Search older scanned documents
Improved text recognition now finds headings in older scanned PDFs.
A filing cabinet full of scans is only useful if someone can find a page again. For years the recognition on older PDFs stopped at the first muddy page: a faded invoice, a map photographed at an angle, a contract where the heading was set in a thin typeface. The file was stored. The words inside it were not.
What changed
The desk now reads those pages a second time, looking specifically for headings, dates and the short labels people actually search for. It does not replace the original scan. The image stays as it was filed. What is added is a text layer you can search, and a note in the activity history that says when that layer was written.
We tried this first on collections that had already been restored once. The pattern was the same in almost every workspace: the recent born-digital files were easy to find, and the archive that justified the whole subscription was a pile of scans from 2009 to 2018. Headings such as "Board minutes", "Lease, unit 4" or "Photo index, March" were visible to a person and invisible to search.
What we leave alone
Handwriting is still handwriting. A pencil note in the margin is not turned into a confident sentence. Pages that are mostly photograph, stamp or diagram keep their preview and are marked as image-only, so a search does not pretend to have read them. If a heading is uncertain, it is stored as a suggestion rather than written over the filename you chose.
Existing filenames, folders and shares are unchanged. New workspaces use the improved reading by default. An administrator can run it on an older collection from the workspace settings, one folder at a time, and stop if the result is not what the owner expected.
How to try it on a real folder
Pick a folder that is already backed up and that nobody is editing this week. Open it, choose the recognition pass, and wait until the activity history shows the pass finished. Search for a heading you can see on the first page of a known scan. Then search for a word that only appears halfway through a longer document. Both should return the file, with the page indicated.
If a result looks wrong, the original file is still the original file. Remove the generated text layer from that document and the search entry disappears with it. Nothing is sent out of the desk to be read elsewhere. The pass runs on the copy that already lives in your archive.
Why this took this long
Earlier recognition was tuned for clean office exports. Those files already contain text, so the hard part never showed up in demonstrations. The requests that reached the help centre were about the other half of a company archive: the scanner in the corner, the borrowed flatbed, the phone photo of a noticeboard. We measured success on those pages, not on a fresh export from a word processor.
The update is on every current plan. There is no separate fee for reading an old folder. Notes about follow-up changes stay on this page, next to the date, rather than being folded quietly into a different behaviour.
Questions about a particular collection belong in the help centre. If you are deciding who should be allowed to start a recognition pass, the guide for archive owners is the shorter checklist.