Vector‑native RAG on Oracle: embeddings, HNSW/IVF, and hybrid search under database governance

Key Takeaways

Storing vectors in an Oracle VECTOR column alongside content, metadata, and provenance means retrieval happens inside the database. Existing governance — row-level security, auditing, data masking — applies to vector queries the same way it applies to any other query.

Hybrid retrieval is ordinary SQL: VECTOR_DISTANCE handles semantic similarity and WHERE clauses handle business predicates in the same statement. Any reviewer who can read a query can understand what rows qualified and why.

HNSW and IVF are index strategies with real trade-offs in recall, memory footprint, and query latency. Choosing between them requires measuring both on your own corpus — they are not interchangeable defaults.

The answer step can stay inspectable too. Connecting retrieval to a Select AI profile with SHOWSQL means the SQL generation step is reviewable, not just the retrieval that fed it.