RAG System

A Retrieval-Augmented Generation (RAG) system implementation in Python supporting multiple LLM providers: Google Gemini, OpenAI, and Anthropic Claude.

Overview

This project implements a powerful RAG (Retrieval-Augmented Generation) system that enhances large language model responses by retrieving relevant context from your documents. It combines efficient document retrieval with various LLM providers to provide accurate, context-aware responses with citations.

Features

• Multiple LLM Provider Support: Use Google Gemini, OpenAI, or Anthropic Claude models
• Document Processing: Ingest PDF documents and split them into manageable chunks
• Vector-based Retrieval: Utilize FAISS for efficient similarity search
• Embeddings Generation: Google's embedding model for high-quality vector representations
• Context-aware Response Generation: Enhance responses with relevant document context
• Citation Support: Responses include page and chunk citations
• Interactive Query Interface: Simple command-line interface for asking questions

Technical Stack

• Language Models:
  • Google Gemini 2.0 Flash
  • OpenAI GPT-4o
  • Anthropic Claude 3 Opus
• Embeddings: Google Generative AI Embeddings (models/embedding-001)
• Vector Database: FAISS (Facebook AI Similarity Search)
• Document Processing: LangChain's document loaders and text splitters
• Environment Management: Python dotenv for API key management

Installation

First, create and activate a virtual environment:

# Create virtual environment
python -m venv venv

# Activate on Windows
venv\Scripts\activate

# Activate on Linux/Mac
source venv/bin/activate

Then install dependencies:

pip install -r requirements.txt

Configuration

1. Copy the .env.example file to .env:

   cp .env.example .env

2. Add your API keys to the .env file:

   # Google API Key for Gemini
   GOOGLE_API_KEY=your_google_api_key_here
   
   # OpenAI API Key
   OPENAI_API_KEY=your_openai_api_key_here
   
   # Anthropic API Key for Claude
   ANTHROPIC_API_KEY=your_anthropic_api_key_here
   
   # Default LLM provider (google, openai, or anthropic)
   LLM_PROVIDER=google
   
   # Default models for each provider
   GOOGLE_MODEL=gemini-2.0-flash
   OPENAI_MODEL=gpt-4o
   ANTHROPIC_MODEL=claude-3-opus-20240229
   

Usage

The system has two main scripts:

1. Create Embeddings: Process documents and generate vector embeddings

   python create_embeddings.py --input docs --output faiss_index

   • --input: Path to directory containing documents or to a single document file (default: "docs")
   • --output: Directory to save the embeddings (default: "faiss_index")

2. Query Embeddings: Ask questions about your documents

   python query_embeddings.py --index faiss_index [--expanded] [--mmr] [--rerank] [--rerank_model MODEL_NAME] [--provider PROVIDER] [--model MODEL_NAME] [--compression] [--retrieval-qa]

   • --index: Directory containing the embeddings (default: "faiss_index")
   • --expanded: Enable expanded mode to generate more comprehensive answers by creating additional related queries
   • --mmr: Use Maximum Marginal Relevance search for more diverse results with reduced redundancy
   • --rerank: Enable CrossEncoder reranking for improved result relevance
   • --rerank_model: Specify which CrossEncoder model to use (default: "cross-encoder/ms-marco-MiniLM-L-6-v2")
   • --provider: Specify which LLM provider to use: "google", "openai", or "anthropic" (overrides .env setting)
   • --model: Specify a specific model to use with the chosen provider (overrides .env setting)
   • --compression: Use document compression to extract the most relevant content from retrieved documents
   • --retrieval-qa: Use LangChain's RetrievalQA chain for structured question answering with detailed references
   • Type 'exit' to quit the application

Project Structure

create_embeddings.py    # Script to process documents and create embeddings
query_embeddings.py     # Script to query the embeddings and get answers
requirements.txt        # Python dependencies
docs/                   # Place PDF documents here
faiss_index/            # Vector database storage
  ├── index.faiss       # FAISS index file
  └── index.pkl         # Pickle file for document metadata
manage.py               # Deprecated main application file
.env.example            # Example environment file with configuration options

How It Works

1. Document Ingestion: PDF documents are loaded and split into chunks
2. Vector Embedding: Text chunks are converted to vector embeddings
3. Storage: Embeddings are stored in a FAISS index for efficient retrieval
4. Query Processing: User questions are converted to embeddings and used for similarity search
5. Response Generation: Retrieved relevant context is sent to the chosen LLM with the user query
6. Output: The system returns context-enriched responses with citations

Contributing

Feel free to open issues and pull requests. Contributions are welcome!

License

MIT License. See LICENSE file for details.