AI
Vector search
Text or images are embedded as vectors and compared by distance, so results are ranked by similarity rather than keyword match.
Approximate indexes change which results come back, not only how fast. That makes recall a property to measure rather than assume, and it makes an embedding a form of the source data for privacy purposes.
Explained in depth
pgvector
Vector similarity search in PostgreSQL - where the data physically lives across heap, TOAST and index, how HNSW and IVFFlat retrieve it, and how pgvector serves as the memory layer in a RAG or agent system.
Vector and Embedding WeaknessesLLM09:2026 - embedding inversion recovers source text, shared indexes leak across tenants, and poisoning steers RAG without any prompt injection.