Chat with PDF

Before the model sees the text, each page that has any is marked with its position in the file, and the model is told to cite the pages it uses as [p.N]. That position counts from the first page of the file, blank pages included, so it can differ from the printed number when a document opens with a cover or roman-numbered front matter. answer.md lists the cited pages that exist in your file, but a citation is still the model's claim: open the page and check it.

The file is most likely a scan or image-only PDF, so there is no selectable text to answer from. The page checks the file for text before sending it, and when it finds none the question stops there and takes nothing from the day's allowance. A file the page cannot open is sent on to our server, and if our server finds no text in it either, that refusal comes after the question has been charged. Run it through OCR PDF, pressing the document's language under Languages, then ask the OCR'd copy; answers are only as good as the recognised text.

Each question reads at most 41,032 characters, the question and the conversation sent with it included; the document gets what is left. Past that point, the model is told not to claim the document lacks something but to say how far it can see, and the notice above the answer says where the reading ended. For a very long file, pull the section you need into its own PDF with Extract Pages and ask that; its citations then count the pages of the extracted copy, not the original.

It starts over. Nothing carries over on our secure server from one question to the next: the conversation lives in this page, which re-sends the recent exchanges with each follow-up and uploads the PDF again. Removing the file or reloading the page starts a new conversation.

Each question is paid for out of 30,774 free characters of AI text a day, which every AI tool draws on, not this tool alone. It is counted per connection, so on a shared network other people's AI jobs count against it too. Every question, follow-ups included, is a run of its own, and one question never costs more than 10,258 characters of that pool, whether the PDF is one page or a thousand. That makes at least 3 questions a day if nothing else has used the pool. Apart from that, a service-wide daily limit on AI processing can pause every AI tool until the next day.

The provider is sent the extracted text, your question and the recent exchanges; the PDF file is not sent to it. Training on what arrives through the API is ruled out by the agreements pdf.xyz uses its providers under, and beyond those terms a provider's handling is not something pdf.xyz can control. Because the same text returns with each follow-up, the provider keeps it in a short-lived cache, measured in minutes rather than days, which lowers the cost of reprocessing it. If that is too much exposure for a document, keep it out of the AI tools.

Up to 100 MB per file.

Yes. 30,774 free characters of AI text a day for your connection. One run never counts more than 10,258 of them, however long the document is, and the allowance resets daily. Each question counts against the allowance as a run of its own, because this PDF's text and the conversation so far are sent again with every question. However long the document, at least 3 chat questions fit in a day if nothing else draws on it. The other AI tools draw on the same allowance.

Up to 41,032 characters for each question, the question and the conversation sent with it included — roughly 6 pages of dense text, or about 23 of sparsely-set pages. The document gets what the conversation leaves; a longer one is read up to that point only, and the answer tells you which page it reached rather than treating the rest as absent.

The recent ones. Each new question is sent with up to the last 6 questions and answers — fewer when they are long — and the panel says when older ones are left out or a long message is shortened.

Yes. The result is written by an AI model from the PDF's selectable text, and a model can miss or misread details. Check anything you rely on against the document itself.