CodeDB 0.2.5853 — hybrid retrieval accuracy

CodeDB 0.2.5853

This release improves default hybrid retrieval using the existing hosted Jina
embedding service and local OpenPuffer index. No configuration change or index
rebuild is required for these ranking changes.

Changes

• Sharper lexical/semantic rank fusion preserves strong semantic matches.
• Unique exact code definitions take priority for definition-oriented queries.
• Implementation and test-finding requests softly favor the requested file type.
  Test intent recognizes explicit test-finding requests, preserving ordinary
  words such as latest and production testing-infrastructure questions.
• Caller/use questions do not force exact definitions ahead of reference sites.
• Adds a reproducible 128-question evaluation suite across six repositories,
  pinned public datasets, live-service comparisons, and a remaining-failure catalog.

Hosted Jina, default hybrid, vector-space calibration, the existing ANN candidate
pool, and hosted exact fallback remain in place. No local embedding model is
introduced.

Accuracy evidence and limits

The successive native experiments improve OpenClaw correct-first results from
26/32 to 30/32 and Express from 8/16 to 12/16. On the expanded suite the final
test-intent change improves Anyhow from 16/20 to 18/20 without per-question rank
regressions. Its fresh Requests holdout retains 17/20 correct-first results and
improves NDCG@5. These comparisons have different baselines; see ADRs 0008–0010
for exact provenance and the OpenClaw tradeoff between intermediate candidates.

These small, hand-labeled file-retrieval fixtures are not a universal accuracy
claim. Fifteen first-result misses remain; one Chi question misses the top five.

Upgrade

codedb update
codedb --version

This is a GitHub-only release; no npm package is published for this version.