This release adds 3 notable features for engineering teams evaluating rollout.
✓ No known CVEs patched in this version
Summary
AI summarysearch_chunks gains configurable score_mode and diagnostic metadata.
Changes in this release
| Type | Severity | Summary | CVE |
|---|---|---|---|
| Feature | Medium |
Added score_mode=auto/raw/centered to search_chunks. Added score_mode=auto/raw/centered to search_chunks. Source: granite4.1:8b-q6_K@2026-05-21 Confidence: high |
— |
| Feature | Medium |
Scoped path searches keep raw cosine by default; centered scoring can be forced manually. Scoped path searches keep raw cosine by default; centered scoring can be forced manually. Source: granite4.1:8b-q6_K@2026-05-21 Confidence: high |
— |
| Feature | Medium |
search_chunks now returns diagnostic score metadata: score, score_raw, score_centered, score_mode, candidate_count, centroid_norm, and score_gap. search_chunks now returns diagnostic score metadata: score, score_raw, score_centered, score_mode, candidate_count, centroid_norm, and score_gap. Source: granite4.1:8b-q6_K@2026-05-21 Confidence: high |
— |
| Feature | Medium |
Added scripts/benchmark_chunk_scoring.py for anonymized raw-vs-centered benchmark reports. Added scripts/benchmark_chunk_scoring.py for anonymized raw-vs-centered benchmark reports. Source: granite4.1:8b-q6_K@2026-05-21 Confidence: low |
— |
| Refactor | Medium |
Unscoped large chunk searches now default to mean-centered cosine scoring. Unscoped large chunk searches now default to mean-centered cosine scoring. Source: granite4.1:8b-q6_K@2026-05-21 Confidence: high |
— |
Full changelog
Changed
- Added score_mode=auto/raw/centered to search_chunks.
- Unscoped large chunk searches now default to mean-centered cosine scoring to reduce anisotropic embedding background.
- Scoped path searches keep raw cosine by default; centered scoring can still be forced manually.
- search_chunks now returns diagnostic score metadata: score, score_raw, score_centered, score_mode, candidate_count, centroid_norm, and score_gap.
Added
- Added scripts/benchmark_chunk_scoring.py for anonymized raw-vs-centered benchmark reports without note paths, headings, titles, or note text.
Verification
- python -m compileall -q nouz_mcp pytest_smoke.py scripts
- python -m pytest -q
- python test_server.py
- live search_chunks smoke on the local SQLite index
- python -m build --sdist --wheel
- python -m twine check
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
About KVANTRA-dev/NOUZ-MCP
Semantic knowledge graph for Obsidian. Three modes (pure graph / semantic classification / strict hierarchy)
Related context
Beta — feedback welcome: [email protected]