[molodec]

It is true that HNSW outperforms LSH on recall and throughput, but for some use cases LSH outperforms HNSW. I just deployed this week to prod a new system for short text streaming clustering using LSH. I used algorithms from this crate that I also built https://github.com/serega/gaoya

HNSW index is slow to construct, so it is best suited for search or recommendation engines where you build the index and serve. For workloads where you continuously mutate the index, like streaming clustering/deduplication LSH outperforms HNSW.