Chat with a PDF: Ask Questions Instead of Reading 100 Pages

The expensive part of a long PDF isn't reading it — it's that you can't tell which pages matter until after you've read them. You pay the full cost of a 100-page report to find the four pages you actually needed, and you only discover which four at the end.

Chatting with a document inverts that. Instead of reading front to back and hoping the relevant part stands out, you ask the question you actually have — "what's the notice period?", "which section covers renewal?" — and go straight to the answer. The document stops being something you consume linearly and becomes something you interrogate.

The AI doesn't read your document the way a person does. The text is extracted from the PDF and sent to the model alongside your question, every turn — the whole document, not a search result from it. The answer is generated from that text rather than from the model's general knowledge, which is why it can quote your file and why it should say so when your file doesn't cover something.

There is a ceiling on how much text fits in one request, so a very long document is cut off at that point rather than being searched for the relevant part. That's why the answer panel reports how many of the document's pages were actually included: on a long file, "the document doesn't mention that" and "the part of the document I was given doesn't mention that" are very different statements, and you should be told which one you're getting. If a document is too long, split out the section you care about and ask about that instead.

It also explains why a scanned PDF returns nothing useful. If the pages are images with no underlying text, there's nothing to extract at all. Run those through OCR PDF first to add a real text layer, then chat with the result — the same coverage report will also tell you how many pages came back with no text, which is how you catch a scan you didn't realise was a scan.

These solve different problems and it is worth knowing which you have. A summary answers "what is in this document?" — you get the whole shape of it, condensed, without having chosen what matters in advance. Chatting answers "what does this document say about X?" — much sharper, but only for the questions you already know to ask.

In practice they work best together on the same file. Run AI Summarizer first to find out what the document contains and which sections are worth your attention, then use Chat with PDF to interrogate those sections in detail. Starting with chat on an unfamiliar document means guessing at questions before you know what is in there to ask about.

Questions that work well are the ones with an answer physically present in the document: obligations, dates, amounts, definitions, who is responsible for what, what a specific section says. The system is locating and explaining real text, and it is very good at that.

Questions that work poorly are ones the document doesn't answer. Asking whether a contract is a good deal, how this year compares to a competitor's results, or what the author should have done instead pushes the model outside its source material — which is exactly where answers get less reliable. If the information isn't in the file, no amount of prompting will retrieve it.

Treat chat answers the way you'd treat notes from a colleague who read the document before you: useful for direction, worth checking before you quote them. For anything with consequences — a legal clause, a payment figure, a deadline — open the document and read the actual passage.

This costs a minute and removes the entire risk category. The speed advantage isn't in skipping verification; it's in knowing exactly which page to verify instead of hunting for it.