VIMbot

VIMbot is a FastAPI-based web application providing an interactive conversational interface for querying and retrieving information from a medical document knowledge base. The knowledge base is built using mkdocs, so this code is easily adaptable to any markdown-based knowledge repository.

Features

Interactive conversational interface for querying a medical knowledge base.
Integration with Ragatouille for document indexing and retrieval.
FastAPI backend for handling HTTP requests and responses.
Jinja2 templates for rendering HTML pages.
HTMX for dynamic interactions, along with supplementary JavaScript.
Tailwind CSS for styling, with additional CSS as needed.
SQLite databases for storing documents, embeddings, and logs.
Docker Compose configuration for easy deployment.

Installation

Clone the repository:

git clone https://github.com/your-username/vimbot.git
cd vimbot

Install the required dependencies using Poetry:
```
poetry install
```

Set up the necessary environment variables in a .env file:

FROM_EMAIL=your-email@example.com
SOURCE_URL=https://your-source-url.com
POSTMARK_API_KEY=your-postmark-api-key
OPENAI_API_KEY=sk-gibberish
ANTHROPIC_API_KEY=sk-mumbo-jumbo
GROQ_API_KEY=gsk_babba_boo

Run the application using Docker Compose:
```
docker-compose up -d
```
Access the application in your web browser at http://localhost:8092.

Usage

Enter your query in the input field on the home page and press Enter or click the Submit button.
The application processes your query, retrieves relevant information from the knowledge base, and displays the response.
Continue the conversation or start a new one.
Provide feedback on the response by clicking the appropriate buttons.
The application includes dropdowns for disclaimers and explanations on how it works—modify these to suit your needs.

Philosophy

strike a balance between off-the-shelf tools (e.g. Ragatouille) and custom-made solutions (e.g. htmx + FastAPI >> streamlit)
for most LLM functionality, prefer base libraries or light derivations over large frameworks
modular API and backend, locality-of-control frontend
pluggable
don't take the philosophy too seriously

Contributing

Contributions are welcome! If you find any issues or have suggestions for improvements, please open an issue or submit a pull request.

License

This project is licensed under the MIT License.

Acknowledgements

Ragatouille
- and Simon Willison's writeup
ColBERT
FastAPI
HTMX
Tailwind CSS

Resources/To-Do

Implement a separate server for the embeddings, a la https://github.com/bclavie/RAGatouille/pull/92/files (will probably speed things up a bit).
Switch over to structured generation, probably using Outlines, a la this HuggingFace cookbook (also may speed things up, and get more metadata goodness)
Consider adding a semantic cache
sqlite streaming backups to S3-compatible object storage using litestream

cbeauhilton/vimbot