Commit History
-
RU/EN bilingual fixes from adversarial audit: tokenizer, numbers, countries, weights, HS guard, multi-item sets, currencies, OCR langs, xlsx intake, sample-doc isolation
Aissugur Kurganbekov +1: Claude Fable 5- retrieval: Cyrillic tokenizer + Russian stopwords (Russian questions no longer auto-abstain) - packet_audit: RU decimal commas (4 120,50 != 412050 bug), country alias map (Kitai==China==CN), per-document value SETS instead of first-match - extractors: weights must contain digits (kills the 'ves brutto, kg' false pass), label-unit-number form, INN/BIN guard on bare 10-digit HS fallback, multi-word countries, currency words normalised to ISO (tenge->KZT), RU question triggers, find_field_values for multi-line-item documents - parsing: OCR lang rus+eng+kaz+chi_sim with graceful fallback; .xlsx intake (openpyxl) - config: multilingual embeddings default (paraphrase-multilingual-MiniLM-L12-v2) - API/UI: Document.is_sample; demo samples stay OUT of a real user's packet selection - tests: 20 regression tests reproducing every audited failure (79 green) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>c86642c -
Enable Thinking mode: LLM7 free proxy config baked into image (no secrets)
Aissugur Kurganbekov +1: Claude Opus 4.8Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>7e9302c
-
Tamga — evidence-backed customs document audit
Aissugur Kurganbekov +1: Claude Opus 4.8Reads a customs declaration packet (commercial invoice, packing list, customs declaration, certificate of origin) and audits it before filing: a deterministic checklist audit per document plus a cross-document packet audit that flags the mismatches that stall clearance (weights, HS code, origin, currency, Incoterms). Every finding cites the exact document and line; when the packet does not contain a value, Tamga reports it as not found instead of guessing. FastAPI + SQLite backend (TF-IDF retrieval with optional dense embeddings, reranking and OCR, all gracefully degrading), React/Vite frontend with a guest demo, evidence reports, and an anti-hallucination evaluation suite (59 tests, golden dataset, A/B harness). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>373030b