 10:41p 11:47p 12:41a\n",
-      "22nd Street 8:50a 10:46a 11:46a 12:46p 1:46p 2:46p 3:46p 4:46p 5:46p 6:46p 7:46p 8:46p 9:46p 10:46p 11:53p 12:46a\n",
-      "San Francisco 8:56a 10:52a 11:53a 12:53p 1:52p 2:52p 3:52p 4:52p 5:52p 6:52p 7:52p 8:52p 9:52p 10:52p 11:59p 12:52aZONE 2 ZONE 3 ZONE 4 ZONE 4 ZONE 3 ZONE 2 ZONE 1 ZONE 12XX Local\n",
-      "2XX Local\n",
-      "EFFECTIVE September 12, 2022  Timetable subject to change without notice. *On SAP Center event days, Train 277 or Train 281departure from San Jose Diridon station may be delayed and will depart no later than 10:30p or 11:30p respectively.\n"
-     ]
-    }
-   ],
-   "source": [
-    "print(base_docs[0].get_content())"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "42203c70-7ca7-4200-bf47-6282eefca3bf",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "base_response = base_query_engine.query(\n",
-    "    \"What are the stops (and times) for train no 237 northbound?\"\n",
-    ")"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "06aa47b6-0f31-4b2d-90f0-bf6c74befd38",
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Train No. 237 northbound stops at the following stations and times:\n",
-      "\n",
-      "- Tamien: 1:05p\n",
-      "- San Jose Diridon: 1:12p\n",
-      "- Santa Clara: 1:18p\n",
-      "- Lawrence: 1:24p\n",
-      "- Sunnyvale: 1:28p\n",
-      "- Mountain View: 1:34p\n",
-      "- San Antonio: 1:37p\n",
-      "- California Ave: 1:42p\n",
-      "- Palo Alto: 1:46p\n",
-      "- Menlo Park: 1:50p\n",
-      "- Redwood City: 1:56p\n",
-      "- San Carlos: 2:01p\n",
-      "- Belmont: 2:04p\n",
-      "- Hillsdale: 2:08p\n",
-      "- Hayward Park: 2:11p\n",
-      "- San Mateo: 2:15p\n",
-      "- Burlingame: 2:19p\n",
-      "- Broadway: 2:22p\n",
-      "- Millbrae: 2:26p\n",
-      "- San Bruno: 2:30p\n",
-      "- S. San Francisco: 2:34p\n",
-      "- Bayshore: 2:41p\n",
-      "- 22nd Street: 2:46p\n",
-      "- San Francisco: 2:52p\n"
-     ]
-    }
-   ],
-   "source": [
-    "print(str(base_response))"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "4f3c1de7-3351-4cd8-991c-34a777952194",
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "base_response = base_query_engine.query(\n",
-    "    \"What are all the trains (and times) that end at Tamien going Southbound?\"\n",
-    ")"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "id": "513b1007-7508-4fb1-836c-de9353433a67",
-   "metadata": {},
-   "source": [
-    "Note that the trains don't line up with the times!"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "id": "108edb92-76af-406b-a139-8b9e7c6528f2",
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "The trains that end at Tamien going Southbound are:\n",
-      "\n",
-      "- Train 224 at 10:15a\n",
-      "- Train 228 at 11:45a\n",
-      "- Train 240 at 2:45p\n",
-      "- Train 252 at 4:45p\n",
-      "- Train 264 at 6:45p\n",
-      "- Train 276 at 8:45p\n",
-      "- Train 284 at 10:45p\n",
-      "- Train 284 at 12:44a\n"
-     ]
-    }
-   ],
-   "source": [
-    "print(str(base_response))"
-   ]
   }
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "llama_parse",
+   "display_name": ".venv",
    "language": "python",
-   "name": "llama_parse"
+   "name": "python3"
   },
   "language_info": {
    "codemirror_mode": {

examples/parse/demo_api.ipynb

@@ -6,32 +6,19 @@
    "source": [
     "# Using the Raw API\n",
     "\n",
-    "This notebook walks through how to use the raw API and how"
+    "This notebook walks through how to use the raw API to parse documents.\n",
+    "\n",
+    "Status:\n",
+    "| Last Executed | Version | State      |\n",
+    "|---------------|---------|------------|\n",
+    "| Aug-18-2025   | N/A     | Maintained |"
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
    "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "--2024-02-02 11:11:39--  https://arxiv.org/pdf/1706.03762.pdf\n",
-      "Resolving arxiv.org (arxiv.org)... 151.101.131.42, 151.101.3.42, 151.101.67.42, ...\n",
-      "Connecting to arxiv.org (arxiv.org)|151.101.131.42|:443... connected.\n",
