This is a strange thing to be impressed by, but a video of a ScanRobot doing 2,500 pages an hour got me. I watched the whole thing twice.
A camera goes down into the gutter of the book, a vacuum pulls the page flat against a glass prism, then a puff of air flips it. The book only opens 60 degrees so old bindings don't crack. They run 14th-century material through it.

Last winter, I had a 40-page document with encoded fonts and lost two days to it. The table of contents came out as symbols. Three passes with different OCR settings before I had text worth anything. Forty pages.

But the scanner isn't really the impressive part. Google was doing this at industrial scale in 2005. What's different now is what you get at the end.
Back then, you got a PDF. You still had to read it or already know which page you wanted.

Now you OCR it, split the text into chunks, embed them, and ask a question the way you'd ask a person. You get back the relevant paragraph with the page number attached. That's not really search anymore. Search assumed you already knew what you were looking for.

So I keep thinking about everything that's technically preserved but practically unreachable. University collections. Twenty years of engineering test reports and failure analyses. Journals that stopped at microfilm and never went further.
None of it is lost. It's just not findable, which for most purposes is the same thing.

The catch is that OCR is still bad at exactly the material you'd want most: handwriting, stamps over text, fax artifacts, two languages on one page. That's where my two days went, and a fast scanner doesn't solve it.

That's mostly changed in the last year. Layout and OCR run as one model now instead of a stack of stages glued together. Marker 2 does about 2.9 pages a second on a single GPU, Docling is in the same range. That's roughly four times faster than the robot turning the pages. The software isn't the bottleneck anymore, now its just a paper. #RAG #OCR #DocumentAI #Digitization #AIEngineering