Keep a bundled PDF from becoming incorrectly separated business records. Verify document boundaries, page ownership and deliberate exclusions before routing anything for extraction.
Two invoices arrive in one PDF, followed by delivery notes. A classifier labels every page correctly as either an invoice or a delivery note. That still does not establish where the first invoice ends, whether the next invoice is a separate instance, or which delivery note belongs to which transaction.
AI document splitting identifies the logical documents inside a packet. The release decision is not simply “did the model find the right document types?” It is “does each proposed document contain the right source pages, in the right order, with no unexplained loss or borrowing?”
This guide proposes a review-first design for operations teams processing bundled business documents. The rules and example are application recommendations, not a customer implementation or a measured model benchmark.
Separate classification, boundaries and business relationships
Classification assigns a type. Boundary detection separates document instances. Business matching connects those instances to a request, order or other record. Treat them as three decisions: two consecutive invoices can share a class while remaining different documents, and an adjacent delivery note does not automatically belong to either one.
The DocSplit research benchmark treats packet processing as identifying document boundaries, classifying document types and preserving page order. Its examples include interleaved and out-of-order pages. That supports testing more than page labels; it does not establish a safe release threshold for your business.
Start with a bounded input contract: permitted PDFs, supported document classes, maximum packet size and an explicit policy for covers, blanks, unknown pages and interleaving. A packet outside that contract should remain visible for manual handling, not be forced into a plausible partition.
Check what the selected service actually returns
Google’s custom splitter documentation distinguishes predicting page groups from physically separating the file. It recommends human review before splitting, while describing alternatives that depend on business requirements and observed errors. A prediction is therefore an input to your acceptance process, not the final exported artifact.
Microsoft’s Document Intelligence custom classifier documentation says the v4.0 2024-11-30 API defaults to splitMode=none; use auto when automatic document splitting is intended. The SplitMode reference distinguishes that from perPage, which treats every page as a separate document. Verify the actual version and request settings rather than assuming a classifier separates packets.
Use a page-accounting release rule
Retain the original packet in approved storage, with its identity, byte hash, revision and observed page count. Systems enumerate the pages. AI proposes document instances, ordered page membership, types and uncertain boundaries. Deterministic validation checks accounting and supported structure. A reviewer resolves meaning and approves the grouping; separate authorization controls any business action.
Use source page positions, not printed page numbers, as the primary references. Several invoices can each contain a printed “page 1.” Preserve printed numbers as evidence, but do not use them as globally unique identifiers.
Let P be all source page positions, D1…Dn the proposed document page sets, E explicitly excluded pages and U unresolved pages. For the simple partition policy below, require:
- Complete accounting: every page in
Pbelongs to one proposed document, an explained exclusion or an unresolved hold. - No accidental overlap: those assignments are pairwise disjoint. A page cannot silently belong to two exported documents.
- Valid references: all positions exist in this exact source revision; no duplicate page reference appears within a document.
- Order and continuity: retain source order. If your supported lane requires contiguous documents, a noncontiguous group goes to review rather than being silently reordered.
- Release eligibility: the release unit has no unresolved pages, every exclusion is accepted, and document boundaries and types have been reviewed under the chosen policy.
These rules catch missing or duplicated pages; they do not prove the grouping is semantically right. A perfect partition can still merge two invoices or cut one in half. Review the proposed joins and breaks using headings, identifiers, continuation text and adjacent page images.
If one cover applies to several documents, store it as shared packet context with explicit links under this policy. Do not duplicate it into each export merely to satisfy accounting. A business that needs shared-page exports requires a different, explicit representation and validation rule.
Normalize provider page references before checking coverage
Google’s splitter output documentation describes zero-based page references and warns that a zero-valued page field can be omitted in JSON. Interpret the provider’s documented representation before converting it to your reviewer-facing page positions. Do not discard an empty page-reference object as if it meant “no page.”
Test the adapter independently with a packet whose first page must be included. A consistent one-page offset can produce plausible groups that are wrong throughout the packet. Store the raw provider response and the normalized mapping so that a reviewer can reconstruct the conversion.
Work through an eight-page packet
Illustrative input: source position 1 is a cover; 2–3 are invoice A; 4–5 are invoice B; 6 is delivery note A; 7 is blank; 8 is delivery note B. Both invoices use the same layout. The process requires review of all proposed boundaries before exporting files.
| Source positions | Accepted grouping | Evidence or decision needed |
|---|---|---|
| 1 | Excluded from document exports; retained as packet context | Reviewer confirms it is a cover, not a substantive document page. |
| 2, 3 | Invoice instance A | Continuation evidence supports the join; the break before page 4 is confirmed. |
| 4, 5 | Invoice instance B | Distinct invoice identity despite the same class and layout. |
| 6 | Delivery-note instance A | Separate document; business linkage must be checked independently. |
| 7 | Excluded as blank | Blankness is verified; exclusion is not inferred from failed OCR. |
| 8 | Delivery-note instance B | Separate instance with its own source reference and later matching decision. |
The accepted document page sets are {2,3}, {4,5}, {6} and {8}; accepted exclusions are {1,7}. Together they account for all eight positions without overlap.
