 enne2
|
128059dd4b
|
feat(gateway): strategie rerank oltre la search (A-F)
- A: gate store con cross-encoder — guardrail.decide async, conferma/scarta
quasi-duplicati (CROSS_DUP_CONFIRMED/WEAK/LOW_COSINE), suggerimento
supersedes in WARN, degrada a cosine-only se il reranker è giù
- B: verifica supersede — cross-score (nuovo,vecchio) sotto soglia →
supersede_warning non bloccante + audit
- C: score composito in search — rerank + importance (nuovo campo payload)
+ recency decay (180gg) + authority, pesi SCORE_W_* da env
- E: multi-query — SearchIn.queries (max 3), pool unito con dedup, rerank
unico; endpoint POST /v1/score come primitiva cross-encoder (F-lite)
- extension search.ts: param queries + rerank_score/composite in output
- D: scripts/consolidate.py — dedup periodico a coppie via cross-encoder
con report ntfy e --apply via gateway
- test: 69 pass (+11 strategie); guardrail_version similarity-v2
|
2026-09-08 12:57:33 +02:00 |
|
|
Matteo Benedetto
|
70f3699396
|
test: suite pytest per il gateway (21 test) + script di verifica pre-push
- conftest: FakeQdrant in-memory + mock embed (nessuna dipendenza da Qdrant/Ollama)
- test: validazione (project_id, kind, expires_at, confidence, lunghezza), auth (422/401/429), idempotency (replay/409/key diverse), supersede (404/409/lineage), filtri search, meta, status, metrics
- scripts/check.sh: esbuild (estensione) + pytest (gateway)
- requirements-dev.txt: pytest
|
2026-08-16 19:53:45 +02:00 |
|