Getting Paige document separation right

Pick the right separation method for each scan job, write AI guidance that describes real boundary cues, and prove the setup on a small test set before trusting large batches.

Administrators30 minutesVerified 2026-08-30

Before you begin

Access to the Paige Settings page for the scan job you are tuning

At the end of this guide your scan job will split incoming pages into the right documents — and you'll have a small, trusted test set that proves it before you run a large batch.

Files are not documents

One uploaded file is not necessarily one document. A single PDF from a scanning session can hold twenty invoices; a single emailed attachment can hold a contract plus unrelated correspondence. Paige separates pages into documents per scan job, and the method you choose changes everything downstream: document classification and field extraction both run per document, so a wrong split produces wrong types and wrong field values no matter how good the rest of your setup is.

Method	Section	Use when
File Boundaries	Simple	Every uploaded file is exactly one document
Single Page	Simple	Every document is a single page
Barcode Separation	Advanced	A barcode marks the start of each document
AI Separation	Advanced	Boundaries must be recognized from page content

Important

Single Page and File Boundaries are exclusive modes. Turning either one on automatically turns off Barcode Separation, AI Separation, and the other simple mode. With File Boundaries active you'll see the banner "File Boundaries mode active: Each uploaded file will be treated as a separate document. No additional separation will be performed." — if your files contain multiple documents, this mode will merge them, by design.

Note

For email scan jobs, the Capture page's Email Capture settings include a switch labeled Allow document separation within each attachment. It controls whether separation runs inside each emailed attachment.

Open the separation settings

Open the Settings page. Your scan jobs are listed in the sidebar under Scan Jobs.

Select the scan job you're tuning, then choose Separation in the left menu.

What you should see: a page starting with "Document separation methods determine how to identify individual documents," split into Advanced Separation Methods and Simple Separation Methods.

Turn on the method that matches your intake. For mixed multi-document files with no barcodes, that's AI Separation ("Let AI identify the start and stop of documents.").

Write guidance that describes boundaries, not one example

With AI Separation on, click Add AI Guidance (or Edit AI Guidance once guidance exists) and enter your rules in the Edit AI Guidance dialog. Saved guidance appears on the card under Current guidance:.

Guidance that names one example document generalizes poorly. Guidance that describes what a boundary between two adjacent pages looks like works far better:

Describe the first page of a new document: a fresh document number, a new date, letterhead or a title block reappearing, Page 1 of N numbering starting over.
Describe continuation pages just as explicitly: running page numbers, "continued" wording, carried-forward totals, exhibit or schedule headings that belong to the document before them.
Call out your hardest case directly. If vendors send packets with several invoices back to back, say that a new invoice number starts a new document even when the vendor and layout are identical.
Say what is not a boundary: terms-and-conditions pages, attachments, and appendices should stay with their parent document.

Build a small test set — with both kinds of mistakes covered

Paige has no built-in accuracy report for separation, so this is a working practice, not a product feature. Gather 15–30 real files and write down the right answer for each before you run anything:

Files holding several documents each — including same-vendor multi-invoice packets (these catch under-splitting, where boundaries get missed)
A few genuinely long single documents — contracts with exhibits, multi-page statements (these catch over-splitting, where false boundaries appear inside one document)
A couple of ordinary one-document files as controls
An answer key: for each file, the page number where each document starts

Important

Real batches usually contain both failure modes at once — some documents merged together and some long documents chopped apart. A test set that only covers one kind will approve a configuration that fails on the other.

Validate on the small set before trusting large runs

Run the test set through the scan job and wait for the Document Separation stage to finish.

In the Documents area, open separation review — the toggle button's tooltip reads Show Separation View.

Compare each file's result against your answer key. Press H for the Keyboard Shortcuts dialog; to fix a missed boundary, select the page where the new document should start (Shift + Click selects a range, Ctrl + Click selects specific pages) and press Enter to add a document break. Ctrl+Z undoes a change.

Tally the errors in two columns — boundaries missed vs. false boundaries added. The two failure modes have different fixes: missed boundaries usually need sharper "what starts a document" cues; false boundaries usually need clearer "what a continuation page looks like" cues.

What you should see: a near-perfect score on the small set. That's your green light for volume — if the small set isn't clean, a large run won't be either, and you'll be reviewing thousands of pages instead of dozens.

Check your answer key before blaming the model

When Paige disagrees with your key, open the actual pages before counting it as an error. In practice, a meaningful share of "model mistakes" turn out to be labeling mistakes in the key — a configuration that looks mediocre can turn out to be scoring near-perfectly once the answer key is corrected. Fix the key, re-score, and only then decide whether the guidance needs another pass.

If clean guidance still misses boundaries

Two settings people reach for — one helps, one is unrelated:

On the Advanced page, under AI Models, raise the Document Separation Model — "The AI model used to determine where one document ends and another begins." The default is Swift - Fast and economical; switch to Precision - Advanced accuracy for difficult batches. The choice is saved per scan job, so only your hard-mix jobs pay the extra cost.

Leave Fields: First Page Only (on the Processing page) out of it. That card controls whether field values are extracted from only the first page of each document — it affects field extraction, never where documents split.

Still stuck?

If your answer key is verified, the guidance describes boundary cues, and the Precision model still gets a file wrong, contact Ademero support with this compact package:

The exact file (the original upload, not screenshots)
The expected boundaries: the page number where each document should start
What Paige produced instead — which documents were merged, and which were split apart
Your current AI guidance text and the selected Document Separation Model