Convert documents and spreadsheets

Three different engines, depending on what you are converting

Office documents — Word, LibreOffice and their relatives — run through headless LibreOffice, the same engine behind the desktop suite rather than a lookalike parser. It reads documents the way an office application does, which is why ordinary files come across faithfully.

Spreadsheet and data conversions between CSV, JSON and workbooks are handled in our own process instead. No office suite is launched, which is faster and also safer: macros in a workbook are never in a position to run, because nothing in that path is capable of running them.

PDF pages are rendered with Poppler where the host provides it and a JavaScript fallback where it does not, so the route works on every deployment rather than only the ones that can install system packages.

Fonts are what move your layout

A document names the typefaces it wants rather than carrying them. When the converting machine does not have one, it substitutes another — and a substitute rarely occupies exactly the same width per character.

Slightly wider characters mean fewer words per line, more lines, and content pushed down the page. That is why a converted PDF occasionally has one more page than the document did, with a single orphaned paragraph on it. Nothing has broken; the text has been set in a different typeface. Sticking to widely available fonts, or embedding them before converting, avoids it entirely.

A scanned PDF has no text in it

Two PDFs can look identical and be completely different files. One created by exporting from a word processor stores characters and their positions. One created by a scanner stores an image of a page — the words are visible to you as shapes, but there are no characters in the file at all.

That decides what is possible. Text extraction and PDF-to-Word need text to work from, so on a scanned document they stop and explain why rather than handing you an empty file. Rendering pages to images works on either kind, because it does not care whether the page was made of text.