Sundial

For Data Teams

Curate context, monitor how the agents are used, and verify quality with evals, the governance loop behind trusted answers.

Sundial is built so business users get instant answers and the data team stays in control of how those answers are produced. This section covers the surfaces and workflows you use to govern the platform: the Context Engine you curate, the Observability you monitor, the Evals you use to verify quality, and the Data Modelling Agent's Playground workflow for testing changes before merge.

The governance loop

These surfaces form a deliberate feedback loop rather than a set of disconnected tools:

  1. Business users ask questions and explore dashboards.
  2. Observability shows where answers are weak, usually missing or incomplete context.
  3. You improve the Context Engine on a branch: semantic definitions, AI context, or playbooks.
  4. The Data Modelling Agent's Playground workflow runs relevant evals against the branch and compares it with main.
  5. You review the result and merge only after confirming the branch does not introduce a regression.
  6. Evals continue to track quality after merge.

Over time this turns reactive firefighting into a proactive practice: you catch and fix gaps before business users hit bad answers.

What's here

SurfaceWhat it does
Context EngineCurate the semantic layer, playbooks, AI context, and warehouse metadata every agent uses
ObservabilitySee how the agents are used, drill into sessions, and act on trust signals
EvalsDefine prompt/response sets that catch quality regressions before they ship
PlaygroundUse the Data Modelling Agent to test branch changes against relevant evals before merge

To model the dimensions and measures the agents rely on, see the Semantic Reference.

On this page