The quality of the AI's answer depends heavily on how you ask. This page gathers practical patterns for text-to-SQL (natural language → SQL) and for the dashboard Q&A (questions about the dashboard), with examples of good and bad questions.
Example workspace: Comércio Aurora Ltda (aurora).
The AI proposes, you confirm
Treat the AI as an assistant that proposes — you review and confirm.
The 5 principles#
- Be specific about the metric and the slice. Say which number, by which dimension, and for which period.
- Use the names from your model. "Revenue," "category," "branch" — the terms from your measures/columns help the grounding (see How the AI understands your data).
- One question at a time. Compound questions ("revenue and also margin and also why it dropped") are confusing; break them into steps.
- Ask for the period explicitly. "Last quarter," "2026," "last 30 days."
- Don't ask for a cause. The AI describes what the data shows; it does not claim that "X caused Y" (and it refuses when someone tries to force it).
Text-to-SQL: good × bad#
| Bad | Why | Good |
|---|---|---|
| "show me everything" | no metric or slice | "total revenue by month in 2026" |
| "sales" | ambiguous | "quantity sold by category last quarter" |
| "which are the best?" | "best" at what? | "top 10 products by revenue in July 2026" |
| "compare with the market" | data you don't have | "this year's revenue vs last year by branch" |
Note
If the question is ambiguous and there's a published semantic model, the AI refuses before spending tokens and lists the options ("did you mean Gross revenue or Net revenue?"). Pick one and ask again — it's cheaper and more accurate.
Dashboard Q&A: good × bad#
The Q&A answers only based on the already materialized snapshot, and every statement comes with a verifiable citation. So:
| Bad | Why | Good |
|---|---|---|
| "why did sales drop?" | asks for a cause | "which category dropped the most vs the previous month on this dashboard?" |
| "what's the forecast for December?" | not in the snapshot | "which customer had the highest revenue in the period shown?" |
| "is this good?" | opinion with no basis | "which items are below the target shown?" |
| "what was the total?" (on a visual that is a sample) | a sample doesn't allow a total | aggregate in text-to-SQL for the exact total |
Read the answer correctly#
- Text-to-SQL returns read-only SQL + an explanation + an estimated cost. Review the SQL (a column/table name may be wrong) before trusting it. There's self-correction (one retry with the real error from the datalake), but it's not a guarantee.
- Q&A includes citations that point to the data backing each sentence. If the
answer comes back as a refusal (
sem_evidencia,fora_do_contexto), it's because there was no backing — don't force it. - On a sample, the AI is forbidden from stating a total/sum/"largest"/"all."
When the AI refuses (and why that's good)#
A refusal isn't a failure — it's honesty. The common refusals:
- Ambiguity → pick one of the options and ask again.
- Out of context → the question doesn't touch any term on the dashboard; rephrase using the names from the model.
- No evidence → the dashboard doesn't support the statement; aggregate in SQL or change the question.
Common mistakes#
- Asking the Q&A for real time: it reads the snapshot from the last refresh, not "now."
- Treating the explanation as proven fact: insights and recommendations are hypotheses to investigate — see Insights.
- Pasting secrets/PII into the question: don't. The text goes to the provider (fenced, but it still leaves) — ask about structure and numbers, not about personal data.
Related#
Last reviewed: 2026-09-08.