Query
↓ Expand queries with LLM (better recall)
↓ Parallel dense + sparse embedding (hybrid search)
↓ Hybrid retrieval (best of both worlds)
↓ Deduplicate (remove redundant results)
↓ Normalize (score consistency)
↓ Cross-encoder rerank (precision boost)
↓ Context reconstruction (better answers)
↓ Context merge (coherent output)
↓ LLM answer (final response)
