Convert PDF to text

Where the words come from, and what order they arrive in

The characters in a PDF are not stored as text. They are stored as glyph numbers pointing into an embedded font, and turning those back into letters depends on a table inside the file that says which character each glyph represents. Most writers include it. When one does not — and cut-down subsets produced by some design software are the usual culprit — the page looks perfect on screen and extracts as gibberish, because the information needed to read it was never written down. No extractor can fix that; it is missing data rather than a hard problem.

Reading order is the thing to check before trusting the output. Text comes out in the order the page draws it, which is the order the software that made the PDF happened to emit — and that is frequently not the order a person reads. A two-column page often extracts as a left column and a right column correctly separated, or as lines interleaved between them, depending entirely on how it was generated. Headers, footers, page numbers and sidebars land wherever they were drawn, usually in the middle of the prose.

A handful of small artefacts are worth recognising rather than puzzling over. Ligatures come through as the single character the font actually used, so "find" may arrive as one glyph rather than f followed by i, and a naive search for "find" then misses it. Words hyphenated at a line break stay hyphenated, because the break is real in the file. And straight quotes are frequently curly ones, which matters if the text is heading into code or a CSV.

A scanned page yields nothing at all, and the tool says so rather than handing back an empty file — there are no characters in a photograph of a document, only pixels. The same is true of text converted to outlines, which designers do deliberately so that a file prints identically everywhere; it is genuinely artwork by then. Both need OCR rather than extraction.

When you want something else

When the columns matter, Poppler's pdftotext -layout in.pdf out.txt is better than anything a browser does: it reconstructs the visual spacing with real whitespace, so columns stay columns and a table stays a table you can read. pdftotext -raw goes the other way and gives the draw order untouched, which is occasionally what a script wants.

For text out of hundreds of files, mutool draw -F txt and Python's pdfplumber are what to build on — pdfplumber in particular hands you every character with its coordinates, font and size, so you can filter out the header and footer by position instead of by guesswork. That is the job this page cannot do, and it is the one that matters at volume.