hnsw

Hierarchical Navigable Small World Graph for fast ANN search

Enable the serde feature to serialize and deserialize HNSW.

Tips

A good default for M and M0 parameters is 12 and 24 respectively. According to the paper, M0 should always be double M, but you can change both of them freely.

Example

To see how this might be used with hamming space, see tests/simple_discrete.rs. To see how this might be used with euclidean space, see tests/simple.rs.

Note that the euclidean implementation in the test may have some numerical errors and fail to implement the triangle inequality, especially on high dimensionality. Use a Kahan sum instead for proper usage. It also may not utilize SIMD, but using an array may help with that.

Please refer to the space documentation for the trait and types regarding distance. It also contains special Bits128 - Bits4096 tuple structs that wrap an array of bytes and enable SIMD capability. Benchmarks provided use these SIMD impls.

Benchmarks

Here is a recall graph that you can compare to its alternatives:

For more benchmarks and how to benchmark, see benchmarks.md.