RAG System

A Retrieval-Augmented Generation (RAG) system for document-based question answering.

Quick Start

Prerequisites

• Python 3.12+
• uv package manager
• Google API key (set in .env)

Installation

# Set up environment
export GOOGLE_API_KEY=your_api_key_here

Running the System

uv run main_api.py

Then open your browser and go to:

• Web Frontend: http://localhost:8000
• Swagger API Docs: http://localhost:8000/docs

Web Frontend

The system includes a modern, professional web interface for testing:

1. Upload Documents - Drag & drop or click to upload PDFs
2. View Documents - See all uploaded documents with metadata
3. Ask Questions - Query your documents in natural language
4. View Results - Get answers with source documents highlighted
5. Manage Documents - Delete documents as needed
6. System Status - Monitor server health and database status

Frontend Features

• 📤 Easy Upload - Drag-and-drop PDF upload with progress tracking
• ❓ Smart Queries - Natural language question answering
• 📚 Source Citations - See which documents your answers came from
• 🎯 Status Monitoring - Real-time system health and document count
• 🎨 Modern UI - Professional, responsive design
• ⚡ Instant Feedback - Real-time status updates

Architecture

Core Components

• Ingestion (ingestion_pipeline/)

  • Document loading and chunking
  • Vector database management

• Retrieval (retriever/)

  • Multi-query expansion
  • Hybrid search (vector + BM25)
  • Cross-encoder reranking

• Generation (generation/)

  • LLM answer generation
  • RAG pipeline coordination

• Backend (backend/)

  • FastAPI routes and schemas
  • Document upload management

• Frontend (frontend/)

  • Single-page HTML5 application
  • Real-time WebSocket-like updates
  • Responsive design with embedded CSS and JavaScript

Configuration

Edit config.py to adjust:

• Model settings
• Chunk sizes
• Retrieval parameters
• Storage locations

API Endpoints

Chat & Queries

• POST /chat - Query the knowledge base
• GET /health - System health check

Document Management

• POST /upload - Upload a PDF
• GET /documents - List all documents
• DELETE /documents/{name} - Delete a document
• POST /rebuild-index - Rebuild the vector index

Web Interface

• GET / - Serve the web frontend
• GET /docs - Interactive API documentation

Workflow

1. Start Server → uv run main_api.py
2. Open Frontend → http://localhost:8000
3. Upload PDFs → Use the upload area (drag & drop supported)
4. Ask Questions → Type your question and get instant answers
5. Manage Documents → Delete documents as needed

Development

Code organization:

• Configuration centralized in config.py
• Shared LLM instance via get_llm()
• Type hints throughout
• Modular, testable functions
• Frontend is fully client-side (no Node.js needed)

Technical Stack

• Backend: FastAPI
• LLM: Google Generative AI (Gemini)
• Embeddings: Hugging Face (BAAI/bge-small-en-v1.5)
• Reranking: Cross-encoders (BAAI/bge-reranker-base)
• Vector DB: Chroma
• Search: Hybrid (vector + BM25)
• Frontend: HTML5, CSS3, Vanilla JavaScript

Notes

• The system starts with an empty knowledge base
• Upload at least one document before asking questions
• The frontend automatically detects when the database is empty
• All timestamps are displayed in local time
• The API automatically handles document chunking and embedding