Embeddings

IVF centroids and PQ codebooks are fitted from a sample of the corpus before ingestion, then frozen. When later data drifts away from that sample, cells become unbalanced and codes reconstruct poorly, so recall degrades quietly while queries keep succeeding.

Derive it from labelled pairs on that exact model and corpus, choose the operating point from the cost of a wrong auto-close, and prefer a margin over the runner-up to a bare number. Re-validate on every model or corpus change.

Compare against the incumbent per query segment, not in aggregate. Semantic search wins on long natural-language queries and can lose on head, identifier and navigational ones, so the decision is which segments improve, which regress, and whether the regressions are acceptable.

Treat the build as a batch job, not an online operation. Shard so builds run in parallel, keep serving the old index while the new one builds on separate capacity, then swap atomically behind the query path. Reserve headroom for two copies.

Whenever the corpus fits in memory and exhaustive scanning meets the latency budget. Exact search gives perfect recall with no training, no tuning, no rebuild and trivial updates — approximation is a cost you should only pay once the arithmetic forces it.

A text encoder and an image encoder are trained together on image-caption pairs so both output into one shared vector space. A description and a matching photo land near each other, so ordinary similarity search over image vectors answers a text query.

That chunk is a hub. In high-dimensional spaces a few vectors sit near the centre of the distribution and land in many nearest-neighbour lists regardless of the query. Generic boilerplate is the textbook hub.