PDF を要約

このツールでできること

This extracts the document's text in your browser, sends it to the AI provider you choose using your own key, and writes the answer out as a Markdown file. The file itself never leaves your device, and the panel shows you the exact text and its size before anything is sent. There is no CekPDF server in that path, which also means no CekPDF account and no CekPDF bill - your provider charges you directly.

Use it when a document is longer than the time you have for it: a hundred-page tender you need to decide whether to bid on, minutes from a meeting you missed, an annual report where you only want the shape of the thing, or a contract you would like a map of before reading the clauses properly.

仕組み

1. Drop the PDF onto this page. Its text is extracted and shown to you, with a character count and a rough token estimate.
2. Choose your provider and paste your API key. The key is kept in this browser's local storage and sent to nobody but the provider it belongs to.
3. Pick a length - roughly 120 words for brief, 350 for standard, 900 with a section per topic for detailed - and a style: bullets, paragraphs, or a nested outline.
4. Set the output language if you want something other than the document's own, and narrow the page range if only part of it matters.
5. Press Summarise. The answer streams in as the model writes it, and Cancel stops the request.

Long documents are summarised in two passes. The text is cut into chunks of about 60,000 characters, each chunk is summarised on its own, and a final request writes one summary from those notes. The alternative most tools take - truncate to fit and summarise whatever survived - produces a confident account of the first thirty pages that reads exactly like an account of the whole document. This one costs more requests and reports how many it made.

There is a ceiling on how much text goes out, 120,000 characters by default, and it exists to stop someone accidentally posting a 900-page book along with the bill that would come back. When a document exceeds it, the text is cut there and the summary file opens with a line saying so - in the output you keep, not in a notification that disappears. Raise the limit under advanced options if you would rather pay for the whole thing.

The instruction the model receives is deliberately narrow: use only what the document says, note where it is unclear rather than filling the gap, and skip the preamble about being an assistant. Temperature is fixed at 0.2, low enough that the same document twice gives close to the same summary. None of that makes a summary trustworthy by itself, so treat it as a way into the document rather than a replacement for the paragraph you are about to quote.

This is one of the few pages here with an amber badge instead of a green one, because something does leave your device. To be exact: the file does not, the images in it do not, and the extracted text does - posted by your own browser, straight to the endpoint you named, with your key in the header. Nothing of ours sits between the two, so there is nothing on this side that could log it or train on it.