Now test the dangerous alternative: the model groups pages 2–5 as one invoice. Page coverage still passes. The reviewer must reject that merge because the boundary before invoice B matters. Conversely, splitting pages 2 and 3 into separate invoices creates two records from one source document even though both labels remain “invoice.”
Only after grouping acceptance should the system materialize the child files and verify their page counts and ordered source mappings. Retain child artifact hashes and stable document-instance IDs alongside the accepted packet revision. Class-based extraction can then operate on each accepted instance.
Matching a delivery note to an invoice remains a separate evidence-backed step. File adjacency and a successful split are not proof of transaction identity. Likewise, correct grouping does not approve an invoice or authorize payment.
Correct a boundary without losing downstream control
Use explicit states: captured → proposed → review_pending → accepted → exported → handoff_verified, with held, rejected and superseded alternatives. A captured packet can stay safely stored while its boundary decision remains open.
Before export
Give the reviewer the whole packet, proposed groups and the pages on each side of a disputed boundary. A correction creates a new grouping revision. Re-run accounting and eligibility checks before acceptance. Unreadable pages need a clearer source or authorized manual treatment, not a guessed label.
Assign held packets to a document-intake owner with a next action and escalation target. If a model or parser is unavailable, retain the packet and use the approved recovery lane. Bounded retries can address a transient processing failure; they cannot establish what an unreadable page means.
After export or downstream use
A changed boundary invalidates affected child artifacts and extraction results. Record which children are superseded and which destinations already received them. Do not silently reuse an approval for a document whose pages have changed. If a business record was already created, obtain a linked correction decision rather than pretending the old action never happened.
Persist the accepted grouping revision and pending handoff before issuing side effects. Use stable action identities within the destination’s supported idempotency contract. An uncertain handoff needs receipt and destination reconciliation before retry; it is not permission to create another child record. Reject late results tied to superseded revisions.
Evaluate packet integrity, not just classification accuracy
Build reviewed fixtures with expected document instances, page membership, order and exclusions. Include consecutive same-type documents, multipage continuations, covers, blank scans, unreadable pages, repeated printed numbers, interleaving and an unknown document class.
- Merge test: two adjacent invoices must remain separate even when their class and layout match.
- False-break test: a continuation page must remain with its original document.
- Coverage test: removing a page from the proposal must leave an explicit unresolved disposition, not silently pass.
- Adapter test: an omitted zero-valued provider page field must preserve the first page under the documented contract.
- Exclusion test: a page with unreadable text must not automatically become a blank-page exclusion.
- Revision test: a corrected grouping must block old exports and late callbacks from advancing work.
Measure exact-packet agreement against reviewed truth, false merges, false breaks and page-assignment errors separately. Define the units and denominators: for example, exact-packet agreement is fully correct packet groupings divided by reviewed packets. Track review minutes, held-packet age and cost per verified packet, including correction work.
Sample accepted packets as well as held ones. A falling review rate can reflect better predictions or a gate that stopped noticing errors. Choose release criteria from observed performance and business consequence; a confidence score is not a universal approval rule.
Where Workhint fits
Workhint’s Workflow Engine describes requiring information or approval steps before work advances and routing approvals to defined roles. Those controls are relevant when a packet needs a boundary reviewer and a recorded next step.
The workflow automation software solution describes structured intake, review stages, owners and status tracking. Use those as requirements for the operational handoff. The cited pages do not establish a native document splitter, page-accounting validator or revision-safe PDF exporter; evaluate those implementation requirements separately.
The broader document extraction workflow provides the surrounding intake-to-record process. Packet acceptance is the narrower boundary before individual document extraction begins.
Frequently asked questions
Does correct page classification mean the split is correct?
No. Two invoices can have identical class labels but require separate instance boundaries. Evaluate the grouping and page membership, not only the labels.
Should blank pages always be deleted?
No. Confirm the page is genuinely blank and apply an explicit exclusion policy. Keep its source position and disposition; failed OCR does not prove blankness.
Can interleaved documents be handled automatically?
Only in a lane explicitly designed and validated for that structure. A contiguous-document policy should hold interleaving for review rather than silently reorder pages.
Does accepting a split authorize the extracted business action?
No. Grouping acceptance establishes which pages form each document. Field validation, identity matching and authorization for downstream actions are separate gates.

Leave a Reply