Semantic layer and skills
A semantic layer is a translation layer that maps business terms, such as “revenue”, “active customer”, or “gross margin”, to the tables, columns, and logic that compute them, so every consumer of the data uses the same definitions.
Without one, each query author decides what a metric means, and two dashboards that should agree end up disagreeing. With one, the definition is written down once and reused.
How it works in dvt
dvt builds on whatever modeling you already have: raw tables, dbt models, or a governed semantic layer in your warehouse all work as the data an agent queries. Alongside that, dvt has its own layer for the knowledge an agent needs at authoring time, called skills.
A skill is an org-scoped, versioned document that holds metric definitions, warehouse semantics, and house rules for how your team builds charts. For example, a skill can say that “churn” means logo churn, that orders with status returned are excluded from revenue, and that currency is always shown in compact USD. An agent reads the relevant skills over MCP before it writes any SQL, so the dashboards it builds follow your definitions instead of guesses from column names. Skills are written once, in the app or by an agent, and every agent with access reads the same version.