Split and sort a scanned stack of documents
One scanner run, forty letters, one PDF. Where does letter two begin? And on Monday someone asks for one contract out of three thousand files called scan_0001 to scan_3000. How long do you search?
Read one page of your stack free and see the top lines with their boxes, the signs a cut rule looks for.
How to find where a new scanned document starts
The boundary is written on the pages themselves.
- A letterhead or logo area at the top.
- Numbering such as "Page 1 of N".
- A change of sender or of date.
Which means: if every page comes back as text with the place of each word, your rule can look at the top of a page and compare it with the page before.
Split a scanned PDF into separate documents in five steps
- Send the stack and ask for
format=json. A dry run counts the pages first and reads nothing. - From each page, take the top lines with their boxes.
- Search them for "page 1" patterns and for dates.
- Compare the top of a page with the top of the page before it.
- Set the cut where the signs add up, and send unclear cases to a person.
The cut rule is yours; the service delivers the text and the places it needs. Separator sheets with a barcode or QR code are cut on the printed text next to the code.
Sort and name scanned documents by what the page says
A date in the letterhead means something else than a date in the footer, so the place of a word matters as much as the word.
Turned pages upright
A pipeline → with page orientation names the turn of each page and reads it upright.
Header and footer apart
Each word has a box, so a date in the header and a date in the footer are two different hits.
- Ask for
format=jsonwith the pipeline that corrects page orientation. - Take the header area of each page and look for date and amount patterns. Keep the box of each hit.
- Let your rule, or your own model, pick the folder or the file name.
- Send pages with unsure words inside the pattern to a folder for review.
Turned pages come back upright, and every page has a status
Scanners feed pages upside down. Of 240 turned pages, the pipeline with page orientation named the right turn for 231 and brought 223 back upright.
Every page of a job has a status: done, failed or pending. A page that could not be read is named in the header X-Pages-Failed and costs nothing, so the stack you cut is complete or tells you what is missing.
Source: own run, 80 public pages, each in three rotations. Read more: rotated pages and word boxes
Split scans with your own rule or a language model
The rule (steps 3 and 4) and the naming scheme are yours. A model can decide where documents end when it runs on top: a language model on the OCR. Documents stay in your network when the same pipelines run on your own server.
Try splitting and sorting scans free
Read one page of your stack free, or open the OCR demo gallery to see turned and mixed pages read with the text a cut rule needs.
Production use, high volume or your own server?
Private use and tests are free. For a company in production, a server of your own or a pipeline built on your pages, send the kind of page and the volume, and get a written offer, or a clear no when the fit is not there.
