Ask AI about a PDF or Word document

Attaching documents

Use the composer's document action, or drag files onto the assistant window. Each becomes its own attachment chip, which you can remove before sending.

Text is extracted on your machine before anything is sent. PDF extraction preserves page markers. DOCX extraction reads text, lists and tables without running macros or loading external resources. If the combined context is too large, the request is rejected with a message rather than silently truncated, so you are never in the position of getting a confident answer about two thirds of a document.

Prompts that work

The difference between a useless and a useful document answer is almost always specificity about the output you want. Compare each pair.

Summarizing

• Weak: *Summarize this report.*
• Better: *Summarize this report in under 200 words for someone who has to decide whether to fund the next phase. Lead with the recommendation.*

Extracting

• Weak: *What are the action items?*
• Better: *List every commitment in this document as a table: what was committed, who owns it, and the stated date. Mark any where the owner or date is missing.*

Reviewing

• Weak: *Is this specification any good?*
• Better: *Read this specification as the engineer who has to build it. List every place where the requirement is ambiguous enough that two developers could implement it differently.*

Comparing

• Weak: *Compare these two proposals.*
• Better: *Compare these two proposals on scope, timeline, cost and assumptions. Give me a table, then note anything present in one and absent from the other.*

The pattern: state the audience, the format and the criteria. *Mark anything missing* is worth adding to most extraction prompts, since gaps are usually what you needed to know.

Working with several documents

Five documents per request is generous enough for real comparison work: two contract versions, a specification and its test plan, three vendor responses against a requirements document.

When you do this, name the documents in your prompt. *Compare the 2025 agreement with the 2026 draft* is easier for a model to follow than *compare these*, especially once the conversation has scrolled and you are on a follow-up question.

If you find yourself attaching the same set of files repeatedly, that is the signal to switch to local folder context, which indexes once and retrieves per question.

The privacy boundary

The original file is never uploaded. Extraction is local, and no network connection is opened during it. What goes to your provider, once you press Send, is the extracted text.

So the question to ask about a confidential document is not *does this app upload my files* but *am I comfortable with this provider receiving this text*. If the answer is no, use a local provider, where the request goes to your own machine.

Extracted text is stored in the conversation archive so saved threads survive the original file moving. Enabling history protection applies DPAPI to document contents too. Markdown export lists attachment names without dumping their extracted bodies.

After the answer

• Save as PDF turns one response into a document you can file or attach elsewhere. See exporting a response.
• Edit and resend the question rather than arguing across three turns when the answer misses. It is faster and keeps the thread short.
• Watch the context budget. A large document plus a long conversation pushes against it. Managing conversations explains compaction.
• Verify anything consequential. A model's reading of a contract or a regulation is a first pass. It is not advice, and it should not be the last word on anything with legal or financial weight.