Document splitting is the step that separates a single multi-document file into its individual documents by detecting where one ends and the next begins. In commercial real estate, a due diligence PDF often stitches a lease, an estoppel, and an amendment into one file, and document splitting divides it into distinct documents before extraction runs.
How Does Document Splitting Work?
Document splitting works by detecting logical document boundaries from content, layout, and structure, not from page numbers or blank pages alone. Systems flag boundaries using visual cues like headers and page breaks, content patterns such as a new party name or a fresh document number, and structural resets where the layout returns to a first-page format, per Hyperscience and Extend.
The task has three parts, per the DocSplit benchmark from Amazon Science: identify boundaries, classify each resulting document's type, and keep the page order correct within each document. Early systems used trained machine-learning classifiers plus domain rules that evolve over time, per US patent literature on document separation. Newer approaches apply large language models, with the DocSplit benchmark evaluating Claude Sonnet 4.5, Claude Haiku 4.5, DeepSeek, Gemma, and Qwen across five datasets of increasing complexity.
Boundary signal | Example in a CRE packet |
Visual cue | A new letterhead or a page break before an estoppel |
Content pattern | A different tenant name or a new document title |
Structural reset | Layout returns to a cover-page or first-page format |
Document number | A fresh amendment or exhibit number appears |
Why Document Splitting Matters
Document splitting matters because extraction assumes it is reading one document at a time. Feed a pipeline a 90-page bundle of a lease plus three amendments and an estoppel as if it were a single lease, and field extraction binds the wrong amendment's rent to the base lease. Splitting first is what keeps each document's values from bleeding into another's.
Getting the split right also decides review cost. When a system correctly separates and classifies each document, extraction and its confidence scores apply cleanly per document. When it misses a boundary, an entire downstream record is wrong in a way that per-field confidence will not catch, because the field was read correctly but assigned to the wrong document. The DocSplit benchmark scores this precisely with page-plus-split-plus-order accuracy, per Amazon Science.
Example
Document splitting is graded page by page. An analyst uploads a 30-page diligence file that actually contains 3 documents: a 20-page lease, a 6-page amendment, and a 4-page estoppel. The system must place each page in the right document and preserve order.
Metric | Value |
Total pages | 30 |
True documents in the file | 3 |
Pages placed in the correct document | 28 |
Page-plus-split accuracy (28 / 30) | 93.3% |
Misplaced pages | 2 |
Documents affected by the 2 errors | 2 |
The system correctly assigns 28 of 30 pages, a 93.3% page-plus-split accuracy, but misplaces 2 pages: the amendment's first page gets grouped with the lease. Even at 93% page accuracy, 2 misplaced pages corrupt 2 of the 3 output documents, because a boundary error contaminates both the document that loses a page and the one that gains it. The page counts are illustrative inputs; the accuracy figure is derived from them. This is why splitting accuracy is measured at the document level, not just the page level.
Variations and Edge Cases
Document splitting difficulty rises with how ambiguous the boundaries are. A packet of clearly different document types is the easy case; near-identical consecutive documents are the hard one.
Variant | Treatment |
Distinct document types | Lease then estoppel; strong type and layout cues, high accuracy |
Repeated same-type documents | Several similar amendments back to back; boundaries blur |
Continuation pages | A page with no header that belongs to the prior document |
Embedded exhibits | Exhibits mid-document that should stay attached, not split off |
Scanned bundles | No digital page structure; boundaries inferred from content alone |
Document Splitting vs Document Classification
Document splitting is often confused with document classification, but they solve different halves of the same problem. Document classification labels what a document is, such as lease or estoppel. Document splitting finds where each document starts and ends inside a multi-document file. Splitting draws the boundaries; classification names the pieces.
They run together on a bundled file: the DocSplit benchmark treats boundary detection, type classification, and page ordering as one combined task, per Amazon Science. A pipeline that classifies without splitting labels the whole bundle as one type; a pipeline that splits without classifying knows where the cuts are but not what each piece is. Both are needed before extraction.
Frequently Asked Questions
What is document splitting in intelligent document processing?Document splitting is the step that separates a multi-document file into its individual documents by detecting where one ends and the next begins. In CRE, it divides a bundled diligence PDF into distinct documents like a lease, an amendment, and an estoppel before extraction runs.
How does a system detect document boundaries?A system detects document boundaries using visual cues like new headers and page breaks, content patterns such as a new party name or document number, and structural resets where the layout returns to a first-page format, per Hyperscience and Extend. It does not rely on page numbers or blank pages alone.
What is the difference between document splitting and classification?Document splitting finds where each document starts and ends inside a multi-document file. Document classification labels what each document is, such as lease or estoppel. Splitting draws the boundaries; classification names the pieces. Both run before extraction, per the DocSplit benchmark.
Related Terms
Intelligent Document Processing
Document Extraction
Optical Character Recognition
Field Extraction
Confidence Score