# How to ask a good question about documents

> Name the document, section and period; say what form the answer should take; attach a file; how to read an answer with a link, and what “not found” means.

The AI agent will ask if something is missing, but the more precise the question, the fewer the rounds and the fewer documents it has to read. The examples below come from a fictional company, “Notes”: a note-taking service for teams with the “Start”, “Business” and “Corporate” plans; the departments’ documents live in the “Company documents” folder.

## Good

- **Name the document, section or period** if you know it: “according to the contract and the plan annex”, “in the 2026 procurement policy”, “for the third quarter”. Then the agent goes to the right folder instead of going through all of them.
- **Say what form the answer should take**: a table, a list, a quote, a short “yes/no” with a link.
- **Name what to check against** if you suspect a discrepancy: “does it match the FAQ?”.

## Bad

- Too general (“what’s our situation with refunds?”) — it is unclear which document and which case are meant.
- Several questions in one — split them, and each gets its own link.
- Hidden context: “like with that client” — the agent does not know which one.

| Bad | Good |
| --- | --- |
| “What’s our situation with refunds?” | “A client asks for a refund 20 days after paying — what do the policy and the contract say?” |
| “Support times” | “What first-response time does the contract promise a client on the ‘Business’ plan, and does it match the FAQ?” |
| “Penalties” | “Which 2026 contracts have late-payment penalties, and at what rates? As a table: counterparty, rate, cap” |

## A file that is not in the sources

Attach it to the question — up to 10 MB: Word, Excel, PowerPoint, PDF, text. The agent reads the content, not the file name: text, tables, slides and their notes. This is how you check something that is not in the shared folder yet: “Check this text against the ‘How we talk about the product’ document” — with the text as an attachment.

You can attach a scan without a text layer, but the agent will not be able to read it: there is no image recognition. If a document exists only as a scan, ask for a text version.

## How to read the answer

- **A link to the document and section.** For a Notion or Confluence page or a document on Drive, it is a link to the page itself; for a file on a server, the path to the file. Follow the link if the answer matters: the agent quotes, but the decision is yours.
- **Two documents disagree — both are named.** “The policy and the contracts promise four working hours, the support FAQ promises eight” — this is not the agent’s mistake but a finding: the documents need to be put in order, and until then the knowledge base should record which one to trust.
- **“Reasoning”** shows what the agent opened and in what order. If it looked in the wrong place, name the folder or section.
- **An answer without a link** is a reason to ask “based on which document?”. An answer based on a knowledge base entry links to the entry, not to a document.

## What “not found” means

The agent did not find a document even though you know it exists. Four usual causes:

1. **The folder or section is not within the connection’s scope** — a different folder is selected as the root, or the one you need ended up among the exclusions. This is a connection setting, and IT changes it: [what to ask for](/en/docs/for-admin).
2. **The connection is not running** — its status and last error are visible in the app’s “Connectors” section; IT starts it.
3. **The account has no access** — the Drive folder was not shared with the connector’s account, the Notion integration was not added to the section, the network folder’s account has no permission to read it.
4. **The document is a scan without a text layer** — the agent sees the file but cannot read it.

If none of these applies, make the question more precise: name the file or folder.

## Follow up rather than starting over

Continuing in the same conversation keeps the context: the agent remembers what it has already read, and a follow-up costs less than a new question. A new conversation starts from scratch. The agent itself may ask a counter-question with options — answer it in the same conversation.

## “Input / output” usage

The conversation header shows two numbers and the cost. **Input** is everything the agent read: your question, the documents opened for the answer, knowledge base entries, earlier messages in the conversation. **Output** is everything it wrote: the answer and the reasoning. The more precise the question, the smaller the input. More in [What a question costs](/en/docs/cost-of-a-question).