Marcel 62be895b9e fix(ocr): drop uvicorn workers from 2 to 1 ...

Two workers × ~5 GB Surya model load = ~10 GB required, exceeding the
8 GB memory cap and causing OOM on the first /train call. Two OS
processes also cause model-state divergence after training, contradicting
the single-node constraint documented in ADR-001.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

2026-04-14 09:55:55 +02:00

..

engines

fix(ocr): use cw-1/ch-1 for synthetic baseline bounds to pass Kraken's >= check

2026-04-13 16:21:00 +02:00

.dockerignore

fix(docker): soften ocr-service dependency and clean up compose

2026-04-13 12:29:21 +02:00

confidence.py

feat(ocr): per-script-type confidence thresholds

2026-04-12 20:50:59 +02:00

Dockerfile

fix(ocr): drop uvicorn workers from 2 to 1

2026-04-14 09:55:55 +02:00

main.py

fix(training): output CoreML format and fix best-model finder

2026-04-13 21:57:42 +02:00

models.py

feat(ocr): add guided OCR mode using existing annotation regions

2026-04-13 15:57:54 +02:00

requirements.txt

fix(ocr): add pyvips for kraken PDF input support

2026-04-12 20:11:14 +02:00

test_confidence.py

feat(ocr): per-script-type confidence thresholds

2026-04-12 20:50:59 +02:00

test_engines.py

test(ocr): add business-logic tests for polygon extraction, Kraken routing, and confidence markers

2026-04-13 10:34:23 +02:00

test_stream.py

feat(ocr): add SSRF protection for PDF URL downloads

2026-04-13 12:29:42 +02:00