RAG Query API

This is a FastAPI-based backend service that powers a Retrieval-Augmented Generation (RAG) pipeline using a local SQLite knowledge base and Pinecone vector search.

Features

• Query a multimodal (text + image) knowledge base via /query endpoint
• Uses Pinecone for vector similarity search
• GPT-4o-mini model via AIProxy (supports vision for image analysis)
• Auto-generated embeddings using OpenAI-compatible embedding model
• Returns contextual answers along with source links
• Handles rate-limiting and retries on failures
• Health-check endpoint at /health

Requirements

• Python 3.8+
• Environment Variables:
  • API_KEY (for AIProxy OpenAI-compatible API)
  • PINECONE_API_KEY (your Pinecone API key)
  • PINECONE_ENV (Pinecone environment, e.g., us-east-1-aws)

Setup

1. Clone the repository
2. Create and activate a virtual environment
3. Install dependencies:

   pip install -r requirements.txt

4. Create a .env file with the required keys:

   API_KEY=your_aipipe_api_key
   PINECONE_API_KEY=your_pinecone_key
   PINECONE_ENV=us-east-1-aws
   

Running the API

uvicorn app:app --reload --port 8000

Endpoints

POST /query

Query the knowledge base with text and optionally an image (base64).

Request:

{
  "question": "Explain cosine similarity",
  "image": "base64_string_if_any"
}

Response:

{
  "data": {
    "answer": "Cosine similarity measures ...",
    "links": [
      {"url": "https://example.com", "text": "Relevant source"}
    ]
  }
}

GET /health

Checks DB connectivity and whether embeddings exist.

Database Schema

• discourse_chunks: stores forum chunks with embeddings
• markdown_chunks: stores markdown doc chunks with embeddings

Notes

• Handles rate limits and retries gracefully
• Image support via GPT-4o multimodal model
• Logs important events and errors with full tracebacks

License

MIT