Hey all, I built brinicle, an in-process C++ vector engine with a python wrapper that consumes substantially less memory, while staying quite fast.
On 1.2 million Amazon products, it achieves sub-ms P99 latency.
It also supports lexical search, and hybrid search.
In the hybrid search, we did not try to build two indexes and then fuse results, we create ONE HNSW Graph and utilize it for semantic, lexical, and hybrid search.
benchmark comparisons: https://brinicle.bicardinal.com/benchmark
hybrid search benchmark comparisons: https://brinicle.bicardinal.com/search_benchmark
Repository::
github.com/bicardinal/brinicle