You have a long contract, a research paper, or a dense report, and you have one specific question. Reading the whole thing to find the answer is slow, and a plain text search only works if you already know the exact word to look for. Chatting with the document is the middle path: you ask a question the way you would ask a colleague, and the answer comes back drawn from the actual text. This piece explains how that works, why the bring-your-own-key model keeps it private, and what it honestly can and cannot do.

How chatting with a PDF works

When you open a document in the Chat tool and ask a question, the relevant parts of the text are sent to a large language model along with your question. The model reads those passages and composes an answer grounded in them, rather than in whatever it happened to learn during training. That grounding is the whole point: you are asking about your document, so the answer should come from your document. Good answers point to where they came from, so you can jump to the passage and confirm it rather than taking the response on faith.

  • Ask specific questions: what is the termination notice period, or which section covers liability.
  • Ask for locations: where does this mention the deadline, so you can verify against the page.
  • Ask it to compare or list: summarize every obligation on the tenant, or list the dates mentioned.
  • Follow up. Chat keeps the thread, so you can drill into an answer without re-explaining the document.

Why bring-your-own-key means private

The AI tools here use a bring-your-own-key model, and the privacy consequence is the important part. You supply your own API key from Claude, ChatGPT, or Gemini. When you chat, your document and question go directly to that provider using your key. They do not pass through a server of ours, and we do not store your document or your key on our end. In practical terms, the only party who sees the document is the AI provider you already chose and have an account with. If you trust them with your key, that is the trust boundary; there is no extra middleman collecting your files.

This is a real distinction from services that upload your document to their own servers to process it. Here, the tool runs in your browser, and the only outbound trip your document takes is the one you initiate, to the provider you named. You can revoke or rotate the key with that provider at any time, and nothing lingers on our side once you close the page.

Beyond chat: summarize, translate, extract

Chat is one of a family of tools that share the same key model, and it helps to know when to reach for a neighbor instead. When you do not have a specific question but want the shape of a document, the Summarize tool condenses it to its key points. When a document is in another language, the Translate tool renders it into one you read. And when you need structured facts rather than prose, the Data Extractor pulls fields into CSV or JSON, which is ideal for turning a stack of invoices or forms into a spreadsheet you can sort and total. Same privacy model throughout: your key, your provider, no server of ours in the loop, and nothing stored once you are done.

Honest notes on cost and limits

Because you use your own key, you pay your provider directly for what you use. Chatting over a document consumes tokens for the text sent and the answers returned; for a typical document this is usually inexpensive, but a very long document asked many questions will cost more, and that bill is yours. It is worth knowing your provider's pricing so there are no surprises. Reach for Chat when you have a large document and targeted questions, and lean on the notes below to keep the answers trustworthy.

  • Verify anything that matters. Models can misread or overstate, so check the cited passage before you rely on an answer, especially for legal or financial detail.
  • Very large documents may need to be split, since providers cap how much text they accept at once.
  • Answers are only as good as the text. A poorly scanned PDF with no real text layer gives the model little to work with; running OCR first helps.
  • The model does not know anything outside the document unless you tell it, which is exactly what you want when the goal is answers grounded in your file.