knowhere-v2.2.2

Release Notes

What's New

• SCANN without refine: (#77)

  • SCANN now supports a with_raw_data config option to determine whether the refine process is needed. Omitting the refine step can dramatically reduce memory usage (to approximately 1/17 of the original).

• Multi devices for GPU: (#83 #146)

  • GPU indexes can now be distributed in different cards to accelerate and the performance and enlarge the capacity.

• Remove IVF_FLAT_NM: (#39):

  • IVF_FLAT_NM was introduced to conserve disk resources by avoiding the encoding of raw data during serialization. However, this approach posed challenges for the support of our new features and when upgrading FAISS. In this release, we've reverted the implementation of IVF_FLAT to the original FAISS version.

• DiskANN filtering search optimization: (#143)

  • We adopt a strategy that randomly determines whether to add a filtered vector to the candidate list, aiming to enhance DiskANN's filtering search performance with a high filtering rate.

• Reprioritize thread pool: ( #130)

  • We deprioritize Knowhere's threads to support better resource allocation of Milvus.

• Support GetVectorById for IVFFLAT Cosine (#80)

Enhancements and Fixes

• Replace OMP with Knowhere's thread pool to better manage thread number. (#120)
• Implement Size() API to estimate the memory usage for HNSW and DiskANN. (#124 #126)
• Fix race condition in GetVectorById of IVF indexes. (#133 #134)

New Contributors

• @alexanderguzhva made their first contribution in #72