-      "HTTP request sent, awaiting response... 200 OK\n",
-      "Length: 2215244 (2.1M) [application/pdf]\n",
-      "Saving to: ‘./attention.pdf’\n",
-      "\n",
-      "./attention.pdf     100%[===================>]   2.11M  --.-KB/s    in 0.08s   \n",
-      "\n",
-      "2024-02-02 11:11:39 (27.3 MB/s) - ‘./attention.pdf’ saved [2215244/2215244]\n",
-      "\n"
-     ]
-    }
-   ],
+   "outputs": [],
    "source": [
     "!wget \"https://arxiv.org/pdf/1706.03762.pdf\" -O \"./attention.pdf\""
    ]
@@ -62,15 +49,23 @@
     "with open(file_path, \"rb\") as f:\n",
     "    mime_type = mimetypes.guess_type(file_path)[0]\n",
     "    files = {\"file\": (f.name, f, mime_type)}\n",
+    "    body = {\n",
+    "        \"parse_mode\": \"parse_page_with_agent\",\n",
+    "        \"model\": \"openai-gpt-4-1-mini\",\n",
+    "        \"high_res_ocr\": True,\n",
+    "        \"adaptive_long_table\": True,\n",
+    "        \"outlined_table_extraction\": True,\n",
+    "        \"output_tables_as_HTML\": True,\n",
+    "    }\n",
     "\n",
     "    # send the request, upload the file\n",
     "    url = f\"{base_url}/upload\"\n",
-    "    response = requests.post(url, headers=headers, files=files)\n",
+    "    response = requests.post(url, headers=headers, files=files, data=body)\n",
     "\n",
     "response.raise_for_status()\n",
     "# get the job id for the result_url\n",
     "job_id = response.json()[\"id\"]\n",
-    "result_type = \"text\"  # or \"markdown\"\n",
+    "result_type = \"json\"  # or \"markdown\" or \"json\"\n",
     "result_url = f\"{base_url}/job/{job_id}/result/{result_type}\"\n",
     "\n",
     "# check for the result until its ready\n",
@@ -82,8 +77,7 @@
     "    time.sleep(2)\n",
     "\n",
     "# download the result\n",
-    "result = response.json()\n",
-    "output = result[result_type]"
+    "result = response.json()"
    ]
   },
   {
@@ -95,27 +89,94 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "                          Provided proper attribution is provided, Google hereby grants permission to\n",
-      "                          reproduce the tables and figures in this paper solely for use in journalistic or\n",
-      "                                                                   scholarly works.\n",
-      "                                                   Attention Is All You Need\n",
-      "arXiv:1706.03762v7  [cs.CL]  2 Aug 2023\n",
-      "                            Ashish Vaswani∗                Noam Shazeer∗               Niki Parmar∗             Jakob Uszkoreit∗\n",
-      "                              Google Brain                  Google Brain             Google Research            Google Research\n",
-      "                        avaswani@google.com             noam@google.com            nikip@google.com            usz@google.com\n",
-      "                                Llion Jones∗                 Aidan N. Gomez∗        †                  Łukasz Kaiser∗\n",
-      "                             Google Research                University of Toronto        \n"
+      "dict_keys(['pages', 'job_metadata'])\n"
      ]
     }
    ],
    "source": [
-    "print(output[:1000])"
+    "print(result.keys())"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "dict_keys(['page', 'text', 'md', 'images', 'charts', 'items', 'status', 'originalOrientationAngle', 'links', 'width', 'height', 'triggeredAutoMode', 'parsingMode', 'structuredData', 'noStructuredContent', 'noTextContent', 'pageHeaderMarkdown', 'pageFooterMarkdown', 'confidence'])\n"
+     ]
+    }
+   ],
+   "source": [
+    "print(result[\"pages\"][0].keys())"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "\n",
+      "Provided proper attribution is provided, Google hereby grants permission to reproduce the tables and figures in this paper solely for use in journalistic or scholarly works.\n",
+      "\n",
+      "# Attention Is All You Need\n",
+      "\n",
+      "**Ashish Vaswani***  \n",
+      "Google Brain  \n",
+      "avaswani@google.com  \n",
+      "\n",
+      "**Noam Shazeer***  \n",
+      "Google Brain  \n",
+      "noam@google.com  \n",
