HYBRID RAG

Retrieval-augmented QA over internal docs: dense and BM25 retrieval fused with Reciprocal Rank Fusion, reranked by Cohere, answered by Claude with inline citations and per-claim source verification.

A FastAPI service that ingests and chunks documents, retrieves with Pinecone embeddings and BM25 over the same chunks, and merges the two rankings with Reciprocal Rank Fusion. Cohere Rerank trims the candidates, Claude answers from context only with inline citations, and a verification pass checks each claim against its cited chunk and flags unsupported ones. Retrieval evals measure recall@5 and MRR across dense, BM25, hybrid and hybrid + rerank.