Extractive, not generated
An extractive summariser selects; it does not write. Every sentence it returns appears character for character in what you pasted, punctuation included. That rules out a whole class of failure: no invented date, no formula that was never on the slide, no confident sentence about a topic your lecturer skipped entirely.
The trade-off is that it can only be as good as the source. If a definition was never stated, no definition comes out. The phrasing stays your lecturer's, which sometimes reads awkwardly in isolation, but is also the phrasing the exam paper and the marking scheme are written around.
There is no language model here and nothing to prompt. Ranking is arithmetic over the words you supplied, which is why the same input always produces the same notes, and why a 4,000-word chapter comes back in well under a second.
How sentences are chosen
The tool counts how often each meaningful word appears across the whole text, ignoring the connective words that turn up everywhere. A sentence's score is the sum of its words' counts, so sentences built from the vocabulary the document keeps returning to rise to the top, and passing asides sink.
Raw totals would simply pick the longest sentences, so each score is divided by the square root of the sentence's length. Dividing by the full length overcorrects and rewards three-word fragments; the square root sits between the two, keeping a 12-word statement and a 40-word one genuinely comparable.
Selected sentences are then returned in document order rather than score order. It is a small decision with a large effect: read in score order, notes leap about, whereas in document order a definition still arrives before the worked example that depends on it.
Getting better notes out of it
Run one topic at a time. Paste an entire module and the word counts flatten out, because terms that dominate a single chapter lose to terms repeated across all twelve. Ten runs of one chapter each, at around fifteen sentences a run, give you far more usable notes than one run over 200 pages.
Clean the input before pasting. Repeated slide footers, page numbers, tutorial questions and a long reference list all add counts for words that carry no meaning in your subject, and they drag sentences containing those words upwards. Thirty seconds of deleting boilerplate changes the output noticeably.
Treat the result as a checklist rather than finished revision. Extraction tells you what the material emphasised; it cannot tell you whether you understand it. The useful step is closing the notes and rewriting each line in your own words, which is the part that shows up in the exam.