IFVPQ

IFVPQ is a high-performance vector indexing method combining inverted file structures with product quantization to enable rapid similarity searches across billion-scale datasets.

IFVPQ (Inverted File with Product Quantization) serves as a cornerstone for modern approximate nearest neighbor search, most notably within the Faiss library. By partitioning the vector space into Voronoi cells and compressing high-dimensional data into compact codes, it achieves a massive reduction in memory footprint while maintaining sub-millisecond query speeds. Engineers leverage IFVPQ to scale recommendation engines and image retrieval systems to 100 million vectors per machine, striking a precise balance between memory efficiency and search recall.

https://github.com/facebookresearch/faiss