Filtering is landed, but is doing 3 reads per vector. We can do much better. The main optimizations are: - Append metadata to the end of the primary vector for search so that we can eliminate 2 reads - Add filter evaluation support on the diskann side, so we can avoid FFI to evaluate the filter
Filtering is landed, but is doing 3 reads per vector. We can do much better.
The main optimizations are: