Production-grade RAG system that classifies query complexity in real time and routes each question to the optimal retrieval strategy. Features hybrid search (dense + BM25 via RRF fusion), cross-encoder reranking, semantic caching (Redis, 50-70% cost reduction), agentic decomposition for multi-hop queries, and web-augmented retrieval fallback. Ships with a React chat UI, Clerk auth, LangFuse observability, and automated RAGAS evaluation.



