The reason your pgvector benchmark is lying to you

pgvector benchmarks on small datasets (10k vectors, 128 dimensions) are misleading and fail to predict production behavior at scale (5M vectors, 1536 dimensions). Key operational practices for success include: running representative benchmarks before committing to an architecture, choosing between IVFFlat and HNSW indexes based on workload needs, tuning parameters like ef_search and M deliberately, leveraging hybrid SQL+vector retrieval with WHERE clause filtering to narrow candidate sets, partitioning tables on query-filter fields rather than just data volume, and using pg_prewarm to avoid cold-cache penalties after deployments. pgvector is a viable production tool when treated with the same operational rigor as any serious Postgres workload, though sub-20ms latency at tens of millions of vectors may eventually require a purpose-built solution.