PyHC-Chat Prototype

Explore and ask questions about the Python in Heliophysics Community and its core packages using OpenAI's GPT-4 language model.

Demo

Prerequisites

• OpenAI API key, set in the environment variable OPENAI_API_KEY
• Activeloop token, set in the environment variable ACTIVELOOP_TOKEN (get one here)

Usage

1. Set the environment variables OPENAI_API_KEY and ACTIVELOOP_TOKEN
2. (Optional) Set your deeplake_username in config.py if you want your vector store online instead of local
3. Run the script: pyhc_chat.py
4. Ask your questions! Type exit() to quit

Key Features

• Has up-to-date knowledge of PyHC and its core packages, facilitated by context retrieval from a DeepLake vector store (this is why an Activeloop token is required)
• Generates detailed answers to user queries based on package repositories' contents
• Spawns helper bots to determine which repos are relevant to the user's prompts and what information should be retrieved from the vector store
• Vector store can be either online or local to your machine
• Uses OpenAI's language model for generating responses
• Optional verbose mode to display intermediate model reasoning before responses

Caveats

• Monitor your OpenAI API usage closely when using the GPT-4 model because it's pretty expensive. It's not hard to rack up a few dollars in usage after just a few conversations.
• When using an online vector store, startup time can take about a minute due to network delays.
• Likewise, responses can be slow because of delays in both querying OpenAI's API and retrieving from the vector store, especially when the helper bots are doing lots of heavy lifting.
• Vector store retrieval doesn't always get the necessary context (it'll tell you when this happens).
• Chat history length is not intelligently managed yet (that's a TODO), so messages can sometimes exceed the model's max token limit which results in errors.
• This PyHC-Chat prototype is only designed to discuss PyHC itself and the seven core packages.
  • GPT-4 has (outdated) knowledge of other PyHC packages baked into its training data (ask it which!), so it may answer some questions about other packages, but results will vary.