+      "\n",
+      "**Niki Parmar***  \n",
+      "Google Research  \n",
+      "nikip@google.com  \n",
+      "\n",
+      "**Jakob Uszkoreit***  \n",
+      "Google Research  \n",
+      "usz@google.com  \n",
+      "\n",
+      "**Llion Jones***  \n",
+      "Google Research  \n",
+      "llion@google.com  \n",
+      "\n",
+      "**Aidan N. Gomez* †**  \n",
+      "University of Toronto  \n",
+      "aidan@cs.toronto.edu  \n",
+      "\n",
+      "**Łukasz Kaiser***  \n",
+      "Google Brain  \n",
+      "lukaszkaiser@google.com  \n",
+      "\n",
+      "**Illia Polosukhin* ‡**  \n",
+      "illia.polosukhin@gmail.com  \n",
+      "\n",
+      "## Abstract\n",
+      "\n",
+      "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks that include an encoder and a decoder. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles, by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.\n",
+      "\n",
+      "----\n",
+      "\n",
+      "*Equal contribution. Listing order is random. Jakob proposed replacing RNNs with self-attention and started the effort to evaluate this idea. Ashish, with Il\n"
+     ]
+    }
+   ],
+   "source": [
+    "print(result[\"pages\"][0][\"md\"][:2000])"
    ]
   }
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "llama-parse-aNC435Vv-py3.11",
+   "display_name": ".venv",
    "language": "python",
    "name": "python3"
   },

examples/parse/demo_astradb.ipynb

@@ -1,295 +0,0 @@
-{
- "cells": [
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "# Using llama-parse with AstraDB"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "In this notebook, we show a basic RAG-style example that uses `llama-parse` to parse a PDF document, store the corresponding document into a vector store (`AstraDB`) and finally, perform some basic queries against that store. The notebook is modeled after the quick start notebooks and hence is meant as a way of getting started with `llama-parse`, backed by a vector database."
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "### Requirements"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "# First, install the required dependencies\n",
-    "%pip install --quiet llama-index llama-parse llama-index-vector-stores-astra-db llama-index-llms-openai"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "### Configuration"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "import os\n",
-    "import openai\n",
-    "\n",
-    "from getpass import getpass\n",
-    "\n",
-    "# Get all required API keys and parameters\n",
-    "llama_cloud_api_key = getpass(\"Enter your Llama Index Cloud API Key: \")\n",
-    "api_endpoint = input(\"Enter your Astra DB API Endpoint: \")\n",
-    "token = getpass(\"Enter your Astra DB Token: \")\n",
-    "namespace = (\n",
-    "    input(\"Enter your Astra DB namespace (optional, must exist on Astra): \") or None\n",
-    ")\n",
-    "openai_api_key = getpass(\"Enter your OpenAI API Key: \")\n",
-    "\n",
-    "os.environ[\"LLAMA_CLOUD_API_KEY\"] = llama_cloud_api_key\n",
-    "openai.api_key = openai_api_key"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "# llama-parse is async-first, running the sync code in a notebook requires the use of nest_asyncio\n",
-    "import nest_asyncio\n",
-    "\n",
-    "nest_asyncio.apply()"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "### Using llama-parse to parse a PDF"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Download complete.\n"
-     ]
-    }
-   ],
-   "source": [
-    "# Grab a PDF from Arxiv for indexing\n",
-    "import requests\n",
-    "\n",
-    "# The URL of the file you want to download\n",
-    "url = \"https://arxiv.org/pdf/1706.03762.pdf\"\n",
-    "# The local path where you want to save the file\n",
-    "file_path = \"./attention.pdf\"\n",
-    "\n",
-    "# Perform the HTTP request\n",
-    "response = requests.get(url)\n",
-    "\n",
-    "# Check if the request was successful\n",
-    "if response.status_code == 200:\n",
-    "    # Open the file in binary write mode and save the content\n",
-    "    with open(file_path, \"wb\") as file:\n",
-    "        file.write(response.content)\n",
