The AI Generic Split action examines a multi-page stack of any document type, works out where one document ends and the next begins, and splits the stack into a new Record per document. It is the general-purpose sibling of AI Invoice Split: instead of being locked to invoices, it takes its sense of what a document looks like from the Record Type you point it at.


Use it for a scanned batch that contains many separate documents of the same kind — a stack of statements, applications, remittances, tax forms, or delivery notes fed through the scanner in one pass.


AZURE ONLY: This is an AI action and runs only on Azure-hosted / AI-licensed sites.


How the split is found

The action looks at the pages and watches the identity fields you name. When a page asserts different values for those fields, that page starts a new document; pages that agree, or that say nothing about them, continue the current one. This is the one thing the AI cannot work out on its own, and it is the only setting the action requires.


Identity fields are field names, not the values to look for. This is the opposite of the Auto-Split family, where you list the values printed on a separator sheet. Nothing has to be printed on the page for the split to work except the identity values themselves — no separator sheets, no barcodes, no page counts.


Use more than one identity field when a single one is not unique on its own. For a stack of statements, Account Number alone repeats across months, so Account Number plus Statement Date is the honest identity.


A Record Type is not required. When one resolves, its field descriptions sharpen the prompt; when it does not, the identity field names alone are enough.


Options

The action's settings are grouped into six tabs.

Document

  • New Document Cat/Sub/Name — The Category, SubCategory and Name applied to each split Record's document. Leave blank to use the defaults.

The Record's active document is the one that gets split, or the Record's first document if none is active.

Identity

  • Identity Fields — Record Type field names, one per line, whose values identify a single document. This is the only required setting.

Splitting

  • Commit Split — On by default: the action creates a new Record per detected document and routes them straight away. Turn it off to stage the detected splits instead, for a later Auto-Split Commit action to commit.

Header Indexing

The AI has to read each document's header to know where the boundaries are, so the values are already in front of it. This tab keeps them instead of throwing them away.

  • Capture Header Fields — Off by default. Turn it on and the same AI call also captures the current Record Type's header fields for each document, writing them to each split Record as Index_<name>. Follow the action with Generic Indexing to move those values onto the real Record fields; turn on its Auto-Map and they are all applied for you.
  • Header Field Instructions — Optional. Tells the AI how to read the fields it captures. Use it to constrain a field the Record Type already defines (allowed values, format, where it sits on the page), or to add a field beyond the Record Type by naming and describing it. Note this is a different box from Extra Instructions on the Advanced tab, which is about finding boundaries rather than capturing values.
  • Exclude Fields — Optional. Record Type header field names, one per line, to leave out of the capture. See below.

This is a token-saving option rather than a convenience. Splitting an eighty-document stack and then extracting each Record separately is eighty-one AI calls; capturing on the way through is one. Generic Indexing makes no AI call, so mapping the values afterwards is free.

Header fields only. The captures come from the boundary walk, which reads document headers, so nothing lands under IndexLine_<name> and Generic Indexing's line item mappings stay empty. A Record Type that needs line item values still runs AI Generic Extraction per Record.

With capture on, a document whose boundary the AI was unsure of gets an Index_AIWarning value on its Record, which you can search or filter on to review those Records by hand.

AI Level

  • Quality — Standard suits most business documents; Advanced costs more and is best for poor scans, dense pages, or subtle boundaries. Vision detection needs a capable model, so there is no Economy option for this action.

Advanced

  • Extra Instructions — Boundary rules specific to your documents, for example "A continuation page repeats the account number in the top-right" or "Terms and Conditions pages belong with the document before them". When in doubt the AI keeps pages together rather than over-splitting.


Exclude Fields

When Capture Header Fields is on, the split captures every header field on the current Record Type. Exclude Fields takes fields back out of that list. Enter Record Type field names, one per line, and those fields are left out of the capture entirely — the AI is never asked for them, so they cost no tokens and cannot come back wrong.


Names must match the Record Type field name exactly, although capitalization does not matter, and wildcards are not supported. Only header fields are involved here, since the split captures header fields only.


Identity fields are never excluded, whatever you list. The split finds its boundaries by watching the identity fields change from page to page, so leaving one out would break the detection rather than simply skip a value. An identity field that appears in your exclude list is still captured.


The box appears only when Capture Header Fields is on. With capture off nothing is being read, so there is nothing to exclude.


What each split Record starts with

Each new Record inherits every data value from the original Record, which is how it keeps the right Record Type and any filter values that Record Type depends on. Captured values are written under the Index_ prefix rather than under the field's own name for exactly that reason: the captures come from arbitrary Record Type fields, so any one of them could collide with a field the Record Type is relying on. The prefix is a separate namespace that cannot collide, and Generic Indexing is what deliberately moves the values across.


TIPS

  1. Choose identity fields that are printed on the page. A calculated field, or an internal status the document never shows, cannot identify anything the AI can see.
  2. Too many Records usually means an identity field is too narrow — something that changes within a single document, like a page number. Too few means it is too broad.
  3. Turn on Capture Header Fields whenever the documents are header-only. It removes an AI call per split Record. If you need line items you are running AI Generic Extraction per Record anyway, and that returns the header fields too, so the option saves nothing in that case.
  4. Set the field Header text on the Record Type. It is sent as each field's description during capture, so a clear Header improves the values you get back without any instructions at all.
  5. Use Exclude Fields to keep the capture list to what the document actually shows. Fields the header never contains cost tokens on every document and give the AI something to guess at.
  6. If the split is close but not right, describe the exception in Extra Instructions before reaching for Advanced quality — a single sentence about continuation pages usually fixes more than a bigger model does.

NOTE ON (AI): Reading the document with AI consumes tokens and may incur extra fees. Workflow tokens are billed separately from End-User or Admin tokens — check with your sales rep for more information.