PP-OCRv6

High-accuracy OCR for images and scanned PDFs, powered by Python PaddleOCR.

Usage

1. Enable PP-OCRv6 in Finch.
2. Share an image or PDF and ask:
   • Extract text from this image
   • Extract text from this PDF

Setup happens automatically on first use (Python venv + pip install).

Tools

Tool	Description
ocr_image	Start async OCR on an image (PNG, JPG, WebP, BMP, etc.) — returns a task ID
ocr_pdf	Start async OCR on a scanned PDF — returns a task ID
check_ocr_task	Poll a running task by task ID and retrieve the result when done
ocr_cache	View cached OCR results (age, expiry)
clear_ocr_cache	Clear all cached results
setup_ocr	Check environment and install dependencies
ocr_status	Quick health check

Async flow: ocr_image / ocr_pdf launch a background process and return immediately with a task ID + estimated time. Finch will automatically call check_ocr_task after the estimated time to retrieve the result.

Requirements

• Python 3.10 – 3.12
• Dependencies (paddleocr, paddlepaddle, PyMuPDF, opencv-python, numpy) auto-installed on first use

Caching

OCR results are cached for 30 days, keyed by SHA-256 file hash. Same file = instant result without re-running OCR.

extensions/pp-ocrv6/
├── cache/
│   ├── index.json         # hash → ISO timestamp
│   └── <sha256>.json      # result file

• Cache files stored in cache/ directory alongside the extension
• index.json records creation time for quick expiry checks
• Expired entries auto-cleaned on each new cache write
• Use ocr_cache to inspect, clear_ocr_cache to purge

Performance & Limits

Type	Suggested Limit	Notes
Single image	Unlimited	Automatically scaled if longest edge >3000px
PDF pages	Hundreds	150 DPI rendering, two-pass blank skip (72 DPI preview then 150 DPI), blank pages auto-skipped
Single PDF page	Up to A0	Very large drawings should be split first

• Model warmup — cold start (~10s) handled once at the start of each PDF; subsequent predictions are instant
• Sequential page processing — pages are processed one at a time with a single model instance, maximizing reliability on macOS ARM (PaddlePaddle multiprocessing is unstable)
• Dynamic time estimation — per-page moving average shown in check_ocr_task progress
• PDF output includes page headers and confidence per page
• Cache stored in extension private data directory

Privacy

All inference runs locally. Network access is only for initial model downloads (~150 MB).

License

MIT