HNSWlib

Approximate nearest-neighbor lookup on layered small-world graphs, the HNSW algorithm, is the whole job here, delivered as C++ headers you drop into a project plus a pip-installable Python module. You choose squared L2, inner product, or cosine distance, set the graph degree M and the ef_construction build parameter, add vectors in multithreaded batches, then query with a tunable ef value that trades recall against speed. Indexes grow incrementally and can be resized; re-adding a label updates its vector, deleted items can be masked and their slots reused, queries accept a label filter, and Python indexes save to disk or pickle. C++ users can plug in custom distance functions and run multi-vector or epsilon searches, while R and Java bindings come from outside projects. The code is Apache-2.0, needs nothing beyond a C++11 compiler, and runs entirely on CPU. A 0.10 release candidate in September 2026 added an exception-free C++ API and stream serialization. It suits engineers embedding vector search inside their own service or library rather than running a separate database.