Ask Better Data Questions
Write the report you need in terms you can check. “Sales by region” leaves the period, currency and sale definition open. “Paid order value by region for August 2026, in EUR before refunds” gives the tool a more specific task.
Select a connection and relevant tables
Each conversation uses one connection. Select it before asking; switching connections starts a new conversation. Discovery lists the available tables and columns.
Table and column selection guide the context used for planning. Source permissions and connection controls determine access.
Review unclear table descriptions in the discovery view. You can edit descriptions and primary/foreign-key hints used in planning. These edits do not change the source database. Optional notes on an analysis can explain business definitions. Do not include secrets in these fields because they are model inputs.
@-mentions: pin tables and columns
Type @ to insert a table or column directly into your question. Useful when:
- A schema has many similarly-named tables and you want to be exact.
- You want to scope an analysis to a specific column.
The mention shows up as a pill in your prompt and is passed along with your question.
Filter pills
Below the input you’ll see chips for the tables in scope. You can:
- Toggle individual tables on or off.
- Toggle individual columns within a table.
- Click “Clear” to reset.
A narrower scope means less schema metadata to search, and fewer chances to pick the wrong table.
Question patterns for reporting and analysis
| Task | Example question |
|---|---|
| Aggregate | “Completed order value by month for Q2 2026, before refunds, in EUR” |
| Ranking | “Top ten customers by paid order value for August 2026” |
| Comparison | “Compare conversion by region using paid orders divided by eligible visits” |
| Trend | “New accounts per complete week for the last three months” |
| Investigation | “List orders overdue for dispatch as of September 8 at midnight” |
State the reporting grain and definition. Analysts can include known tables, join keys or exclusions when exploring a schema; business users can state the desired outcome and confirm definitions with the source owner.
Read the result and execution details
The answer can contain text, a table or chart, source names, returned-row count and runtime. Open the query and trace details to inspect the executed SQL/KQL and the recorded steps. If a query was revised after an error, review the final executed version.
A correct-looking SQL statement is not proof that the result is right. Check its business meaning, source freshness and relevant records. Keep the query with the result when asking another person to review it.
Include the decision-making details
| Detail | Example |
|---|---|
| Measure | Paid order value before refunds |
| Period and time zone | August 2026, UTC, excluding September 1 |
| Grouping | One row per region |
| Exclusions | Cancelled orders and test accounts |
| Output | Table first, then a bar chart |
If a term such as “active customer” has several meanings, state the rule: a paid order in the previous 90 days, for example. Clarifying a term does not establish that the source contains the fields needed to calculate it.
Inspect the answer
For a completed data query, review the table or chart and open the query details. The source tables show where the data came from; row count and runtime describe the execution. A grouped result can contain two rows even when thousands of orders were read.
Check dates, labels and a known total yourself. Ask a SQL reviewer to inspect join keys, deduplication or a complex calculation. A plausible query can still contain a mistake; an unexpected answer may also reflect stale data or a different business definition.
Refine and chart
Ask one follow-up at a time: “Only wholesale customers,” then inspect the changed result. Selected values from single-row answers can enter follow-up context. Data handling
Use the chart control or ask for a chart. Confirm its values and units match the table before adding it to a dashboard. Interactive charts use a declarative specification; Python runs only when the plan needs an analysis step.
Resolve common failures
| Result | Next check |
|---|---|
| No rows | Confirm period, filters, permissions and source freshness; broaden only if it matches the task |
| Timeout | Narrow the period or ask the source owner about the query workload |
| Invalid plan | Simplify or clarify the request and retry |
| Connection failure | Test the connection and review network/authentication settings |
| Permission failure | Check the identity’s grants with the owner |
| Validation rejection | Review whether the query requests a write or goes outside the allowed scope |
Each error message, and what clears it, is explained in Troubleshooting. Conversations are private to your account; you can bookmark their /c/<id> address to reopen them.