Give recurring questions SQL your team has reviewed.

Choose a question your team asks regularly, like “What was net revenue in September?”, add the SQL and run it against your data. Check the result before you save it as an example for the agent to reuse.

• Save & Run Query runs the SQL before you save it
• Each example keeps the question, the SQL and its source

-- Net revenue for September 2026 (sample)
WITH paid AS (
  SELECT COALESCE(SUM(i.amount_cents), 0) AS cents
  FROM invoices i
  JOIN accounts a ON a.id = i.account_id
  WHERE i.status = 'paid'
    AND i.paid_at >= '2026-09-01'
    AND i.paid_at <  '2026-10-01'
    AND NOT a.is_test
),
refunded AS (
  SELECT COALESCE(SUM(r.amount_cents), 0) AS cents
  FROM refunds r
  JOIN accounts a ON a.id = r.account_id
  WHERE r.refunded_at >= '2026-09-01'
    AND r.refunded_at <  '2026-10-01'
    AND NOT a.is_test
)
SELECT (paid.cents - refunded.cents) / 100.0 AS net_revenue
FROM paid, refunded;

An example records more than a number: the joins, date boundaries and exclusions behind it. When the agent reads a table, the examples on it show the filters your team applies, like leaving out test accounts.

• Examples on a table show the agent its usual filters
• The agent reads your examples before it answers

When the SQL behind an answer matches an example, the answer is marked as a verified query and rated high confidence. Add the same question as an eval, and rerun it after changes to check the agent still answers it your way.

• Marked as a verified query, rated high confidence
• Add the question as an eval to keep checking it