Retrieval-Augmented Generation

This project implements a Retrieval-Augmented Generation (RAG) workflow that uses,

• 🕸️ Web search via Tavily
• 📚 Document retrieval via vector database
• 💬 Local LLMs via Ollama
• 🔀 Dynamic reasoning and control flow using LangGraph
• 📊 Tracing and evaluation using Langfuse

🚀 Features

• Hybrid RAG with both web and vector-based retrieval
• Automatic query reformulation
• Relevance filtering before answer generation
• Hallucination detection and correction
• Tracing via Langfuse
• Modular node/edge logic using LangGraph

This RAG system is built with a hybrid retrieval architecture combining multiple search strategies:

• ChromaDB Vector Search: Retrieves documents based on semantic similarity using dense embeddings from a Chroma collection.
• BM25 Lexical Search: Identifies relevant documents using traditional keyword-based matching with BM25 scoring.
• Hybrid Retrieval: Integrates results from both semantic and lexical searches using Reciprocal Rank Fusion (RRF) for balanced and robust retrieval.

🔐 Environment Variables

Create a .env file in the project root:

TAVILY_API_KEY=your_tavily_api_key
LANGFUSE_PUBLIC_KEY=your_langfuse_public_key
LANGFUSE_SECRET_KEY=your_langfuse_secret_key