The DiskANN library: Graph-Based Indices for Fast, Fresh and Filtered Vector Search

Approximate nearest neighbor search has become a core component of AI systems on cloud and edge, spanning extremes of scales and form factors. We overview the DiskANN library of graph-based indices and algorithms that enable the practical construction and deployment of approximate nearest neighbor search indices across a variety of such systems. Specifically, we present indices
that are capable of running efficiently out of an SSD, preserving recall over a stream of updates, and incorporating attributes alongside vector data to support predicate filters. They also support performance at least on par with other “in-memory” graph-based indices. Interestingly, all these algorithms arise from a variation of the prune procedure used in most graph-based indexing algorithms