-    "    print(\"Download complete.\")\n",
-    "else:\n",
-    "    print(\"Error downloading the file.\")"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "Started parsing the file under job_id ce3909a7-54cf-438b-849a-fe9a903b0c71\n"
-     ]
-    }
-   ],
-   "source": [
-    "from llama_cloud_services import LlamaParse\n",
-    "\n",
-    "documents = LlamaParse(result_type=\"text\").load_data(file_path)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [
-    {
-     "data": {
-      "text/plain": [
-       "'rmer - model architecture.\\nThe Transformer follows this overall architecture using stacked self-attention and point-wise, fully\\nconnected layers for both the encoder and decoder, shown in the left and right halves of Figure 1,\\nrespectively.\\n3.1   Encoder and Decoder Stacks\\nEncoder:     The encoder is composed of a stack of N = 6 identical layers. Each layer has two\\nsub-layers. The first is a multi-head self-attention mechanism, and the second is a simple, position-\\nwise fully connected feed-forward network. We employ a residual connection [11] around each of\\nthe two sub-layers, followed by layer normalization [1]. That is, the output of each sub-layer is\\nLayerNorm(x + Sublayer(x)), where Sublayer(x) is the function implemented by the sub-layer\\nitself. To facilitate these residual connections, all sub-layers in the model, as well as the embedding\\nlayers, produce outputs of dimension dmodel = 512.\\nDecoder:    The decoder is also composed of a stack of N = 6 identical layers. In addition '"
-      ]
-     },
-     "execution_count": null,
-     "metadata": {},
-     "output_type": "execute_result"
-    }
-   ],
-   "source": [
-    "# Take a quick look at some of the parsed text from the document:\n",
-    "documents[0].get_content()[10000:11000]"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "### Storing into Astra DB"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "from llama_index.vector_stores.astra_db import AstraDBVectorStore\n",
-    "\n",
-    "astra_db_store = AstraDBVectorStore(\n",
-    "    token=token,\n",
-    "    api_endpoint=api_endpoint,\n",
-    "    namespace=namespace,\n",
-    "    collection_name=\"astra_v_table_llamaparse\",\n",
-    "    embedding_dimension=1536,\n",
-    ")"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "from llama_index.core.node_parser import SimpleNodeParser\n",
-    "\n",
-    "node_parser = SimpleNodeParser()\n",
-    "\n",
-    "nodes = node_parser.get_nodes_from_documents(documents)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "from llama_index.embeddings.openai import OpenAIEmbedding\n",
-    "from llama_index.core import VectorStoreIndex, StorageContext\n",
-    "\n",
-    "storage_context = StorageContext.from_defaults(vector_store=astra_db_store)\n",
-    "\n",
-    "index = VectorStoreIndex(\n",
-    "    nodes=nodes,\n",
-    "    storage_context=storage_context,\n",
-    "    embed_model=OpenAIEmbedding(api_key=openai_api_key),\n",
-    ")"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "### Simple RAG Example"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "query_engine = index.as_query_engine(similarity_top_k=15)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "\n",
-      "***********New LlamaParse+ Basic Query Engine***********\n",
-      "Multi-Head Attention is also known as multi-headed self-attention.\n"
-     ]
-    }
-   ],
-   "source": [
-    "query = \"What is Multi-Head Attention also known as?\"\n",
-    "\n",
-    "response_1 = query_engine.query(query)\n",
-    "print(\"\\n***********New LlamaParse+ Basic Query Engine***********\")\n",
-    "print(response_1)"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [
-    {
-     "data": {
-      "text/plain": [
