Skip to content

v0.3.7

Choose a tag to compare

@ScionKim ScionKim released this 06 Sep 21:53
· 92 commits to main since this release
d7d9cd5

Reduce redundant fitted-index storage when using donor_policy="complete" with index_factory="Flat".

  • Skip training and donor insertion for the unused full-dimensional Flat index. Transform continues to build projected indexes for observed-feature patterns.
  • Retain factory and metric metadata in an empty fitted index. Its index_.ntotal is now zero.
  • Preserve existing training and insertion validation for other index factories.

Validation: 196 tests passed, including regressions for L2/IP imputation, empty fitted Flat storage, IVF training and metric metadata, and cleanup after factory failures.

In a 100,000-row, 16-feature example, serialized fitted-index size decreased from 6,400,045 to 45 bytes with identical transform outputs. This measures index storage, not whole-process memory.