Ask and analyze
How to get accurate results
Write clear questions, check results, and keep common definitions for reuse.
The agent works out what to query from your question. The more specific the question, the more accurate the result and the fewer back-and-forth confirmations. These methods apply to every way of asking, including conversations, dashboards, and Feishu.
What a good question contains
| Element | Vague | Clear |
|---|---|---|
| Time range | Recent sales | Sales in October 2026 / sales in the last 30 days |
| Metric and definition | Sales | Sales (amount paid minus refunds) |
| Grouping | How are stores doing | Grouped by store |
| Filters | East China data | East China only, completed orders |
| Sort and count | Which stores are good | Top 10 by sales, highest first |
| Output form | — | Show the daily trend as a line chart |
For example:
Sales (amount paid minus refunds) of completed orders in East China in October 2026, grouped by store, top 10 from highest to lowest.
Use business terms and state definitions
- Use names that exist in the data, or terms your team commonly uses. If the data calls it “GMV”, do not just say “turnover”.
- When a word can be calculated in several ways, say how, such as “average order value = amount paid divided by number of buyers”.
- For time, say which date to use, such as “by order date” or “by payment date”.
Break complex questions apart
Ask one thing at a time, then follow up in the same conversation:
- “What were sales for each store in October?”
- “Which 5 stores fell the most compared with September?”
- “For these 5 stores, which product categories fell the most?”
The agent keeps the context. To refer to earlier content, select it and click “Quote”. For analysis where you want to keep every intermediate result, use the Analysis canvas.
Check results
- Read the SQL first. Expand the result to view the executed SQL and confirm the time range, filters, and calculation.
- Look at row counts and totals. When the row count or total is far from what you expect, the filters or joins are usually wrong.
- Choose accurate options when clarifying. When the agent asks back, your question has several possible readings. Choosing “Skip” lets it decide, and the result may not be what you want.
Make future questions more accurate
Definitions you keep repeating in questions should be stored for the agent:
| Content | Where |
|---|---|
| Field meanings, code values, and units, such as “status=3 means completed” | Table and field notes |
| Business rules and terms, such as “sales = amount paid - refunds” | Business documents |
| Fixed analysis steps and output format, such as “a weekly report first shows totals, then the top 10 stores” | Analysis skills |
Choose a run mode
The default deep thinking mode explores and verifies more, which suits complex questions. When you only need one clear number, turn on “Fast mode” in the “+” menu for faster, more direct answers. If the result looks wrong, turn fast mode off and ask again.