Docs › Asking good questions
Asking good questions
Say what you mean by the words
ProLyt writes SQL from your question, so the question has to carry the definition. "Net revenue" can mean revenue minus cost or revenue minus returns; "last quarter" needs a date range if the data spans years. When a term has a house definition, a workspace owner can record it in the metric glossary (Workspace page); the glossary is shown to the model as guidance — it steers the query, it does not compute the number.
Good: *"Total order value per region for 2025, regular orders only (status = completed)."* Vague: *"How are regions doing?"*
Scope the question
- The data panel lets you scope a question to one folder or to a few tables; everything else is left out of the search.
/tablesin the question box inserts a table name exactly as stored;/tilesinserts the title of a saved dashboard tile, which then runs that tile's exact query — a way to reuse a definition someone already checked.
Follow-ups
On paid plans a chat keeps memory: ProLyt sends recent questions in the same chat, together with the SQL that answered them, to the model. The Explore plan is stateless — each question stands alone. Answers that could not be verified are never fed back as context.
Reading the answer
Under each answer:
- SQL — the query that ran. Copy it, or Download full result to re-run it and export the complete rows.
- Tables read — what the executed query actually touched, as opposed to what the search offered.
- Passages — the document excerpts the answer cites.
- Chips flag things worth a look: a join that may double-count rows, figures in a document-based answer that were not found in the passages read, a list built from only part of your documents, or a credit refund.
If an answer is wrong, the fastest fix is usually a more precise question. Edit re-asks with a changed wording (the original question and answer stay in the chat); Retry asks the same question again.
Predictions and forecasts
Questions that ask for a prediction ("what will sales be next quarter?") are declined in the chat at no charge: the model is not allowed to invent a forward number. You get three options: open the Forecast page (a deterministic projection of a time series with an uncertainty band, no AI model involved), record your own prediction in the Predictions ledger, or rephrase as a question about the past.
Charts
Ask for a chart by name when you want one ("as a bar chart", "line chart by month"). Charts are drawn from the rows the query returned; a chart that would need columns the query did not return is refused and the model is told what it can use. Charts can be pinned to a dashboard and included in a deck.