light-ocr: extract text locally and check what it missed

I would test one invented timetable before using OCR on anything important. Keep a typed copy as ground truth, photograph the page, extract the text and compare every line. Start with ordinary type before adding glare, small print or a tilted camera. That gives you a useful error record rather than an impressive-looking box overlay.

The three things this guide helps you do are set up the documented package, extract a first image or PDF and check the mistakes. “Offline” here refers to local recognition after package acquisition; the initial installation still needs the package files.