What Google Document AI OCR does
Enterprise Document OCR is one of the processors in Google Cloud's Document AI. You send a PDF or image (PDF, GIF, TIFF, JPEG, PNG, BMP or WebP; DOCX is in private preview) and get back a Document object in JSON with the text, pages, blocks, lines, tokens and their positions. It reads printed and handwritten text, can merge embedded text from digital PDFs with OCR results (native PDF parsing), and scores each page's image quality from 0 to 1, listing likely defects such as blur, glare or cut-off text when the score is below 0.5. Google says it handles more than 200 languages; handwriting is marked unsupported for a dozen of them, including Hebrew, Thai, Tamil and Telugu.
Setup is the usual Google Cloud flow: create a project, enable the Document AI API, create an Enterprise Document OCR processor in a region (US, EU and several others) and call it from the REST API or client libraries. Online (synchronous) requests take up to 15 pages and 40 MB (30 pages in imageless mode); batch requests take up to 500 pages per document, 1 GB per file and 5,000 files per request. Stable model versions are pretrained-ocr-v2.0-2023-06-02 and pretrained-ocr-v2.1-2024-08-07, with v2.1.1 as a release candidate outside the US and EU.
