Skip to content

v1.2.0 — Hybrid search + pipeline hardening

Choose a tag to compare

@solarcloud7 solarcloud7 released this 23 Jun 05:00
a0af766

v1.2.0 — Hybrid search + pipeline hardening

New

  • Hybrid retrieval across all five search tools: RRF reranking over the
    ingest-built FTS index + dense vectors, with transparent fallback to pure
    vector where no FTS index exists. (make eval reports recall@k for vector vs
    FTS vs hybrid on a golden set.)

Fixed (ingestion / data correctness)

  • .lua files are now AST-chunked (were forced through the TypeScript grammar).
  • file_path stored POSIX-normalized; repo_url keyed by owner/repo slug
    (unique, no .git-strip collisions) with a one-time migration of old keys.
  • Per-file dedup (store-wide dedup had collapsed distinct files sharing content).
  • Orphan reconcile on file rename/delete; idempotent no-op re-ingest.
  • Token-correct, prefix-aware chunk cap (≤510 incl. context prefix); multi-line
    import-aware AST coverage.

Fixed (server)

  • Escaped SQL LIKE filters (%/_/\ + ESCAPE) for plugin/repo scoping.
  • ensure_databases extracts only missing stores (never clobbers a local data/).

Quality

  • Permanent chunk-health gate (ChunkAuditor) + a measure-once --dry-run
    mode; 64 offline tests incl. gate-first invariants; validated by a 46-agent
    code review (1 high + 7 medium findings, all fixed).

The bundled LanceDB stores (factorio_lancedb.zip) were rebuilt for this release.
Existing installs keep their local data/; delete a store dir to pull the refresh.