-       "'We used beam search as described in the previous section, but no\\ncheckpoint averaging. We present these results in Table 3.\\nIn Table 3 rows (A), we vary the number of attention heads and the attention key and value dimensions,\\nkeeping the amount of computation constant, as described in Section 3.2.2. While single-head\\nattention is 0.9 BLEU worse than the best setting, quality also drops off with too many heads.\\nIn Table 3 rows (B), we observe that reducing the attention key size dk hurts model quality. This\\nsuggests that determining compatibility is not easy and that a more sophisticated compatibility\\nfunction than dot product may be beneficial. We further observe in rows (C) and (D) that, as expected,\\nbigger models are better, and dropout is very helpful in avoiding over-fitting. In row (E) we replace our\\nsinusoidal positional encoding with learned positional embeddings [9], and observe nearly identical\\nresults to the base model.\\n6.3    English Constituency Parsing\\nTo evaluate if the Transformer can generalize to other tasks we performed experiments on English\\nconstituency parsing. This task presents specific challenges: the output is subject to strong structural\\nconstraints and is significantly longer than the input. Furthermore, RNN sequence-to-sequence\\nmodels have not been able to attain state-of-the-art results in small-data regimes [37].\\nWe trained a 4-layer transformer with dmodel = 1024 on the Wall Street Journal (WSJ) portion of the\\nPenn Treebank [25], about 40K training sentences. We also trained it in a semi-supervised setting,\\nusing the larger high-confidence and BerkleyParser corpora from with approximately 17M sentences\\n[37]. We used a vocabulary of 16K tokens for the WSJ only setting and a vocabulary of 32K tokens\\nfor the semi-supervised setting.\\nWe performed only a small number of experiments to select the dropout, both attention and residual\\n(section 5.4), learning rates and beam size on the Section 22 development set, all other parameters\\nremained unchanged from the English-to-German base translation model. During inference, we\\n                                                             9\\n---\\nTable 4: The Transformer generalizes well to English constituency parsing (Results are on Section 23\\nof WSJ)\\n                          Parser                           Training             WSJ 23 F1\\n            Vinyals & Kaiser el al. (2014) [37]    WSJ only, discriminative         88.3\\n                 Petrov et al. (2006) [29]         WSJ only, discriminative         90.4\\n                   Zhu et al. (2013) [40]          WSJ only, discriminative         90.4\\n                   Dyer et al. (2016) [8]          WSJ only, discriminative         91.7\\n                  Transformer (4 layers)           WSJ only, discriminative         91.3\\n                   Zhu et al. (2013) [40]              semi-supervised              91.3\\n               Huang & Harper (2009) [14]              semi-supervised              91.3\\n               McClosky et al. (2006) [26]             semi-supervised              92.1\\n            Vinyals & Kaiser el al. (2014) [37]        semi-supervised              92.1\\n                  Transformer (4 layers)               semi-supervised              92.7\\n                 Luong et al. (2015) [23]                  multi-task               93.0\\n                   Dyer et al. (2016) [8]                  generative               93.3\\nincreased the maximum output length to input length + 300. We used a beam size of 21 and α = 0.3\\nfor both WSJ only and the semi-supervised setting.\\nOur results in Table 4 show that despite the lack of task-specific tuning our model performs sur-\\nprisingly well, yielding better results than all previously reported models with the exception of the\\nRecurrent Neural Network Grammar [8].\\nIn contrast to RNN sequence-to-sequence models [37], the Transformer outperforms the Berkeley-\\nParser [29] even when training only on the WSJ training set of 40K sentences.\\n7    Conclusion\\nIn this work, we presented the Transformer, the first sequence transduction model based entirely on\\nattention, replacing the recurrent layers most commonly used in encoder-decoder architectures with\\nmulti-headed self-attention.\\nFor translation tasks, the Transformer can be trained significantly faster than architectures based\\non recurrent or convolutional layers.'"
