0 Releases 11 Tags

RSS Feed

• v0.7.0 078ec029da

  Compare

  v0.7.0 — brain hybrid retrieval + reranker + OAuth + structured taxonomy + knowledge/ vector coverage

  mathias released this 2026-05-19 19:31:41 +00:00 | 25 commits to main since this release

  Highlights since v0.6.0:

  Brain — retrieval quality

  • feat(brain): hybrid BM25 + pgvector retrieval (opt-in via BRAIN_PG_DSN +
    BRAIN_EMBED_URL). RRF merge of BM25 top-k and vector top-k.
  • feat(brain_answer): Qwen3-Reranker cross-encoder filter (opt-in via
    BRAIN_RERANKER_URL). Prunes BM25 top-20 to relevant top-5 before
    LLM synthesis.
  • feat(brain): cross-wing tunnels — bidirectional wikilinks + auto-detect.
  • feat(brain): structured wing/hall taxonomy + obsidian-compatible layout.

  Brain — coverage / observability (infra#37 part 1)

  • fix(ingestion): embed sync also scans brain/knowledge/ + per-item
    WARN log on failure. 110 curated knowledge entries now embedded
    (was 0); brain_embeddings 31 → 141 rows. Vector arm of hybrid
    retrieval no longer blind to the most-important brain content.

  Brain MCP — auth

  • feat(brain-mcp): OAuth 2.0 client_credentials flow for claude.ai
    custom-MCP UI.
  • fix(brain-mcp): static Bearer short-circuits before OAuth challenge.

  Tooling / project hygiene

  • feat(mcpclient): fail-fast on empty bearer token.
  • fix(project_create): commit staging namespace directly to infra main.
  • fix(ingestion): redact password from BRAIN_PG_DSN log line.
  • fix(scripts): brain-embeddings-init.sql psql-level conditionals.
  • test(routing): de-flake TestRoutingPodEndToEnd.

  Carry-over

  • infra#37 part 2 (long-file chunking before embed) — 3 wiki/knowledge
    files still exceed nomic-embed ctx and skip embedding.

  Downloads

  • Source Code (ZIP)
  • Source Code (TAR.GZ)