Document QA

Document Q&A

When to use

The user has attached a document (or referenced one in their files) and is asking a specific question about its content.

Steps

1. Identify the document — confirm what file the user is asking about. If multiple files are attached, ask via ask_questions which one to focus on.

2. Read the document:

   • For attachments in the current message: the file is already in context (use directly)
   • For files in the user's R2 storage: call fs_read with the path
   • For URLs: call browser_markdown or use convertToMarkdown server-side

3. Answer the question — based on what's actually in the document:

   • Quote specific passages when answering yes/no questions
   • For "what does it say about X?" — find the relevant section, quote or paraphrase
   • For "summarise" — use the summarise-url skill pattern
   • For "find me X" — search the text and report what's there

4. Cite specifics — when you can, reference the page, section, or paragraph: "On page 3, the report states..."

5. Be honest about limits:

   • If the document doesn't contain the answer, say so explicitly
   • If the document is unclear, acknowledge it
   • If it's image-based and OCR is poor, note that

Style

• Direct quotes in quotation marks; paraphrases without
• Page/section references when available
• Don't speculate beyond what's in the document
• If asked your opinion on the document's claims, separate the question of "what it says" (factual) from "is it correct?" (analytical, may need fact-check)

What not to do

• Don't make up content that wasn't in the document
• Don't blend prior knowledge with document content without flagging it ("the document says X, though more recent data suggests Y")
• Don't claim to have read more than was actually extracted (if the document is huge and only part loaded, say so)