Text summarizer that picks sentences, not paraphrases

No. There is no model, no API call and no account behind this page — it scores your sentences with arithmetic that has been in the literature since the 1950s and picks the highest-scoring ones. That is a real limitation: it cannot rephrase, merge two sentences or write a smoother opening. It is also the reason it cannot hallucinate a fact you never wrote, cannot quietly change your meaning, cannot send your document to anyone, and produces exactly the same summary every time you run it.

Literally, yes — every sentence in it is one you wrote, copied out unchanged and left in the order you put it in. Nothing is paraphrased or generated. That makes it safe to quote in a citation and safe to hand to a lawyer, but it also means the result can read jumpy where two chosen sentences were three paragraphs apart in the original. Smoothing the joins is the one job left to you.

It weights each word by how distinctive it is across the text, then scores each sentence by the weight of the words in it. A word appearing in nearly every sentence carries almost nothing; a word concentrated in a few sentences carries a lot, which is how the topic-bearing sentences rise. Position adjusts the result — an opening sentence, and the first sentence of a paragraph, get a bonus, because writers put their point there — and very short lines are penalised so headings and captions do not crowd out real content.

Because the scoring reads vocabulary, not grammar, and a sentence can be dense with the text's key terms while depending on the sentence before it for its subject. This is the characteristic failure of every extractive summariser and the reason the marked-up original is on the page: it takes seconds to see that a chosen sentence has been orphaned, and either to move the slider up so its neighbour comes along, or to patch the opening clause by hand.

Somewhere between 20 and 30 per cent for most documents, which is where research summaries have conventionally sat. Below about 10 per cent an extractive summary stops being prose and becomes a list of assertions, because the connective sentences are always the first to be dropped. The other way of choosing is to work backwards from the reading time you want: the panel shows how many minutes of reading the current setting saves.

Not directly — it takes pasted text, so open the document, select the part you want and copy it in. That is usually an improvement, because it means the running headers, footnote markers, page numbers and cookie banners that come with a whole file stay out of the summary. If a PDF pastes with a line break at the end of every line, the sentence detection still works, since sentences are found by their punctuation rather than by their line endings.

Extractive summarising selects existing sentences; abstractive summarising writes new ones. This page is firmly the first kind, which is the older and more conservative of the two. Abstractive systems read better and can compress much further, at the cost of sometimes stating things the source did not — a well-documented failure mode that matters most in exactly the situations where summaries get used, such as medical, legal and financial documents. Extractive output can be checked against the original in a glance, which is what the highlighting above is for.