Understand plans, answers, and provenance
A Cassis answer is inspectable. Check the plan for assumptions, the SQL for execution, and the context objects for provenance.
Check the plan
The plan is Cassis’s interpretation of the question. Read its assumptions first: “revenue last quarter” may mean booked revenue, calendar quarter, excluding refunds.
When Cassis surfaces a plan for approval, nothing runs until you approve it. Plans without a genuine fork run automatically. If an assumption is wrong, state the correction; you do not need to rewrite the whole question.
Check the answer
- Answer
- The conclusion and numbers, in the language you write in. A very short message, such as “ok”, keeps the language of your earlier messages.
- SQL
- The exact generated statement. On a connected warehouse it produced the rows; on a schema-only project it was not executed and
resultsisnull. - Objects used
- The tables, joins, and metrics that grounded the answer. An unexpected object is a reason to inspect the result more closely.
- Domains consulted
- The domain context Cassis read, in the order it explored it. Click a domain to inspect its definitions. This list also appears when a question cannot be answered.
- Context version
- The published context the conversation is using. Follow-ups stay in the same conversation and keep that context.
In a long conversation, Cassis can lose sight of earlier exchanges and says when that limits its answer. Restate a definition or constraint if it is needed for the next question.
Correct a wrong answer
Rate it down and explain the missing or incorrect rule in the conversation. The correction becomes evidence; recurring evidence is grouped into a detected issue for the data team.
In the web app, Add as test case captures a candidate regression case. A reviewer must validate both the SQL and the business interpretation before it becomes approved truth.
“Not answerable” is useful. It means the required definition, relationship, or data is missing. For an undefined business term, reply with its definition or add it to the context for future questions. If the data cannot express what you asked, Cassis explains the limitation; a definition alone cannot supply missing data.