- Add app/fallback_matcher.py for intelligent crisis entity matching with token and latency observability
- Route Jev failure/timeouts in app/matching.py to DeepSeek fallback engine
- Handle MatchingFailureError defensively in handlers/__init__.py with honest error message
- Enforce strict 'new' entity status in handlers/quote.py and implement defensive get_or_create on commit
- Clean up app/resolve.py to eliminate difflib and character ratio heuristics
- Pass all smoke tests (59/59), fallback matching tests (15/15), and matching unit tests (13/13)
- tools/seed_mock.py: idempotent additive mock data (uuid5 ids) covering all
tables (units, brands, specs, suppliers, materials + EAV/brand/unit links,
projects, warehouses, inventory, quotes with termsheets/line items/decisions,
receipts) so phases 6-7 can be exercised end-to-end
- tools/calibrate_matching.py: A/B difflib vs Jev on 35 scenarios (31 positive
+ 4 negative); collects RAW picks with widened bands, classifies against the
configured thresholds, sweeps EXACT 0.55-0.95 (negatives count as binds),
recommends FLOOR strictly above the highest wrong pick with margin for
±0.05 run-to-run confidence jitter; markdown report + real traces
- matching.py: below-FLOOR picks no longer suppress difflib — they log a
routing.decision event (source=jev-lowtrust) and leave no result so the
existing difflib path keeps its chance (0.54 «پردیس» was lost to «new»)
- config.py: JEV_CONF_FLOOR 0.60 -> 0.65 (removes the جعبه→بسته @0.61 wrong
candidate offer; correct sub-floor picks باق اسپارت/کیلو are difflib-covered)
- test_matching.py: +2 low-trust cases (13/13 mock, 16/16 live)
- docker-compose.yml: mount bot/tools into the bot container
Phase 2 — jev.py client: POST /v1/systemone, PINNED model jev-1.13.0 (drift warning when the response model differs), one retry on 429/5xx, fail-soft (None on any error), usage/latency returned for observability. Config adds TYPESAFE_API_KEY / JEV_* (thresholds 0.85/0.60, 15 questions per field); no daily cap by owner decision. bot/.env is now mounted read-only into the container so the key never enters compose/git.
Phase 3 — matching.py: برونگروهی (one request per field, parallel), درونگروهی (one choice question per mention sharing the same full candidate list + «_none»), درونواریانسی (all STT spellings in one question). Routing: conf>=exact → bind; floor..exact → ambiguity; else/_none → new; missing/failed field → difflib fallback (bug fixed during testing: a failed request used to record 'new' instead of falling back). Resolver gains full cached lists for suppliers/materials (paged) plus brands (API+DB) and specs (EAV values + draft specs). NLU prompt now emits entity_mentions (6 fields) with the voice-variant grouping rule; sanitize_mentions guards the shape. handle_message runs matching only for register_quote/update_draft/add_material behind JEV_ENABLED; jev_map threads into quote/material resolve_payload (Jev hit skips difflib; exact writes ids; ambiguity skips difflib). Spec effect per owner: spelling normalization only (normalize_brand_spec).
Phase 4 — entity-match ambiguity sub-messages: create_ambiguities() builds deduped entries {kind: entity_match, field, variants, candidates, details with Persian percentages}; new keyboards.ambiguity_candidates_keyboard renders one button per candidate + «هیچکدام — ایجاد جدید»; callbacks handle amb⛏️{prefix}:{amb_id}:{idx} (applies the candidate via matching.apply_choice then re-resolves/renders) and amb:new (rejected → auto-create path at commit). Every user choice lands in observability (ambiguity.user_choice with arg+candidates).
Phase 5 — voice variants: NLU groups different STT spellings of one entity into one variants[] (verified live: مصطفوی/مستفدی → one supplier mention); tools/test_matching.py live section (RUN_LIVE=1) exercises real DeepSeek + real Jev without touching Bale.
Tests: test_jev.py 5/5 (mock HTTP: no-key, happy path, 429-retry, 400, exception), test_matching.py 11/11 mock + 3/3 live (real NLU produced 7 mentions with correct fields; real Jev matched «50 کیلویی»→«۵۰ کیلویی» 0.98 and «کیسه» 1.0 and correctly returned new against the emptied catalog); observability regression 8/8.
whqreza