Vector Search (ANN)

Vector search is nearest-neighbor retrieval in an embedding space. The query becomes a vector; the result is the objects nearest to it under a chosen metric. On small collections this is brute force; in production it is Approximate Nearest Neighbor (ANN) search with HNSW, IVF, PQ indexes — near-instant retrieval at a controlled accuracy cost.

How it differs from classical retrieval: a relational query answers “does this exact record exist”, vector search answers “what is semantically similar”. These are different query classes that do not replace each other. In production they almost always live together: the relational store holds entities and links, vector search delivers semantic results filtered by metadata (author, date, project, access).

Where it works: RAG systems, recommendations, deduplication and entity linking, clustering of inbound, anti-fraud. Without metadata filters the result list quickly turns into mush — “similar” lands in the wrong collection, the wrong freshness or the wrong user, and no re-ranker fixes that.