Vector Database

Stores embeddings and does nearest neighbor search at scale.

Why not Postgres?

Postgres with pgvector works up to ~1M vectors. Beyond that, specialized indexes win: HNSW, IVF-PQ.

Indexes

• HNSW (Hierarchical Navigable Small World) — Graph. Fast, high recall, high memory. Used by Qdrant, Weaviate.
• IVF (Inverted File) — Cluster vectors into Voronoi cells. Search only closest cells. Needs nprobe tuning.
• PQ (Product Quantization) — Compress vectors.

Tradeoffs

Index	Query	Memory
HNSW	~1ms	high
IVF+PQ	~5ms	low

Links

• PostgreSQL Extensibility — pgvector as extension
• rate-limiter — rate limiting embedding API calls

Seed from reading: 2026-08-23