# Changelog

## v0.3.0 (2026-05-18)

It adds new search capabilities first, then the Python-facing stack built on top of them, plus the refactors and tests needed to support updates and filtering consistently.

### Feature highlights

#### Speculative filtering

- Added speculative filtered ANNS for arbitrary attribute constraints.
- Uses lightweight in-memory probabilistic filters to explore a superset of valid vectors, then verifies final candidates exactly against SSD-resident attributes.
- A cost model chooses speculative pre-filtering, speculative in-filtering, or post-filtering per query.
- Supports label filters, range filters `[l, r)`, and Boolean combinations through native selectors.
- Added typed attributes (`Attributes`, `AttrsVec`), on-disk attribute indexes, dense-neighbor support (`range_dense`), and JSON filter config loading.

#### OOD refinement

- Added NGFix-style out-of-distribution graph refinement.
- Exposed `train_...