Executive summary
Production RAG quality depends less on the vector database brand and more on retrieval evaluation, source freshness, reranking, and clear fallback behavior. Strong systems measure whether the correct evidence was retrieved before scoring the generated answer.
What consistently works
Hybrid retrieval combines semantic vectors with keyword matching, then reranks a wider candidate set. Chunking should follow document structure instead of fixed token windows, while metadata filters keep access control and tenant boundaries deterministic.
Recommended operating model
Start with a small golden dataset, track retrieval recall and citation correctness, and inspect failed queries weekly. Add query rewriting only after the baseline is measurable. Route low-confidence requests to clarification or human review instead of forcing an answer.