...
- Add int8 scalar quantization to the HNSW vector format. This optionally allows for more compact lossy storage for the vectors, requiring about 75% approximately 4x less memory for fast HNSW search.
HNSW graph now can be merged with multiple
threadthreads, leveraging the same infrastructure that inter-segment concurrency utilizes.
Improvements
- Speed up Panama vector support, use FMA, and test improvements.
- FSTCompiler can now approximately limit how much RAM it uses to share suffixes share suffixes during FST construction using the suffixRAMLimitMB method.
...