IVF-Flat

IVF-Flat is a vector indexing algorithm that accelerates similarity searches by clustering high-dimensional data into Voronoi cells and performing exact distance calculations only within relevant partitions.

IVF-Flat (Inverted File Flat) optimizes vector search by partitioning the embedding space into a user-defined number of clusters (nlist) using k-means. Unlike exhaustive brute-force methods, it limits search operations to the most relevant clusters (nprobe) identified during the query phase. This approach maintains high precision because it stores vectors in their original, uncompressed form within each cluster, avoiding the accuracy loss typical of product quantization. It is a standard choice for datasets where memory is sufficient to hold raw vectors but latency requirements demand sub-linear search times: common in PostgreSQL's pgvector, Faiss, and Milvus deployments.

https://milvus.io/docs/index.md#IVF_FLAT