Trò chuyện với một PDF

Công cụ này làm gì

This puts a document's text in front of a language model and lets you ask it questions. The text is extracted in your browser and shown to you before anything is sent; the file itself never leaves your device. Every question and every answer is written into a Markdown transcript you can download, so a conversation is something you keep rather than something that vanishes with the tab.

It suits documents you want an answer out of rather than a reading of: a lease where the notice period is somewhere in the small print, a tender with a deadline buried in eighty pages, minutes you need one decision from, or five supplier contracts you want to compare on the same clause.

Cách hoạt động

1. Drop up to five PDFs onto this page. Their text is extracted and shown, with the total that will be sent.
2. Choose your provider and paste your API key. You only do this once - the AI tools share it.
3. Leave cite page numbers on. The text is sent with page markers and the model is told to reference them.
4. Type your question and press Ask. The answer streams in as it is written, and a follow-up carries the earlier turns with it.
5. Download the transcript, which holds every question, every answer and the citations.

Citation is the point of the design rather than a feature bolted onto it. A model handed a document will answer confidently whether or not the answer is in there, and nothing about the prose tells you which kind of answer you received. So the extracted text carries a marker at the start of every page, the instruction is to cite a page for every factual claim in the form (p. 12) and to say plainly when the document does not answer the question, and the transcript preserves those references. One answer you can check on page 14 is worth several that read well.

Several documents share one character budget rather than each receiving the full allowance. Ask across five files and each contributes a fifth - 30,000 characters apiece out of 150,000 - because the request that goes out is the combined one, and dividing it is what stops a five-file question quietly costing five times what you agreed to. Each document is labelled with its file name in what the model sees, so an answer can tell you which one it came from.

When a document is longer than its share, the surplus is not sent, and the transcript opens with a line saying the answers do not draw on the whole document. That line is in the file you download rather than in a message that disappears - a month later, reading the transcript again, the caveat is still attached to the answers it applies to.

Creativity defaults to 20%, low enough that the same question twice gives close to the same answer; raising it loosens the wording rather than improving the reasoning. None of that fixes the real risk, which is that a model can misread a table, miss a clause contradicting the one it found, or cite a page that says something adjacent to what it claims. Open the page it cites.