Technical 21 October 2026 3 min read

A Stack of Receipts, Now Sorted Receipt by Receipt

Receipts are the paperwork almost everyone keeps and almost no one enjoys sorting. Scanning a whole pile in one go should leave you with one file per receipt. Until today, PaperSweep sometimes got that wrong. Here's what we changed.

Diesen Beitrag auf Deutsch lesen →

TL;DR — When you scan many single-page receipts together, PaperSweep now checks the date on each page, whether a page looks like the start or the end of a receipt, and asks the AI when it isn't sure. In our test, a 19-page stack of receipts that used to come back as 8 files now comes back as 19, one per receipt.

What went wrong

We scanned a stack of 19 receipts ourselves: fuel, groceries, a car wash, a café. Several came from the same shop on different days. PaperSweep returned 8 files instead of 19. Some files held receipts from different shops; one combined three receipts from the same supermarket, months apart.

The reason: PaperSweep was built mainly for letters, contracts and statements. Those usually say "Page 2 of 3" or repeat a contract number, so it's clear which pages belong together. A receipt has none of that. When PaperSweep found no clear sign of a new document, it assumed the page belonged to the previous one. To make it worse, every receipt from the same shop carries the same tax number, and PaperSweep mistook that for a document reference.

What PaperSweep looks at now

  • The date. Two neighbouring pages with different dates are very likely two different receipts.
  • How a page begins and ends. A shop name, address and phone number at the top look like a fresh start. A total, the VAT line or "thank you for your purchase" at the bottom look like a finished receipt.
  • A second opinion when it's unclear. If these signs disagree, for example on a page with no date, PaperSweep asks the AI model it already uses for sorting to compare the end of one page with the start of the next and decide whether they belong together.

Multi-page documents are still handled as before: "Page 1 of 3" still counts for more than a date. We re-ran an earlier test scan with letters and an insurance policy, and it still comes back as the same four documents.

A bonus fix: text inside tables

Receipts often print their content as a table. On some pages, PaperSweep was only seeing an empty placeholder where the table was. One receipt in our stack looked almost blank because of it. PaperSweep now reads the table text too. That also helps it name and categorise receipts more accurately.

Try it with your own pile

The update is live on papersweep.eu. If a scan still comes back grouped the wrong way, let us know. A short description of what you expected helps us a lot.