Docs · Step 4
Ask a question
Plain English in, a checked answer out, with the query and rows behind it when there are any.
1. Open Chat
- Select the dataset (or datasets) you want to ask about.
- Open the Chat page.
2. Ask a question
Type a question the way you'd ask a colleague: "what's the average final grade for students who attended office hours", "how many rows are missing a value in the income column", "what predicts whether a patient was readmitted". No query language, no column-name syntax required.
3. Read the answer
CLAIR answers directly, and marks an answer verified when it came from a query CLAIR actually wrote and ran against your table (or a saved model, or a rule-based computation), rather than from the language model alone. A verified answer shows the query and the rows it ran against; click through to open both.
An unmarked answer isn't necessarily wrong
It just means CLAIR couldn't check it deterministically, most often a genuinely open-ended or exploratory question. Treat it the way you would a knowledgeable colleague's answer: useful, but worth checking against the data yourself for anything that matters.
Some questions are declined outright rather than answered with a guess: if a question names a value your data doesn't actually hold, or asks about a group CLAIR can't map onto a real column, it says so and names the mismatch instead of returning a confident wrong number.
4. Ask a follow-up
Chat keeps the last question in mind for a natural follow-up ("now break that down by year") without repeating the full context yourself.
What "verified" means, precisely
- A bare aggregate (a sum, average, count) over a table that also contains a total or rollup row is checked before it's allowed to verify, so it can't silently double-count that row.
- A question about "how many people" against a dataset with more than one row per person (repeated measurements, say) is checked, and the badge is withheld if the raw row count would answer a different question than the one you asked.
- A group comparison excludes a rollup label (like an "All Regions" summary row) from the statistical test it runs, rather than treating it as a real group.