-      ]
-     },
-     "execution_count": null,
-     "metadata": {},
-     "output_type": "execute_result"
-    }
-   ],
-   "source": [
-    "# Take a look at one of the source nodes from the response\n",
-    "response_1.source_nodes[0].get_content()"
-   ]
-  }
- ],
- "metadata": {
-  "kernelspec": {
-   "display_name": "Python 3 (ipykernel)",
-   "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"
-  }
- },
- "nbformat": 4,
- "nbformat_minor": 4
-}

examples/parse/demo_basic.ipynb

@@ -4,7 +4,14 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "# LlamaParse Usage"
+    "# LlamaParse Usage\n",
+    "\n",
+    "This notebook walks through the basic usage of LlamaParse.\n",
+    "\n",
+    "Status:\n",
+    "| Last Executed | Version | State      |\n",
+    "|---------------|---------|------------|\n",
+    "| Aug-18-2025   | 0.6.61  | Maintained |"
    ]
   },
   {
@@ -13,32 +20,14 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "%pip install llama-index llama-parse"
+    "%pip install \"llama-index>=0.13.2<0.14.0\" llama-cloud-services"
    ]
   },
   {
    "cell_type": "code",
    "execution_count": null,
    "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "--2024-02-02 11:10:10--  https://arxiv.org/pdf/1706.03762.pdf\n",
-      "Resolving arxiv.org (arxiv.org)... 151.101.131.42, 151.101.3.42, 151.101.67.42, ...\n",
-      "Connecting to arxiv.org (arxiv.org)|151.101.131.42|:443... connected.\n",
-      "HTTP request sent, awaiting response... 200 OK\n",
-      "Length: 2215244 (2.1M) [application/pdf]\n",
-      "Saving to: ‘./attention.pdf’\n",
-      "\n",
-      "./attention.pdf     100%[===================>]   2.11M  --.-KB/s    in 0.08s   \n",
-      "\n",
-      "2024-02-02 11:10:10 (25.9 MB/s) - ‘./attention.pdf’ saved [2215244/2215244]\n",
-      "\n"
-     ]
-    }
-   ],
+   "outputs": [],
    "source": [
     "!wget \"https://arxiv.org/pdf/1706.03762.pdf\" -O \"./attention.pdf\""
    ]
@@ -49,11 +38,6 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# llama-parse is async-first, running the sync code in a notebook requires the use of nest_asyncio\n",
-    "import nest_asyncio\n",
-    "\n",
-    "nest_asyncio.apply()\n",
-    "\n",
     "import os\n",
     "\n",
     "os.environ[\"LLAMA_CLOUD_API_KEY\"] = \"llx-...\""
@@ -68,14 +52,21 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "Started parsing the file under job_id dd0b8e31-0c09-4497-b78a-cc1c92f1d6cf\n"
+      "Started parsing the file under job_id ebc7e76e-addb-429b-8666-bee9c5832a84\n"
      ]
     }
    ],
    "source": [
     "from llama_cloud_services import LlamaParse\n",
     "\n",
-    "documents = LlamaParse(result_type=\"text\").load_data(\"./attention.pdf\")"
+    "result = await LlamaParse(\n",
+    "    parse_mode=\"parse_page_with_agent\",\n",
+    "    model=\"openai-gpt-4-1-mini\",\n",
+    "    high_res_ocr=True,\n",
+    "    adaptive_long_table=True,\n",
+    "    outlined_table_extraction=True,\n",
+    "    output_tables_as_HTML=True,\n",
+    ").aparse(\"./attention.pdf\")"
    ]
   },
   {
@@ -87,23 +78,68 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "ad\n",
+      "1 Introduction\n",
+      "\n",
+      "Recurrent neural networks, long short-term memory [13] and gated recurrent [7] neural networks\n",
+      "in particular, have been firmly established as state of the art approaches in sequence modeling and\n",
+      "transduction problems such as language modeling and machine translation [35, 2, 5]. Numerous\n",
+      "efforts have since continued to push the boundaries of recurrent language models and encoder-decoder\n",
+      "architectures [38, 24, 15].\n",
+      "Recurrent models typically factor computation along the symbol positions of the input and output\n",
+      "sequences. Aligning the positions to steps in computation time, they generate a sequence of hidden\n",
+      "states ht, as a function of the previous hidden state ht−1 and the input for position t. This inherently\n",
+      "sequential nature precludes parallelization within training examples, which becomes critical at longer\n",
+      "sequence lengths, as memory constraints limit batching across examples. Recent work has achieved\n",
+      "significant improvements in computational efficiency through factorization tricks [21] and conditional\n",
+      "computation [32], while also improving model performance in case of the latter. The fundamental\n",
