Voyager

Spotify built Voyager to run nearest-neighbor lookups over embedding vectors held in memory, and the company relies on it behind many user-facing features at very high query volume. Under the hood it implements HNSW graphs, extending the hnswlib codebase, and gives Python and Java the same feature set (Scala projects can use the Java artifact), so an index saved from one language loads unchanged in the other. Indexes support Euclidean, inner product and cosine distance, tunable graph parameters such as M and ef_construction, multithreaded batch inserts, deletion by marking IDs, and reduced-precision storage in 8-bit fixed point or E4M3 floating point to cut memory use. Install from PyPI or add the com.spotify artifact to a Maven or sbt build; prebuilt wheels cover Python 3.9 to 3.13 on Linux, macOS and Windows, with ARM builds on Linux and macOS. No GPU or server process is involved, which makes it a library to embed inside an application rather than a database to operate. Apache-2.0 licensed for commercial use. A good pick for teams wanting better recall than Spotify's older Annoy without running a vector service.