Summarise a PDF
Pull the key sentences out of a long document, on your own device.
Or drop one anywhere on this page
What to know about Summarise PDF
Find the sentences that carry a long document's substance, ranked by how central they are to the rest of it. Every line comes out of the document word for word — this selects, it does not write — so there is nothing invented to check. It runs on your device and nothing is uploaded.
The honest description of this tool is that it selects rather than writes, and that is worth putting first because it is not what most people expect from the word "summarise" any more.
What it does is rank the sentences of your document against each other. Every sentence becomes a node in a graph, connected to every other by how much vocabulary they share, and then the ranking that made search engines work on links is run over that graph: a sentence is important when it resembles other important sentences. It sounds circular and it converges, and what it converges on is the part of a document carrying its recurring subject matter rather than its asides. In a contract that means the payment, term and liability clauses. In a quarterly report it means the revenue figures and the conclusions, not the boilerplate about forward-looking statements.
Two smaller pieces of work matter more than the ranking. Text extracted from a PDF is not prose — it is lines that were positioned on a page, so it arrives with breaks in the middle of sentences and words hyphenated across them, and a sentence splitter fed that raw produces fragments. And running headers are a genuine hazard rather than an untidiness: a line printed on every page shares vocabulary with the entire document, so it is the single most connected node in the graph, and the first version of this tool cheerfully summarised a contract as its own title repeated four times. Lines that repeat across most pages are found and removed before anything is ranked.
The obvious question is why this is not a language model, and it deserves a real answer rather than a shrug about privacy.
There are two ways to have a model write your summary. The first is to send the document to somebody's server. That is the one thing this site promises never to do, and the promise is not decorative — the documents people want summarised are contracts, medical letters, case files and reports under embargo, which is precisely the category that should not be posted to a third party to save someone ten minutes. The second is to run a model in your browser. That is genuinely private, and it costs hundreds of megabytes of download before anything happens at all, on a site whose tools start working in under a second; and a model small enough to be worth downloading produces summaries measurably worse than the sentences already sitting in the document.
Bring-your-own-key was considered and rejected for the same reason as the first option. It would be easy to add a field for an API key and it would produce better summaries. It would also mean that one page on a site built around a single unambiguous promise quietly became an exception to it, and a visitor would have to notice. A privacy guarantee with an exception is not a privacy guarantee.
So the tool does the thing that can be done without qualification, and gets one real advantage out of the constraint: **nothing is invented**. Every sentence is in your document word for word. A generated summary has to be checked against the source, because a model that has misread a clause will state the misreading fluently and with no indication that anything is wrong. This cannot misstate what the document says, because it only ever quotes it.
How to summarise a PDF
- 1
Add the document
Choose a PDF or drop it here. The text is read on your device and no copy is sent anywhere.
- 2
Choose how much you want
Anything from a few sentences to a page. More sentences means more detail and less compression.
- 3
Read the key sentences
They appear in the order they occur in the document, so the summary reads as a sequence rather than a list.
What Summarise PDF accepts and produces
- Input
- One PDF containing real text. A scan has no text layer until OCR has been run on it.
- Output
- The most representative sentences from the document, in the order they appear.
- Limits
- It selects existing sentences rather than writing new ones, so it cannot paraphrase, draw conclusions or answer questions about the document. It reads English well; other languages work but less reliably, because the ranking uses English stop words.
- Offline
- Yes, entirely, after your first visit.
Why Summarise PDF here
- Nothing is invented — every sentence is in the document word for word
- No account, no key, no credits and no per-page charge
- The document is never uploaded, which is the point for anything confidential
- Running headers and page numbers are removed, so they cannot crowd out the content
- Works instantly on a long document, with no model to download first
Where your file goes
Nowhere. This tool runs inside your browser, and the page is served with a policy that forbids it from sending your document anywhere. You can confirm that yourself: open your browser’s developer tools, watch the Network tab, and process a file.
Questions
Does this use AI to write a summary?
No, and it is worth being direct about that because it is what most people expect from this button. It ranks the sentences already in your document and returns the most central ones, word for word. It cannot paraphrase, cannot draw a conclusion the document does not state, and cannot answer a question about it.
Why not use AI, then?
Because both ways of doing it break something. Sending your document to a language model means uploading it to a company's servers, which is the one thing this entire site promises never to do — and a confidential contract is exactly the sort of document people want summarised. Running a model in your browser instead means downloading hundreds of megabytes before anything happens, for output that at that size is noticeably worse than the sentences already in the file. Neither trade is worth making, so the tool does the thing that can be done honestly and says so.
Can I use my own OpenAI or Claude key?
No, and this was considered properly rather than dismissed. Adding a key field would be easy and would produce better summaries. It would also mean this page starts sending the text of your documents to a third party — and a visitor who has read that nothing here leaves their device would have to notice that one page is now an exception. A privacy guarantee with an exception is not a privacy guarantee, so the answer is no, permanently.
How does it decide which sentences matter?
It builds a graph where every sentence is connected to every other by how much vocabulary they share, then runs the same kind of ranking that search engines use on links. A sentence scores highly when it resembles other high-scoring sentences, which selects the parts of a document carrying its recurring subject matter rather than its asides. In a contract that surfaces the payment, term and liability clauses; in a report, the figures and the conclusions.
Will it work on a scanned document?
Not until the scan has been through OCR. A scanned page is a picture with no text layer, so there is nothing to read and nothing to rank — the tool counts the pages with no text and tells you when this applies rather than returning an empty summary. Run OCR PDF first and summarise the searchable version.
Does it work in languages other than English?
It runs, and the results are weaker. The ranking discards a list of common English words before comparing sentences, so in another language those words are still counted and every sentence looks more similar to every other than it really is. The sentence splitting also assumes Western punctuation. Treat a summary in another language as a rough guide rather than a reliable one.
Why is a heading or a caption in my summary?
Because it shares vocabulary with a lot of the document, which is exactly what the ranking rewards. Repeated page furniture is removed automatically — a line appearing on most pages is dropped before ranking begins — but a one-off heading can still score well. Asking for more sentences usually pushes it into better company.
Is my document uploaded to be summarised?
It is not. Reading the text and ranking the sentences both happen in a worker inside this tab, and there is no model, no key and no request. Open your browser's network panel and summarise a file — nothing goes out, and with the network switched off it works exactly the same.