Header Detection
Running titles and letterheads at the top of the page — furniture to strip, signals to use.
Header detection is the identification of the repeating elements at the top of document pages — running titles, chapter names, letterheads, logos, fax transmission stamps, "Page X" lines that some layouts place top-right — and their separation from body content. It is footer detection's twin, sharing the core method (positional priors plus cross-page repetition analysis: elements recurring at the same position with constant or patterned content are furniture) and the same downstream stakes: headers left in the content stream contaminate extraction, chunking, and search with text that belongs to the page, not the document's meaning.
The header zone carries richer signals than the footer's, which makes detection's second job — mining rather than merely stripping — more valuable here. Letterheads identify the document's issuer (the bank, the firm, the agency — a classification feature and an entity extraction in itself); running section titles reveal document structure that aids navigation and chunk labeling; fax stamps carry transmission metadata (sender, timestamp, page count) that intake workflows use; and header changes across pages are among the strongest document-boundary signals bundle splitting has — a new letterhead mid-file announces a new document more reliably than most content analysis.
The distinctions that keep detection honest: the first page's header is usually different and content-bearing (the letterhead belongs to the document's identity even as it's excluded from body text; a report's title page is content, not furniture), headers vary by section in structured documents (per-chapter running titles — furniture with structure), and top-of-page content that isn't repeated (a date line, an addressee block) must not be swept up by positional enthusiasm. As with footers, mature parsers label rather than delete — emitting header elements as typed metadata so each consumer applies its own policy: stripped from the RAG chunk, parsed by the intake router, preserved by the archival converter.
The page number, the confidentiality stamp, the running title — recognized as furniture, not content.
Before reading the words, a model has to see the page the way a human does — headings here, table there, footnote at the bottom.
Finding where one document ends and the next begins — inside the 100-page scan that arrived as one file.
Proof Perimeter runs document AI inside your own perimeter — with a provenance record on every field.
Book a demo