Data context and governance

    Pointing Genie at raw tables is a fun way to get confident wrong answers

    Databricks Genie lets people ask questions of their data in plain language. Its answers hold up when the metrics and relationships it uses are defined in a semantic layer, such as Unity Catalog metric views, and when each person sees only the rows they may see. Raw tables plus a hopeful prompt is how the numbers drift.

    By Toby Mather · Published

    Key takeaways

    • Genie One is where business users ask questions. Data teams build Genie Agents (formerly Genie spaces) with the datasets, sample queries, instructions and metrics behind the answers.
    • Genie only queries the data you attach, up to 50 tables, views or metric views, all registered in Unity Catalog.
    • Attach certified metric views for the numbers people argue about. Genie can then use the approved calculation in place of rebuilding it from raw columns. Test that it actually picks the metric view.
    • Data access is evaluated with each end user's own Unity Catalog permissions, so row filters and column masks belong in Unity Catalog, where they hold for every question.
    • Agents outside Genie, such as shared or scheduled ones, have no person behind them and need their own identity and log.
    1. 01Business question

      Asked in Genie One

    2. 02Genie

      Reads attached data

    3. 03Metric views and access controls

      In Unity Catalog

    4. 04Answer with certified metrics

      Tested against the report

    A business question goes to Genie, which reads certified metric views and checks the asker's access before it answers.

    What is Databricks Genie?

    Genie is the natural-language question answering built into Databricks. Business users ask questions in Genie One. Behind it, data teams set up Genie Agents, which Databricks describes as domain-specific environments where the team configures trusted data, metrics and business rules. Databricks renamed Genie spaces to Genie Agents, so older guides use the old name.

    An author builds a Genie Agent by attaching datasets, adding sample queries and instructions, and adding metrics, business rules and verified answers to improve response quality. Genie writes SQL against what is attached and returns the answer with the query it ran.

    Why does Genie need a semantic layer to answer correctly?

    Raw tables are ambiguous in ways a model cannot see. An orders table might hold test transactions, refunds as negative rows and three currency columns. Ask Genie for last quarter's revenue over that table and it will pick an interpretation and answer with full confidence. The answer reads well and may still differ from the finance report by a margin nobody notices until the board meeting.

    A metric view removes the guesswork for the numbers that matter. Revenue is defined once, with the filters, joins and currency rules your finance team approved, and Genie asks for it by name. Instructions in the Genie Agent also help, though they are prompt text that only Genie reads. A definition in Unity Catalog is the same for Genie, dashboards, notebooks and alerts.

    How do you set up Genie on metric views?

    For a first Genie Agent, with your data already in Unity Catalog, I would follow this order.

    1. List the questions the team asks every week, and the metrics behind them. Start with the handful people argue about, such as revenue, active customers and churn.
    2. Define each metric in a metric view, with SQL DDL or in Catalog Explorer. Agree the definition with the person who owns the number, and check it against a report they already trust.
    3. Grant SELECT on the metric views, and on any tables Genie will use, to the groups who will ask questions. Put row filters and column masks on the source tables in Unity Catalog.
    4. Create a Genie Agent and attach the metric views and the few tables people need for detail. Genie does not query tables that are not attached, and the limit is 50 tables, views or metric views.
    5. Add instructions and example SQL queries, which Databricks recommends for accuracy. Use them for business vocabulary and preferred filters, and keep calculations in the metric views.
    6. Test with real questions from the list, compare the answers with the trusted report, and accept good suggested queries so they become part of the agent's context.

    Who can ask which questions?

    Databricks separates compute from data access. The author's compute credentials are embedded in the Genie Agent so everyone can use the warehouse, while data access is always evaluated with each end user's own Unity Catalog permissions. End users need SELECT on the data and CAN VIEW or CAN RUN on the agent, and they do not need direct warehouse permissions.

    That means a sales rep and a finance analyst can share one Genie Agent and see different rows, as long as the row filters live in Unity Catalog. A row rule written into Genie instructions is a request to the model, which a carefully worded question can get around, while a Unity Catalog row filter applies to every query.

    What still goes wrong when the metric is not certified?

    Two teams define active customer differently, one as a metric view and one in a dashboard, and Genie answers with one while the weekly report shows the other. A table gets attached for detail, and Genie builds revenue from it instead of using the metric view, because the question mentioned a column only that table has. An instruction that explained a business rule goes stale after a schema change, and nobody notices because Genie still answers.

    The fix is the same each time: certify the definition, give it one owner and make every tool read that one copy. Agents outside Genie need the same treatment, plus an identity of their own. A scheduled agent, or one a whole team shares in Slack or Notion, has no end user for Unity Catalog to evaluate.

    Can the same metrics serve Genie and your other tools?

    Yes, and they should. Databricks documents connecting metric views to external BI tools such as Power BI, Tableau and Sigma, and metric views are queryable from SQL editors, notebooks, dashboards and alerts. You can also export a Genie Agent to a metric view, which creates one from the data and semantic context configured in the agent. That is a quick way to turn a working Genie setup into a definition other tools can share.

    How Rig does it

    Rig connects to a Databricks SQL warehouse with a read-only identity, reads Unity Catalog and builds a context layer over your tables, joins and metric definitions. Rig Map keeps certified metrics with their SQL and descriptions in an open, OSI-compatible format, alongside the tables, joins and business terms around them, and the same definitions are available in Claude, Cursor and ChatGPT over MCP.

    Gateway covers the callers Genie does not. People connect with their own sign-in and see what their role allows. Agents get their own role, token and daily call cap, and every call lands in the audit log under the agent's name. Teams use Genie for questions inside Databricks and Rig for agents and AI tools that work across systems, on the same certified definitions.

    Tools and approaches

    Databricks Genie
    Natural-language questions in Genie One, answered by Genie Agents that data teams configure with datasets, sample queries, instructions, metrics and verified answers.
    Databricks metric views
    Metric definitions in Unity Catalog that Genie Agents can attach as data sources, with measures separate from dimensions.
    Unity Catalog
    Privileges, attribute-based policies, row and column filters, lineage and auditing for the data Genie reads.

    Common questions

    Does Genie replace a semantic layer?

    No. Genie answers from the tables and metric views you attach, so the quality of its answers depends on the metric definitions you give it.

    Can Genie query tables that have no metric view?

    Yes, if you attach them. Genie writes SQL against attached tables and views, and it does not query tables that are not attached. Attach metric views for the numbers that must match your reports.

    How do I stop Genie seeing the wrong rows?

    Put row filters and column masks on the tables in Unity Catalog. Genie evaluates data access with each end user's own Unity Catalog permissions, so those rules apply to every question.

    Can I use the same metrics in Genie and in Power BI?

    Yes. Databricks documents connecting metric views to external BI tools, including Power BI, so both can read the same definition.

    Should I point Genie at raw tables?

    Only for detail lookups. For any number someone will quote in a meeting, attach a certified metric view and check the answer against a trusted report.

    Sources

    Vendor documentation checked on 20 November 2026. Confirm current features and terms with the vendor.

    About the author

    Toby Mather, Co-founder and CEO, Rig. Toby co-founded Rig after seeing how hard it was for business teams and AI to work with complex internal data. He was previously co-founder and CEO of Lingumi and CPO at Novakid.

    Talk to Rig

    See your stack drawn as one diagram

    An architect drafts a one-page diagram of your warehouse, sources and access path, then offers a 30-minute walkthrough.