Convert PDF to text

There are two kinds of PDF and they look the same from outside. One was made from a document — exported from Word, printed to PDF, generated by a system — and it contains real characters. The other was made by a scanner or a phone camera, and it contains a picture of a page. Extraction works perfectly on the first and returns nothing at all from the second, because there is nothing there to return.

You can tell in five seconds without converting anything. Open the PDF and try to select a line of text with your mouse. If individual words highlight, the characters are in the file and this page will get them out. If the whole page highlights as a single rectangle, or nothing highlights, it is an image. That case needs optical character recognition — software that looks at the shapes and guesses which letters they are — which is a different job with a different error rate, and this converter does not attempt it.

Plain text is a sequence of characters and nothing else. It has no notion of a column, a table, a heading, a page or a font. So converting a PDF to text is deliberately lossy in a way that is not a limitation to be worked around — it is what you asked for.

The consequences are worth expecting. A two-column academic paper comes out as one continuous sequence, with the left column of a page followed by the right rather than interleaved line by line. A table becomes its cells as words, without the grid that made them mean something. Headers and footers appear inline on every page. If any of that matters, you want the PDF converted to a document format instead, where the structure has somewhere to live.

A PDF does not store sentences or paragraphs. It stores instructions of the form "draw these glyphs at this position on the page", and the order those instructions appear in is usually, but not necessarily, reading order. Most documents are generated by software that writes them in the obvious sequence, and extraction is exact.

Heavily designed pages are where it breaks down. A magazine layout with a pull quote in the middle of a column, a form with fields positioned around labels, a newsletter with boxes — these can come out interleaved, because the file genuinely does not record which block is meant to be read first. Nothing can reliably fix that from the file alone; the information was never written down.

Look at what people extract text from. Contracts, tenancy agreements, medical letters, bank statements, invoices, unpublished manuscripts, papers under review. The category is unusually skewed towards documents whose contents are somebody’s business and nobody else’s.

Every other text extractor on the internet asks you to upload the file. This one reads it with pdf.js — the engine Firefox uses to display PDFs, which is open source and Apache licensed — running inside the page you already have open. The document is never sent, and you can verify that rather than believe it: open your developer tools, watch the network tab, and convert something.

This has become one of the common reasons for the conversion, and plain text is the right format for it. Models read text; they do not benefit from the fonts and positioning a PDF carries, and pasting a PDF into a chat window usually means the tool did this conversion for you, less carefully.

Two things help. Convert the whole document rather than screenshots of pages, because text is exact where an image needs recognising. And check the beginning of the output — if the first page is a cover with a title and nothing else, the text will start with very little, which is a sign the extraction worked rather than a sign it failed.

Extraction is exact about what the file says, which occasionally means being exact about something odd. A word broken across a line in the original keeps its hyphen. Some documents encode fi and fl as single glyphs, and those arrive as one character rather than two letters.

None of this is damage — it is the document being reported faithfully. A search and replace over the output handles the hyphens, and most editors and language models cope with the ligatures without being asked.

The output is UTF-8, which is what you want in essentially every case: it holds accented letters, Greek, Cyrillic, Japanese and emoji in one file, and every current tool reads it.

The one place this still bites is older Windows software, and Excel is the usual culprit, because double-clicking a text or CSV file makes it guess a legacy code page instead. Opening the file through Excel's import dialogue and choosing UTF-8 fixes it, as does opening it in almost anything else.

If the selection test showed you an image, this page cannot help, and saying so is better than handing back an empty file. Recognition is a genuinely different operation with a genuinely different failure mode: it does not fail loudly, it quietly guesses wrong.

The realistic routes are your operating system's built-in recognition — both macOS and recent Windows will pull text out of an image — or a dedicated OCR tool. Whichever you use, read the result before trusting it, because a misrecognised digit in an invoice looks exactly like a correct one.

Extraction is quick because it is reading structure rather than drawing anything. A few hundred pages is unremarkable, and the work is roughly proportional to the amount of text rather than to the size of the file.

What slows it down is a document that is large because it is full of images, since those still have to be parsed past. A 200 MB scanned brochure takes a moment and then, correctly, returns almost no text — which is the selection test giving you the same answer the slow way.