How to chat with a PDF: from first question to checked answer

Chatting with a PDF means a language model reads the text inside your file and answers questions about it. In the Chat with PDF tool you add one file, type a question, and the answer streams in. You can follow up as often as you like.

The key word is text. If you can select and copy words from your PDF, it has a text layer, and that layer is what the model gets. It doesn't look at the page the way a person does. A scanned contract or a photo of a letter has no text layer, and there's no OCR here, so the tool says "No text found" and stops.

The model isn't browsing the web or remembering your other files. It sees this document and the conversation so far. That's handy, because everything it tells you should trace back to something on a page. It's also a limit: if the document doesn't say it, the model shouldn't either, though sometimes it does anyway. Most of this guide is about catching that.

Three checks before you upload anything.

• Is it text? Open the PDF and try to select a sentence with your cursor. If words highlight one by one, you're fine. If the whole page highlights as a single block, it's an image.
• How big is it? The AI jobs take about 5.7 MB per PDF, and a PDF counts as roughly 6,000 input tokens. A 40-page policy of plain text is usually fine. A 300-page report isn't, so pick a page range (how to summarize a long PDF shows how).
• Is it password-protected? We can't remove passwords. If the file is yours to open, re-save a copy without the password and use that.

One check isn't about the file at all: should it leave your device? Chat uploads the PDF to the model provider you picked. If you wouldn't email it to an outside service, stop here.

Say you have a 40-page home insurance policy and a leaking pipe. That's an example, not a real file, but it shows the rhythm. Most useful questions fall into four kinds.

Orient. Ask what the document is before you ask anything specific. "What is this, who issued it, and what period does it cover?" You confirm that you and the model are looking at the same thing, and you get a map for later questions.

Locate. Ask where something lives. "Which section covers water damage from pipes?" You want a heading or a page, not a paraphrase of the whole policy.

Extract. Ask for specific facts: the deductible, the claim deadline, the exclusions. Say what shape you want back, such as a short list, or one number with the line it came from.

Challenge. Ask the model to argue with itself. "Quote the sentence that supports that, and say if anything else in the policy limits it." Insurance is full of "covered, except when...", and the exception is usually the part you needed.

The order matters. Orient, then locate, then extract, then challenge the answers that will cost you or save you money.

Follow-ups are where chat beats a search box. Your next message builds on the earlier ones, so you can write "and for a leak that went on for weeks?" without retelling the whole situation. If an answer is too broad, say how: "Shorter, one sentence, kitchen only." If it's thin, say what's missing.

Keep each message to one job. Two questions in one message tend to get one good answer and one vague one.

Then check. Answers cite pages like "(p. 4)" only when the text itself shows page numbers, so a document without printed numbers usually means no citations. When there is one, open that page in your own PDF viewer and read the sentence. Look for three things: the number matches, the condition is still attached ("only if", "unless"), and the sentence says what the model claims. One pitfall: the printed page number can differ from your viewer's count when there's a cover or front matter.

No page given? Ask the model to quote the sentence, then search the PDF for those words. If the quote isn't there, say so in the chat. A good reply is a correction, or an admission that the document is silent. The same habit works for contracts, which understand a contract with AI covers in more depth.

The file is sent again with every message. Each message costs Gems by model tier, and the cost is shown before you run it. On GPT-4o mini that's 1 Gem per message. A smart model such as Claude Haiku 4.5 is 5, and premium models are 25 on paid plans.

So a session costs roughly the number of messages times the tier. Ten loose questions on a smart model add up to 50 Gems. Four focused questions on GPT-4o mini add up to 4. New accounts start with 30 Gems, which covers a first session comfortably if you stay on target.

A few habits help. Write your questions down before you start. Combine two requests only when they come from the same section. Stop when you have the answer. And stay on the default model unless its answers are actually poor; you can switch in the dock. Pricing lists the packs and plans.

• A scan or a photo. No text layer, no OCR. Ask whoever made the document for the original file, or find a version where you can select text.
• A huge file. Past about 5.7 MB, or far beyond roughly 6,000 input tokens, it gets cut or refused. Pick a page range, or cut the file into chapters with Split PDF, which runs in your browser and uploads nothing.
• A sensitive file. Medical records, a client's documents, anything under an agreement not to share: don't upload it. The page tools don't send anything anywhere, and you can read the file yourself.
• A job with its own action. For the gist of a whole document, use Summarize PDF. For tables you want as a spreadsheet, use PDF to CSV. For two versions of one file, use Compare PDFs.

Chat is for questions. If your real goal is a shorter version, a table or a diff, the matching action will do it with less back and forth.