2 / 5 Vector Databases & ANN Indexes

Turning the Dial (2)

An HNSW index serves p99 latency of 12 ms against a 50 ms budget, but recall@10 measured against exact search is 0.82 and the product needs 0.95. Which change gets you there first, without rebuilding the index?

A

Switch the distance metric from cosine to Euclidean, which is better behaved on high dimensional embeddings

B

Raise the query-time candidate list size (ef_search), which walks more of the graph per query and spends the latency headroom

C

Raise efConstruction, which builds a better connected graph and lifts recall for every query that runs against it afterwards

D

Lower M so each node carries fewer edges, which lets the search descend to the base layer sooner and examine more candidates there