Why it matters

IVF scales ANN to billions. Understanding shapes vector DB.

Advertisement

The architecture

K-means centroid partitioning.

Query: find nearest nprobe centroids + search those clusters.

IVF ANN structureK-means clusteringcentroidsQuery: nprobe nearestclusters searchedRefinetop-K in clustersFaiss + Milvus; combined with PQ for memory-efficient billion-scale
IVF math.
Advertisement

How it works end to end

Number of clusters nlist.

nprobe = clusters searched.

Trade recall vs speed via nprobe.

Combine with PQ.