ducoterra
dbf2af26c6
Standardize on the .agents/ directory (shared with project skills): phases/, user_stories/, reports/, screenshots/, validate.sh, and phase-sessions/ + pipeline.log all move to .agents/ (git mv preserves history; runtime artifacts move alongside). Updates every reference in AGENTS.md, README.md, .gitignore, app docstrings, and test story headers. Historical KB content in data/ and the runtime pipeline.log transcript are left untouched.
1.8 KiB
1.8 KiB
Phase 09 is complete. Final report:
Phase 09 report — commit 7e8d147 (36 files, +2021/−293)
- A7 hybrid retrieval: cosine top-100 ∪ Postgres FTS top-30 (stored
tsvector+ GIN, OR-tsquery), RRF k=60, LOCKED tiebreak. Fusion iteration = vector window 30→100: measuredgitlab.md:1double-hit 0.02212 > template 0.01639 → "How did I install gitlab?" now ranks gitlab.md #1 ✓ - A8 gate: LOW only when best cosine < 0.62 and
fts_hits == 0;fts_hitsin per-turn log +query_log(migration 0002, nullable) - A9 scope:
md,markdown,txt,yaml,yml,json,py(stdlib-only chunkers: YAML---/top-level keys, JSON top-level keys,astdefs/classes), dot-prefixed paths skipped,--prunedrops filter-outs, per-format summary counts - Live KB re-import: 520 docs / 2774 chunks / 0 errors in 72s; zero dot-prefixed paths; k8s/backup/dns questions all ground correctly
- Gates: 174 unit+integration passed · coverage 98% (>90%) · ruff + pyright clean · all 9 E2E suites green in isolation, incl. new
test_retrieval_quality.py(4 tests: multi-format import, gitlab grounding, FTS-beats-vector, sourdough deflection)
Deviations (flagged per protocol, not silent):
- "sourdough → LOW" is unmet at the locked 0.62: measured off-topic best-cosine floor with the real
embedmodel is 0.73 (battery: off 0.61–0.73, on 0.65–0.84 — the phase's 0.41–0.58 assumption was too low), so the gate answers. Fix when you want it:BOR_RELEVANCE_THRESHOLD≈0.77(then salsa/travel/sourdough deflect; common-word off-topic like tire/pizza still answers via the owner-approved FTS-OR). gitlab-compose.yamllands #3, not top-2 (sits at vector rank 483; no sane window lets it outrank the template's double-free 1/61 term).
Next pending: 10_story_document_viewer.md.