Vector stores and indexes

Approximate, on purpose

Exact nearest-neighbour search compares the query against every vector and is fine up to a point. Beyond that, approximate indexes trade a small amount of recall for orders of magnitude less latency.

Index parameters are a dial between speed and completeness. Tune them against your own evaluation set rather than accepting defaults, because the right point depends on how much recall your application can lose.

Ask how the store handles updates before you pick it. Some indexes degrade as documents are deleted and need periodic rebuilds, which is an operational cost you inherit.