Show HN: HNSW index for vector embeddings in approx 500 LOC

a year ago

HNSW is a graph structure with levels that are sparsely populated at the top and densely populated at the bottom.
Nodes in a layer connect to nearby nodes on the same level.
Insertion involves picking a random level and inserting the node there and all levels below.
Searches start at the top level, descend, and track the K nearest nodes.
The implementation is a single-header, modern C++ library with about 500 lines of code.
It uses Eigen for SIMD acceleration to speed up distance calculations.
Example usage includes creating an index, adding vectors, and searching for nearest neighbors.

Hasty Briefsbeta