Golden Assets

Golden assets teach the agent how your team actually answers questions. The system prompt encodes general rules and the data model encodes structure; golden assets encode patterns. Each one is a simple chain: a recurring question your team gets asked → the confirmed, trusted SQL that answers it correctly → and optionally the chart that presents it best. When a user asks something, the agent searches your golden assets for similar questions and uses the closest matches as references while it writes its own SQL and picks a chart. Each one is a demonstration of correct behavior the agent can generalize from.

​

Golden assets are test cases

You don’t keep golden assets in a separate library. A golden asset is a test case with Use as golden asset switched on, so one example does two jobs:

• It teaches the agent: the agent retrieves it as a worked example for similar questions.
• It checks the agent: every test run asks the question and grades the answer against the same SQL.

Turning the switch off keeps the test and stops the agent from retrieving it. A test case that isn’t a golden asset still checks the agent; it just isn’t shown to it. The amber bookmark beside a test case on the Tests page marks a golden asset.

​

Add a golden asset

1. Open your workspace and go to Develop → Tests.
2. Click Add test case.
3. Enter the Question the way a user would phrase it (for example, “Show sales by owner segmented by product”).
4. Enter the SQL query that answers it correctly. It must run before you can save.
5. Optionally, add a Chart when the right answer is a particular visualization.
6. Switch on Use as golden asset.
7. Click Add test case. It’s saved to the draft. Save a workspace version, then promote it to production (or, with environments, deploy it to your users’ environment) to make the golden asset available to them.

There are quicker ways too:

• From a good answer. In Chat History or the Playground, open a trace and click Save as test case. The question, the SQL and any chart are filled in for you; switch on Use as golden asset.
• With the Builder agent. Ask the Builder agent to “add a golden asset for monthly spend by category”, or let guided setup write a first set.

​

What makes a good golden asset

Write the question the way a real user would ask it, not as a technical specification: the agent matches on meaning, not keywords. Write the SQL your team would actually run, with every filter, join and calculation that makes it correct for your data. Prioritize:

• Questions with non-obvious joins or filters specific to your schema.
• KPI calculations with a precise definition (for example, “revenue” as a specific formula, not just a column name).
• Questions whose correct answer depends on business rules the agent can’t infer from column names.
• Questions the agent previously got wrong, found in Chat History or a failing test.

Add a chart when the right visualization isn’t obvious from the shape of the data: for example, when revenue trends should always be a line chart by month, or a comparison should be grouped bars rather than stacked. 15–25 well-chosen golden assets covering your core questions do far more for accuracy than 100 generic ones.

​

Check that it’s used

Open the trace of an answer in the Playground or Chat History. Its retrieval step shows which golden assets the agent found for the question. If a common question retrieves none, add a golden asset phrased the way your users ask it. See Agent Observability.

​

Next steps