Poe Knowledge Bot with LlamaIndex

A knowledge-augmented Poe bot powered by LlamaIndex and FastAPI.

Easily ingest and chat with your own data as a knowledge base!

Quick Start

Follow these steps to quickly setup and run the LlamaIndex bot for Poe:

Setup Environment

Install poetry: pip install poetry
Install app dependencies: poetry install
Setup environment variables

Name	Required	Description
`POE_API_KEY`	Yes	This is a secret token that you need to authenticate Poe requests to the API. You can generate this from https://poe.com/create_bot?api=1.
`OPENAI_API_KEY`	Yes	This is your OpenAI API key that LlamaIndex needs to call OpenAI services. You can get an API key by creating an account on OpenAI.

Run API Server

Run the API locally: poetry run start

INFO:poe_api.llama_handler:Creating new index
INFO:poe_api.llama_handler:Loading data from data/
INFO:llama_index.token_counter.token_counter:> [insert] Total LLM token usage: 0 tokens
INFO:llama_index.token_counter.token_counter:> [insert] Total embedding token usage: 19274 tokens
2023-04-17 15:24:05,159 - INFO - Application startup complete.

Make the API publicly available with ngrok: in a different terminal, run ngrok http 8080

Connect Poe to your Bot

Create your bot at https://poe.com/create_bot?api=1
Interact with your bot at https://poe.com/

Test Your LlamaIndex Bot

To quickly verify if your bot is up and running, go to the Swagger UI at http://localhost:8080/docs, authenticate with your POE_API_KEY and issue a query (satisfying the Poe Protocol)

Alternatively, to use a sample query, replace <add your POE_API_KEY here> in Makefile with your POE_API_KEY, then run:

make try

Customize Your LlamaIndex Bot

By default, we ingest documents under data/ and index them with a GPTSimpleVectorIndex.

You can configure the default behavior via environment variables:

Name	Required	Description
`LLAMA_LOAD_DATA`	Optional	Whether to ingest documents in `DATA_DIR`.Defaults to `True`
`LLAMA_DATA_DIR`	Optional	Directory to ingest initial documents from. Defaults to `data/`
`LLAMA_INDEX_TYPE`	Optional	Index type (see below for details). Defaults to `simple_dict`
`INDEX_JSON_PATH`	Optional	Path to saved Index json file. `save/index.json`

Different Index Types By default, we use a GPTSimpleVectorIndex to store document chunks in memory, and retrieve top-k nodes by embedding similarity. Different index types are optimized for different data and query use-cases. See this guide on How Each Index Works to learn more. You can configure the index type via the LLAMA_INDEX_TYPE, see here for the full list of accepted index type identifiers.

Read more details on readthedocs, and engage with the community on discord.

Ingesting Data

LlamaIndex bot for Poe also exposes an API for ingesting additional data by POST to /add_document endpoint.

You can use the Swagger UI to quickly experiment with ingesting additional documents:

Locally: http://localhost:8080/docs
Publiclly via ngrok: https://<instance-id>.ngrok-free.app/docs