PDF to TXT

This is the thing that makes text extraction harder than it sounds. A PDF stores instructions of the form "draw these glyphs at this coordinate". There are no sentences, no paragraphs, often no spaces — a space is frequently just the next fragment being drawn a little further right.

So extraction is reconstruction. Fragments are grouped into lines by their vertical position, sorted left to right, and joined with spaces inserted where the horizontal gap implies one. Lines are then grouped into paragraphs using the signals a typesetter actually leaves: a larger vertical gap, a short line followed by an indent, a line ending in a full stop.

The result reads as prose for a normal document. It reads less well for multi-column layouts, where the reading order genuinely is ambiguous from geometry alone, and for tables, where the visual structure is the meaning.

Reflowing destroys a table: the cells become a run of words. Switch to keep the page's line breaks and each visual line comes out as its own line, which preserves the row structure even though the columns are only aligned by spaces.

The same applies to source code listings, forms, poetry and anything where the line breaks carry meaning.

There are two entirely different things called a PDF.

A born-digital PDF was made by software — exported from Word, generated by a report tool, printed to PDF. It contains real text with real font information, and it converts cleanly in every direction.

A scanned PDF is a photograph of a piece of paper wrapped in a PDF container. There is no text in it at all — only an image — regardless of how clearly you can read it on screen.

The difference decides what is possible. Rasterising works identically on both. Extracting text works only on the first, and a scan yields nothing at all — which is reported plainly rather than handing back an empty file.

If the PDF is a scan, extraction returns nothing — and you are told that explicitly rather than being handed an empty file and left to wonder.

Reading a scan requires OCR: recognising characters in an image. That is a genuinely different job and needs a language model of ten to fifteen megabytes per language, which is not something to download on the chance a visitor needs it. The practical route is to render the pages with the PDF to JPG tool and run those images through an OCR tool.

A mixed document — mostly digital with a few scanned pages inserted — extracts the digital pages and reports how many were empty.

Everything that is not the words: fonts, sizes, bold and italic, colours, images, tables as tables, headers and footers as distinct from body text, and footnotes as distinct from the paragraph above them. Hyphenation at line ends is not undone, so a word split across two lines stays split.

Plain text is the format that keeps the words and nothing else. That is usually the point.