changelog
What changed
Dated, from the repository. Breaking changes are marked. ← API reference
- v2.1.0
PP-OCRv6 engine, PDF text layer, quality signals
- PaddleOCR 3.x with PP-OCRv6 medium detector and recognizer, pinned by name. Word F1 0.969 on the benchmark set.
- source field on every extract response: text, ocr or hybrid. A sound PDF text layer is used verbatim.
- quality block: blur, long_side_px, low_confidence_ratio and warnings[] so clients can ask for a rescan.
- v2.0
Platform conformance
- Idempotency-Key with a 24 h replay window; 409 idempotency_conflict on a different payload.
- Schema managed by alembic; standard src/ layout and uv lockfile.
- v1.8
Error envelope and dev environment
- One error envelope everywhere: { error, message }, and X-Request-ID on every response.
- Isolated dev environment on the GPU box: PRs deploy to api-dev.optii.eu, tags to prod.
- v1.6
chut.app platform
- Gateway HMAC trust path and the OpenAPI tool manifest: optii.ocr_extract, optii.ocr_classify, optii.health_probe.
- v1.5
Classifier v2
- XLM-RoBERTa base, 32 Belgian document types, confident flag and top-k predictions on POST /ocr/classify.
- v1.4
Structured extraction
- ?structured=true: Belgian invoice fields with VAT, IBAN, structured communication and totals checks.