v0.5.0

What's New

Composable Incremental Index

IncrementalDiskANN now supports filtered search (labels) and quantized search (F16/Int8/PQ), independently or combined on a single index — without code duplication.

New API

• build_with_labels() — incremental + filtered
• build_quantized_f16() / build_quantized_int8() / build_quantized_pq() — incremental + quantized
• build_full() — incremental + filtered + quantized (all three)
• search_filtered() / search_filtered_with_dists() — label-based filtered search on incremental indexes
• add_vectors_with_labels() — add labeled vectors without rebuilding
• IncrementalQuantizedConfig, QuantizerKind — configuration types

Internal Architecture

• GraphIndex trait — internal abstraction for searchable graph indexes, enabling code reuse across all index types
• UnifiedView adapter — presents base + delta layer as a single graph for multi-seed beam search
• Multi-seed beam_search — explores disconnected base/delta subgraphs in a single pass
• Extracted free functions — filtered_search() and quantized_search() shared by all index types

Serialization

• New incremental format with INCR magic number and versioning
• Full backward compatibility with v0.4 format (auto-detected on load)
• Labels and quantizer state survive to_bytes() / from_bytes() roundtrips

Bug Fixes

• Tombstone filtering now integrated into beam search filter (not post-hoc), returning correct k results
• Fixed lock ordering in add_vectors / add_vectors_with_labels to prevent label/vector misalignment under concurrency
• stats().total_live uses saturating subtraction to prevent underflow

No Breaking Changes

All existing public APIs remain unchanged. All 138 tests pass (75 unit + 63 integration).

Full Changelog

v0.4.0...v0.5.0