PDF to Markdown

Columns are found by hunting gutters: vertical strips no line crosses, spanning most of the text band. Width alone proves nothing. A single-column report with one two-cell grid has narrow gaps scattered everywhere, and a threshold on width would call each of them a column boundary, so what settles it is that a real gutter is one every body line respects and the share of height covered on both flanks decides. Running banners are matched with digits normalised away, since "Page 3 of 40" and "Page 4 of 40" are one footer. Six pages out of ten. Below that, the line stays. Grids are the honest part. Three or more consecutive lines whose fragments begin at roughly matching x positions form a table; prose does not. Boundaries cluster with a tolerance proportional to type size, and every table emerges carrying a confidence figure: the share of its lines that genuinely respect them. Below 0.6 it still gets written, flagged and raised as a warning. Inventing cells would be worse. Pictures come cropped from the rendered page at 216 dpi rather than pulled out of the original stream. That costs native resolution, and buys survival across every encoding a PDF can carry. The bundle holds one chunk per line in JSONL, each with its heading path, pages, token estimate and SHA-256, plus a map naming which page produced each block. Chunks break on block edges. Always. Splitting a paragraph mid-sentence to hit a token count yields fragments nobody reads alone, which defeats retrieval entirely.