How it works: the recognition engine downloads the first time you use it, along with a language pack — expect a few seconds and a handful of megabytes on that first run, then it is cached and instant afterwards. Accuracy depends almost entirely on the picture: straight-on, well lit, and in focus beats a high-resolution photo taken at an angle every time. Printed text comes out well, handwriting far less so. Scanned PDFs are handled a page at a time. Nothing is uploaded — the engine runs inside this tab, so payslips, medical letters and anything else private stay on your machine.