+      "constraint of sequential computation, however, remains.\n",
+      "Attention mechanisms have become an integral part of compelling sequence modeling and transduc-\n",
+      "tion models in various tasks, allowing modeling of dependencies without regard to their distance in\n",
+      "the input or output sequences [2, 19]. In all but a few cases [27], however, such attention mechanisms\n",
+      "are used in conjunction with a recurrent network.\n",
+      "In this work we propose the Transformer, a model architecture eschewing recurrence and instead\n",
       "relying entirely on an attention mechanism to draw global dependencies between input and output.\n",
       "The Transformer allows for significantly more parallelization and can reach a new state of the art in\n",
       "translation quality after being trained for as little as twelve hours on eight P100 GPUs.\n",
-      "2    Background\n",
+      "\n",
+      "2 Background\n",
+      "\n",
       "The goal of reducing sequential computation also forms the foundation of the Extended Neural GPU\n",
       "[16], ByteNet [18] and ConvS2S [9], all of which use convolutional neural networks as basic building\n",
       "block, computing hidden representations in parallel for all input and output positions. In these models,\n",
       "the number of operations required to relate signals from two arbitrary input or output positions grows\n",
       "in the distance between positions, linearly for ConvS2S and logarithmically for ByteNet. This makes\n",
       "it more difficult to learn dependencies between distant positions [12]. In the Transformer this is\n",
-      "reduced to a constant number of operations, albeit at the cost of reduced effective res\n"
+      "reduced to a constant number of operations, albeit at the cost of reduced effective resolution due\n",
+      "to averaging attention-weighted positions, an effect we counteract with Multi-Head Attention as\n",
+      "described in section 3.2.\n",
+      "Self-attention, sometimes called intra-attention is an attention mechanism relating different positions\n",
+      "of a single sequence in order to compute a representation of the sequence. Self-attention has been\n",
+      "used successfully in a variety of tasks including reading comprehension, abstractive summarization,\n",
+      "textual entailment and learning task-independent sentence representations [4, 27, 28, 22].\n",
+      "End-to-end memory networks are based on a recurrent attention mechanism instead of sequence-\n",
+      "aligned recurrence and have been shown to perform well on simple-language question answering and\n",
+      "language modeling tasks [34].\n",
+      "To the best of our knowledge, however, the Transformer is the first transduction model relying\n",
+      "entirely on self-attention to compute representations of its input and output without using sequence-\n",
+      "aligned RNNs or convolution. In the following sections, we will describe the Transformer, motivate\n",
+      "self-attention and discuss its advantages over models such as [17, 18] and [9].\n",
+      "\n",
+      "3 Model Architecture\n",
+      "\n",
+      "Most competitive neural sequence transduction models have an encoder-decoder structure [5, 2, 35].\n",
+      "Here, the encoder maps an input sequence of symbol representations (x1, ..., xn) to a sequence\n",
+      "of continuous representations z = (z1, ..., zn). Given z, the decoder then generates an output\n",
+      "sequence (y1, ..., ym) of symbols one element at a time. At each step the model is auto-regressive\n",
+      "[10], consuming the previously generated symbols as additional input when generating the next.\n",
+      "\n",
+      "                                                 2\n"
      ]
     }
    ],
    "source": [
-    "print(documents[0].text[6000:7000])"
+    "documents = result.get_text_documents(split_by_page=True)\n",
+    "print(documents[1].text)"
    ]
   },
   {
@@ -115,54 +151,66 @@
      "name": "stdout",
      "output_type": "stream",
      "text": [
-      "Started parsing the file under job_id d4531453-1bbb-48c4-8324-ae9fea9f2fa2\n"
+      "\n",
+      "Provided proper attribution is provided, Google hereby grants permission to reproduce the tables and figures in this paper solely for use in journalistic or scholarly works.\n",
