Summarize PDF

Each point of the summary carries the sentence it came from and its page number, and anything the document does not back up is marked as unsupported.

Open in PDF ARENA

Every point cites

A verbatim quote and its page number, or an explicit label saying no backing was found.

Checked here, not claimed there

The quotes are searched in the text this browser extracted, so a made-up one gets no page.

You decide before anything leaves

The document stays put until you accept. What gets sent, where, and for how long is stated first.

A summary you can check

Ask any model to summarise a contract and it produces something perfectly readable. Some of it will not be in the contract. That is not a bug you can prompt your way out of, and a summary you have to re-read the original to trust saves nobody any time. So the deal here is different. Each point must arrive with a verbatim quote. Then that quote is searched in the text this browser pulled out of your PDF, page by page. Found it? The point gets that page number. Not found? The point gets a red label saying so, and stays on screen instead of being quietly dropped. Three numbers sit at the top: points with a located quote, points whose quote was not found, points that came with no quote at all. That is the whole trick.

How to summarize a PDF with citations

1

Open your PDF

The text is extracted here, on your device. Nothing has left yet.

2

Read the consent panel

What gets sent, to whom, and what happens to it afterwards. Then choose the kind of summary and how many points.

3

Let it run

The document goes up, the summary comes back, and every quote gets searched against your local text.

4

Look at the red ones first

Points whose quote was not located are exactly the ones worth checking. Download the Markdown or the JSON with the whole verdict.

How the checking works

The request asks for one point per line, each ending in a literal quote between angle quotes, eight to forty words, and it tells the model to write no point it cannot quote. That part a prompt can do. The rest is arithmetic. Each quote and each page get normalised: lowercase, accents stripped, punctuation gone, whitespace collapsed. An exact match after that is a hit. Failing that, a sliding window the size of the quote runs across the page and word-trigram overlap gives a similarity figure. Above 0.82 it counts. Below, it does not. Quotes under twenty-four characters are refused outright: a short phrase matches somewhere in any document by chance. Word trigrams rather than characters, and measured against the quote rather than the union. The page is far longer than the quote, and a union would drown every real match. Document text is data, never orders. The request says so, and separately the pages get scanned for the usual attack phrases: ignore your previous instructions, reveal your system prompt, send this document to. Finding one changes nothing about how the summary was produced. It gets reported, because somebody built that PDF to manipulate whoever opened it.

Why this one

Each point carries a quote, and each quote is searched in your own document.

Nothing is sent until you read what gets sent and press accept.

Attempts to instruct the model from inside the PDF are reported to you.

Questions about summarizing PDFs

Short answers, limits included

Does my document leave my computer?
Yes. Three of the 35 tools do it: this one, chat with a PDF, and translate a PDF. It says so on screen before anything moves: the file goes to our server, the relevant fragments go to DeepSeek, a provider outside the European Economic Area, our copy is deleted afterwards, and the quote check happens back in your browser. Press cancel and nothing left.
What does a red point mean?
That the quote attached to it was not found in your document. It does not prove the point is false, only that nothing here backs it. Those are the ones to check by hand.
Can I summarize only part of a document?
Yes. Pick the pages and the request is limited to them. The quote check still runs against every page, so a quote pulled from elsewhere is still caught.
Does it work on scanned PDFs?
No, and it stops rather than pretending. Without a text layer there is nothing to check quotes against, which would leave you with an unverifiable summary. Run OCR first.
What if the PDF contains instructions aimed at the model?
The document is passed as content, and the request states plainly that anything inside it is data. On top of that, the pages get scanned for known attack phrases and whatever turns up is shown with its page number.

Updated on

Do more with your PDFs

Discover all the professional tools we have prepared for you. All in one place.