pdf-ocr-skill

PDF OCR Skill extracts Chinese and English text from scanned (image-based) PDFs and image files, packaged as a Python tool with a PDFOCRProcessor class and a command-line entry point. It solves the problem of getting selectable text out of photocopied documents, scanned contracts, books, reports, and screenshots where copy-paste does not work. It converts PDF pages to images and runs OCR, preserving reading order and structure.

It offers multiple engines selectable by an OCR_ENGINE environment variable or a CLI argument: RapidOCR as a free, offline local default that needs no API key; a RapidDoc enhanced engine adding layout analysis, table, and formula recognition; a PaddleOCR local engine; and a SiliconFlow cloud engine that calls a large model (for example DeepSeek-OCR) for high-accuracy recognition. It includes intelligent fallback, switching to the cloud API when local initialization fails, and a documented pattern for trying RapidOCR first and escalating to the cloud when the extracted text looks too short. The documentation covers installation (pymupdf, pillow, requests, python-dotenv, plus optional rapidocr_onnxruntime), environment configuration through a .env file, quick-start snippets, command-line usage, batch processing of directories, saving intermediate images, supported formats (PDF plus JPG, PNG, BMP, GIF, TIFF, WEBP), and an engine comparison covering cost, speed, accuracy, and network dependence.

The target users are developers and analysts who need to digitize scanned documents or images, individually or in batch. Credentials for the optional cloud engine (a SiliconFlow API key) are configured via environment variables in a .env file, which is standard and safe practice; the local engines run entirely offline. Choosing the cloud engine sends document images to a third-party OCR service, which is an intended feature rather than a covert behavior, so users should be mindful of document sensitivity and API costs.