+      "\n",
+      "# Attention Is All You Need\n",
+      "\n",
+      "**Ashish Vaswani***  \n",
+      "Google Brain  \n",
+      "avaswani@google.com  \n",
+      "\n",
+      "**Noam Shazeer***  \n",
+      "Google Brain  \n",
+      "noam@google.com  \n",
+      "\n",
+      "**Niki Parmar***  \n",
+      "Google Research  \n",
+      "nikip@google.com  \n",
+      "\n",
+      "**Jakob Uszkoreit***  \n",
+      "Google Research  \n",
+      "usz@google.com  \n",
+      "\n",
+      "**Llion Jones***  \n",
+      "Google Research  \n",
+      "llion@google.com  \n",
+      "\n",
+      "**Aidan N. Gomez* †**  \n",
+      "University of Toronto  \n",
+      "aidan@cs.toronto.edu  \n",
+      "\n",
+      "**Łukasz Kaiser***  \n",
+      "Google Brain  \n",
+      "lukaszkaiser@google.com  \n",
+      "\n",
+      "**Illia Polosukhin* ‡**  \n",
+      "illia.polosukhin@gmail.com  \n",
+      "\n",
+      "## Abstract\n",
+      "\n",
+      "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks that include an encoder and a decoder. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English-to-German translation task, improving over the existing best results, including ensembles, by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.8 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data.\n",
+      "\n",
+      "----\n",
+      "\n",
+      "*Equal contribution. Listing order is random. Jakob proposed replacing RNNs with self-attention and started the effort to evaluate this idea. Ashish, with Illia, designed and implemented the first Transformer models and has been crucially involved in every aspect of this work. Noam proposed scaled dot-product attention, multi-head attention and the parameter-free position representation and became the other person involved in nearly every detail. Niki designed, implemented, tuned and evaluated countless model variants in our original codebase and tensor2tensor. Llion also experimented with novel model variants, was responsible for our initial codebase, and efficient inference and visualizations. Lukasz and Aidan spent countless long days designing various parts of and implementing tensor2tensor, replacing our earlier codebase, greatly improving results and massively accelerating our research.  \n",
+      "† Work performed while at Google Brain.  \n",
+      "‡ Work performed while at Google Research.\n",
+      "\n",
+      "31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA.\n",
+      "\n"
      ]
     }
    ],
    "source": [
-    "from llama_cloud_services import LlamaParse\n",
-    "\n",
-    "documents = LlamaParse(result_type=\"markdown\").load_data(\"./attention.pdf\")"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [
-    {
-     "name": "stdout",
-     "output_type": "stream",
-     "text": [
-      "ction describes the training regime for our models.\n",
-      "\n",
-      "##### Training Data and Batching\n",
-      "\n",
-      "We trained on the standard WMT 2014 English-German dataset consisting of about 4.5 million\n",
-      "sentence pairs. Sentences were encoded using byte-pair encoding [3], which has a shared source-\n",
-      "target vocabulary of about 37000 tokens. For English-French, we used the significantly larger WMT\n",
-      "2014 English-French dataset consisting of 36M sentences and split tokens into a 32000 word-piece\n",
-      "vocabulary [38]. Sentence pairs were batched together by approximate sequence length. Each training\n",
-      "batch contained a set of sentence pairs containing approximately 25000 source tokens and 25000\n",
-      "target tokens.\n",
-      "\n",
-      "##### Hardware and Schedule\n",
-      "\n",
-      "We trained our models on one machine with 8 NVIDIA P100 GPUs. For our base models using\n",
-      "the hyperparameters described throughout the paper, each training step took about 0.4 seconds. We\n",
-      "trained the base models for a total of 100,000 steps or 12 hours. For our big models,(described on the\n",
-      "bo...\n"
-     ]
-    }
-   ],
-   "source": [
-    "print(documents[0].text[20000:21000] + \"...\")"
+    "documents = result.get_markdown_documents(split_by_page=True)\n",
+    "print(documents[0].text)"
    ]
   }
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "Python 3 (ipykernel)",
+   "display_name": ".venv",
    "language": "python",
    "name": "python3"
   },