Annoy

Spotify uses Annoy to find similar tracks and users for recommendations, and the design reflects that job: you add integer-keyed vectors, build a forest of random-projection trees, write the index to a file, then memory-map it from many processes. Because the file sits in the OS page cache, many workers can query one index without private copies, and an index can be built once and shipped like any other artifact. Metrics are angular (cosine), Euclidean, Manhattan, Hamming and dot product. Recall is tuned in two places: extra trees at build time improve it but enlarge the file, and a higher search_k per lookup improves it but slows queries. The catch is immutability: a built index cannot take new items, so changing data means a rebuild, though an on-disk build mode handles datasets larger than RAM. The core is C++ included through one header, with official Python bindings on pip and community ports for R, Java, Go, Rust and Node. It runs on CPU, needs no server, and is Apache-2.0 licensed. A good fit for batch-built embedding indexes of moderate dimension; for frequent updates or metadata filtering, use a vector database.