Problem
Teams that serve dashboards and AI chat (or agents) from the same data often define KPIs in the BI tool, at the end of the pipeline, where the platform cannot see them. Engineers have to translate measures from DAX into SQL to check a number. Chat tools and agents have no definitions to rely on, and they answer confidently even when KPIs are unclear or several near-identical KPIs exist.
Example from the article: in classic BI each ad hoc question tends to become its own report. The report is rarely opened again, yet an organization can end up with thousands of them.
Solution in brief
Define measures and dimensions once as Metric Views, which live in Unity Catalog next to the tables. Dashboards, Genie, the SQL editor and notebooks then query the same definitions. Build small, specialized Metric Views, each centered on measures that belong together, and combine several of them per report or Genie space instead of growing one big model.
Questions this case answers
- How to define KPIs once in the data platform so that dashboards, natural-language questions and agents return the same numbers
- How to decide between many small specialized semantic models and one large shared model
- How to define KPIs once so that dashboards, chat with data (Genie) and AI agents return the same numbers
- How to decide between small specialized semantic models and one big shared model
Tools and methods
- Metric Views (Databricks) — a semantic layer in Unity Catalog; sources, dimensions, joins and measures defined in SQL and YAML
- Unity Catalog — metadata, column-level lineage and access control, as for tables
- Genie (chat with data) — for ad hoc questions, and a way to see what users ask
- Dashboards — for core KPIs and operational monitoring
Google or a one-time email code. No password.