Skip to content

v0.4.1

Choose a tag to compare

@Udjin79 Udjin79 released this 03 May 17:11
· 147 commits to main since this release

Documentation-only release. No code changes vs v0.4.0.

This release ships an overhauled README and a docs sweep across ARCHITECTURE.md, benchmarks/README.md, SECURITY.md, and the integration guides. Feature behaviour and APIs are identical to v0.4.0 — if you don't read the docs, you don't need to upgrade.

What changed

  • README: unified benchmark tables (consistent Strict / Combined terminology, symmetric historical columns, cat_1..cat_5 ordering), reframed the no-cat_5 columns as signal-only with a clear noise explanation, added an honest-numbers disclaimer to the needle benchmark, fixed heading hierarchy in the intro, surfaced v0.4.0 wins in the Features list (sliding window, query expansion + smart retry, lazy ingestor), added the missing v0.3.0 synthesize() bullet, added MCP Desktop config paths for macOS / Windows / Linux, and dropped the "alpha — API may change" framing in favour of "actively developed; public API stable".
  • benchmarks/README: aligned headline numbers with v0.4.0 (82.5% strict / 92.2% combined), bridged terminology to the main README, documented the gemini-3-flash-preview judge change, and explained the cat_5 empty-ground-truth artifact.
  • ARCHITECTURE: expanded the recall-pipeline section to cover query expansion, weighted/adaptive RRF, the named 8-stage pipeline with stateful learning, GraphResurrection, the LLM reranker, and answer-mode retry. Listed all six MCP tools (health, search, answer, feedback, graph_query, graph_add_triple) instead of an outdated subset. Added vector.window_size to the config example.
  • SECURITY: bumped supported-versions table from 0.2.0a3+ to 0.4.x and dropped the "alpha" wording.
  • integrations/*: refreshed the MCP-tools list in the README and the Cursor / Claude Code skills (was 2-3, now 6).

Note

PyPI rejects re-uploads of an existing version, which is why this docs work ships as 0.4.1 rather than amending 0.4.0.