Unlimited OCR 2 parses a 59-page PDF in one request, but default settings corrupt tables

Unlimited OCR 2, an MIT-licensed document parser from Baidu's team, can process an entire PDF in a single request instead of page by page. Tested on one RTX 4090 with the 59-page IRS employer's tax guide, the default tile-based mode produced severe errors on a table spanning pages 46 to 52, including runaway repetition. Switching to crop_mode=false, which uses only the overview image, was faster and far more accurate. The server also silently truncated output at about page 27 because of a context-length flag, leaving no client-side error. The author recommends counting pages in the output and checking table figures in every run.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)
Log in to join the discussion and vote.
Log in