Set refine_factor between 5 and 50 on IVF-PQ queries. It re-scores the top rf times k candidates against the full vectors, which recovers most of the lost recall for a few milliseconds of latency. The docs report over 0.95 recall under 10ms on GIST-1M with about 50 probes and refine_factor 50. Measure the latency impact on your own data before deploying.

Context: Official docs (FAQ): documents the gotcha that IVF-PQ indexes in LanceDB compress vectors with product quantization, which is lossy and hurts recall. Without a refine_factor, searches rank on compressed vectors only and miss results they should have found. Agents that build an IVF-PQ index and never tune this will see mysteriously low recall.