Skip to content

SplitPDF / Scanned PDF

Split a scanned PDF into separate documents

You fed a stack of receipts, letters and forms through the scanner and got one 60-page PDF. Let a vision model mark where each document begins, correct anything it got wrong, and save one file per document.

Loading the splitter…

Tips for a clean result

  • Remove blanks first if you scanned double-sided: blank backs confuse document boundaries. Use remove blank pages, then come back with the cleaned file.
  • Upright pages help. A sideways page is still read, but labels may be less accurate.
  • Fixed-length documents (for example, every form is 2 pages) don't need AI at all: use every N pages.
  • Digital PDFs with real text are faster and private with split by text.

Why the preview images are small: 560 pixels wide is enough to see a letterhead or a "Page 1 of 3" line, keeps the request light, and sends far less detail than the original scan.

Questions about splitting scans

How do I split one scan into separate documents?

Open the scan, tick the consent box and press "Find where each document starts". A vision model looks at each page and marks where a new invoice, letter or form begins. Check the cuts in the preview, fix any by tapping the gap between pages, then press Split PDF.

Is my scan uploaded?

In this mode only, small preview images (about 560 pixels wide) of up to 120 pages are sent to our server and passed to an AI vision model through Vercel AI Gateway. They are not stored. The splitting itself, and the full-quality file, stay on your device.

How accurate is it?

It works well when documents have clear first pages: letterheads, titles, "Page 1 of 2". It can miss a boundary between two similar forms or cut a long letter in two, which is why every cut is shown for review before anything is saved.

Can I do it without AI?

Yes. If every document has the same number of pages, use every N pages. Otherwise, tap the gaps between pages in custom ranges mode. Both are fully offline.

Does it read or keep my text?

The model is asked only for page numbers and a short label per document (like "Invoice 1043 Acme"), used for the file names. Nothing else is transcribed or kept.