Wire real PaddleOCR into /ocr #22

New issue

Closed

opened 2026-07-05 19:20:23 +01:00 by TimCane · 0 comments

TimCane commented

2026-07-05 19:20:23 +01:00

(Migrated from github.com)

Goal

PP-OCRv4 English det+rec models baked into the image, CPU single worker, top-to-bottom line ordering with axis-aligned boxes, the 200 shape and 422/413/500 errors including the 8000x8000 header-dimension decode-bomb check.

References

Done when

Advances M3 A1/A3 (real inference, no hang).

Part of M3: OCR pipeline.

## Goal PP-OCRv4 English det+rec models baked into the image, CPU single worker, top-to-bottom line ordering with axis-aligned boxes, the 200 shape and 422/413/500 errors including the 8000x8000 header-dimension decode-bomb check. ## References - [docs/06-ocr-service.md#stack](https://github.com/TimCane/bill-splitter/blob/main/docs/06-ocr-service.md#stack) - [docs/06-ocr-service.md#post-ocr](https://github.com/TimCane/bill-splitter/blob/main/docs/06-ocr-service.md#post-ocr) - [docs/10-security-privacy.md#upload-hardening](https://github.com/TimCane/bill-splitter/blob/main/docs/10-security-privacy.md#upload-hardening) ## Done when - Advances M3 A1/A3 (real inference, no hang). --- Part of **M3: OCR pipeline**.