Multi-Column Document Parsing
Read down, not across — parsing the layouts where naive left-to-right produces word salad.
Multi-column document parsing is the correct handling of columnar layouts — academic papers, newspapers and newsletters, legal codes, brochures, statement designs — where text flows down one column before continuing at the top of the next, and where the naive left-to-right, top-to-bottom serialization that flat text extraction performs interleaves the columns into fluent-looking nonsense: half a sentence from column one spliced mid-thought into column two's content, every line. It is among the most common and most consequential parsing failures, because its output looks like text — downstream search, extraction, and RAG consume the scramble without complaint, and the errors surface as mysteries far from their cause.
The parsing requires layout analysis to precede serialization: column detection (whitespace gutters, alignment consistency, ruling lines) segmenting the page into column regions; flow analysis determining their reading sequence — usually left-to-right across columns, but complicated by the real repertoire: full-width elements (titles, figures, tables) interrupting the columns, articles that jump ("continued on page 7"), sidebars and pull-quotes that belong outside the main flow entirely, and mixed layouts where column count changes mid-page. Modern layout models handle the standard cases well; the document-specific traps (the two-column contract with full-width recitals, the statement whose columns are actually a table) reward document-type awareness — and VLM-based parsers, reading the rendered page holistically, have made columnar reading order one of their clearest wins over coordinate-heuristic pipelines.
Verification is cheap and worth automating: serialized text should be linguistically coherent (perplexity and sentence-integrity checks catch interleaving), and hyphenation should resolve at column breaks (the "docu-/ment" split rejoined). For PDF text-layer extraction the same discipline applies — the layer stores glyphs positionally, not logically, so column structure must still be inferred, and the born-digital document is no immunity against reading-order salad.
Which block comes next — the sequencing decision that turns detected columns into readable text.
The sequence a human eye would follow — inferred from layout, and essential before a word is serialized.
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