Document Segmentation
Cutting the content at its joints — pages into regions, files into documents, text into sections.
Document segmentation is the umbrella term for dividing document content into meaningful units, and it operates at three distinct granularities that share a name but not a method. Page-level segmentation partitions a page image into regions — text blocks, tables, figures, headers — the visual task underlying layout analysis. File-level segmentation splits multi-document files into their constituent documents — the scanned bundle problem. And content-level segmentation divides a document's text into logical sections — chapters, clauses, topics — the structural task underlying chunking, navigation, and section-targeted extraction. Which sense applies is usually clear from context, but conflating them causes real confusion in requirements and vendor conversations.
Each granularity has its own signals and failure modes. Page-level segmentation reads visual structure (whitespace, alignment, typography) and fails on dense or unconventional layouts. File-level segmentation reads discontinuities (letterheads, page numbering resets, format shifts) and fails on documents without distinctive boundaries. Content-level segmentation reads discourse structure (headings, numbering schemes, topical coherence — legal documents' clause hierarchies being the canonical hard case, where section 12.3(b)(ii) must attach to its parents) and fails when formatting doesn't mirror logic: the heading-free contract, the report whose structure lives in prose transitions.
What unifies them is consequence: segmentation defines the units everything downstream operates on, so its errors are inherited rather than caught. A table split across two regions extracts as two half-tables; a bundle boundary missed merges two customers' documents into one case; a clause severed from its exception chunk-by-chunk inverts a contract's meaning in retrieval. The engineering response is likewise uniform — segment confidence tracked and thresholded, uncertain boundaries routed to cheap human confirmation, and downstream validation (does the extraction reconcile? does the section parse coherently?) treated as a second check on cuts the segmenter got wrong.
Pixel by pixel: this is text, that's a table, there's a stamp — the map beneath layout analysis.
Finding where one document ends and the next begins — inside the 100-page scan that arrived as one file.
Before reading the words, a model has to see the page the way a human does — headings here, table there, footnote at the bottom.
Proof Perimeter runs document AI inside your own perimeter — with a provenance record on every field.
Book a demo