Approximate Nearest Neighbor Search

Approximate nearest neighbor search finds vectors similar to a query without comparing it against every item, making large-scale embedding retrieval practical. Common approaches include graph indexes such as HNSW and DiskANN, inverted-file indexes with product quantization, locality-sensitive hashing, and learned routing; many compress vectors or rerank a shortlist to save time and memory. Methods are compared using recall@k, query speed, memory, and index-building cost on embedding benchmarks and workloads such as billion-vector, filtered, and streaming search. A central challenge is balancing these costs as data changes, while ensuring that exact-neighbor overlap reflects useful semantic retrieval.