ALI-OUALA/dzdoc-engine

Deterministic Arabic-French document intelligence engine for Algerian documents # DzDoc Engine Independent source-available foundation for Arabic–French document intelligence. Current deterministic foundation provides: - versioned canonical document models with raw, normalized, and search text; - safe local document ingestion and SHA-256 fingerprinting; - optional PDFium native PDF inspection and a native-text quality gate; - explicit hybrid-pipeline protocols with no vendor imports in core models; - page-level native-text routing and rendering of rejected pages only; - one PP-OCRv5 text-detection pass per OCR page; - Arabic-first region routing to pinned Arabic/Latin PP-OCRv5 recognizers; - deterministic candidate fusion, digit disagreement warnings, alternatives, provenance, confidence, and RTL/LTR geometric reading order; - a disabled-by-default, region-capped PaddleOCR-VL-1.6 fallback whose prompt, immutable model revision, decoding, raw response, trigger, decision, and latency are retained as processing events; - an Algerian invoice pack for supplier/buyer, invoice number/date, NIF, NIS, RC, line items, HT, TVA, TTC, currency, source boxes, confidence, and arithmetic checks; - JSON export and a neutral DZ-Bench `Predictions` artifact. - a versioned FastAPI surface, tenant-scoped API keys, idempotent asynchronous jobs, DB leases, CPU/GPU worker capabilities, content-addressed local/S3 storage, signed webhooks, corrections, usage, retention, and deletion; - Python and TypeScript SDK foundations, OCI images, a one-server Docker Compose profile, and a responsive Arabic/French review workspace. No model weight is bundled or downloaded at install/test time. Native-text PDFs can be inspected and processed without OCR assets. Raster pages fail closed unless `DZDOC_MODEL_DIR` (or `--model-dir`) points to the reviewed asset directories; the engine never performs an implicit unpinned download. ## Quick start ```powershell uv run dzdoc inspect .\document.pdf uv run dzdoc process .\document.pdf --output .\document.json uv run dzdoc export-predict …