Split
Category: Processing
The Split activity groups the transaction's pages into separate documents. A single upload (or batch of uploads) often contains several documents, and Split decides where each one begins and ends.
Configuration
Every setting here belongs to the Split step, its own classes and guidelines included. With Class Source set to classification it reads the project's classes, which the Classify step reads too.
| Setting | Description | Default |
|---|---|---|
| Enable Splitting | When set to No, the step is passed over and the documents stay as they reach it: one per file, as Import created them. | Yes |
| Split Mode | How pages are grouped into documents (see below). | One document per file |
| Document Classes | Candidate document classes used by intelligent splitting. | (optional) |
| Split Fields | Fields that help the model decide boundaries. | (optional) |
| Guidelines | Extra instructions for intelligent splitting. | (optional) |
| Window Size | How many consecutive pages the model looks at in one go when deciding where a document ends. A larger window gives more context around a possible boundary and costs more per window. | 2 |
| Window Step | How far the window moves between checks, from 1 up to one less than the window size, so every boundary is checked and the last window ends on the last page. Changing Window Size sets the step to that maximum, the cheapest choice: each boundary is checked once. A smaller step checks each boundary in several windows. | 1 |
| Send Images | Send page images to the model. Intelligent modes only; at least one of Send Images / Send Text must stay on. | Off |
| Send Text | Send OCR text to the model. Intelligent modes only. | On |
| Send Images of Pages Without Text | With text only: still send the image of a page that has no text (a blank page, a photo), so the model can see it. See Pages without text. | Off |
| Class Source | Where the candidate classes come from: the Classify step's class list, or the split step's own separate list. Prefer the classification list unless splitting genuinely needs a coarser set. | Classification |
Intelligent splitting always analyzes every page of the transaction. There is no setting to look at only the first few pages, because a page the model never sees could not be placed in any document.
Split modes
| Mode | Behavior |
|---|---|
| One document per file | Each file added to the transaction becomes one document, keeping its page order. Documents never span files. |
| One document per page | Every page becomes its own single-page document. |
| All pages in one document | The whole transaction is a single document, in the order the files were added. |
| Intelligent LLM splitting | The model detects document boundaries from content. All pages are analyzed as one stream, so a document may span several files. |
| Intelligent LLM splitting within files | The model detects document boundaries inside each file. File borders are always document borders, so documents never span files. |
Pages always have to be grouped into documents, so there is no "no split" outcome: Import makes each file one document. A Split step that is off leaves the documents as they are, so changes an earlier step made to them (a Data Transform, or a reviewer's splits and merges before a Send To back to this step) are kept. The mode selector is shown locked to One document per file.
Per-file split policies
API clients and connectors can pin a split constraint on individual files at upload time, overriding the step's mode for just those files. A file can be marked as exactly one document, or as self-contained (one or more whole documents, analyzed inside the file only). The email import connector uses this to keep the email body separate from its attachments. See Per-File Split Policies.
Pages without text
A page with no text is a normal page: a blank page, a photo, a signature sheet. OCR reads it and finds nothing, and that is not an error. This applies to the intelligent modes, which send pages to the model.
- Some pages have text: splitting works as usual. Each page without text is marked to the model as "(no text on this page)". A blank page between two documents often helps the model see the boundary.
- No page has text, and page images are sent (Send Images on, or Send Images of Pages Without Text on): the model splits from the images.
- No page has text, and no image is sent: there is nothing for the model to read, so it is not called. Each file becomes one document, with split confidence 0, so the Human Review step flags the split for a person to check. The Activity Log names the pages.
A page that never went through OCR is different. If the step sends text and a page has no OCR result at all (there is no OCR step before Split, or OCR is off in it), and no image of the page is sent, the step fails at once and names the pages. Add an OCR step before Split, or send page images.