Adaptive RAG
Corrective Retrieval-Augmented Generation · Agentic AI · LangGraph
Project Info
- Domain: GenAI / Agentic AI / RAG
- Stack: Python, LangGraph, Claude API, ChromaDB, Tavily, FastAPI, Streamlit
- GitHub: View Repository
The Problem
Standard RAG pipelines blindly pass whatever they retrieve to the LLM. When knowledge bases are stale, incomplete, or queries are out-of-distribution, models either hallucinate confidently or give non-answers — with no mechanism to detect the failure.
What I Built
- Designed a self-correcting retrieval loop using LangGraph
StateGraph: an LLM judge (Claude Haiku) scores each retrieved document 0.0–1.0 in parallel, and routes to live Tavily web search when average relevance falls below 0.6 - Implemented the CRAG decompose-and-recompose technique — strips noise from mixed-source context by filtering to only query-relevant sentences before final generation with Claude Sonnet
- Exposed the full pipeline via a FastAPI REST API and an interactive Streamlit dashboard with live pipeline visualization and source attribution
Impact
Eliminates confident hallucinations by generating answers only from verified, high-relevance context — with automatic live fallback and full source attribution on every response.