Databricks Genie can now reuse Power BI DAX measures in Beta

2 minutes ago

8 min read

Cover with the title Databricks Genie can now reuse Power BI DAX measures in Beta on a teal data skyline, Data Studios original cover

Databricks Genie can now read the measure definitions stored in Power BI semantic models and use them when it answers business questions. Databricks listed the capability as "External knowledge sources" in its platform release notes on November 5, 2026, with Beta status. Power BI is the only external source documented so far, and the metadata arrives through a Unity Catalog connection that a Power BI task in a Databricks job already uses or will use.

The practical effect is that a company with years of curated DAX in Power BI does not have to rewrite its revenue, margin or churn definitions before Genie can answer questions about them. According to the documentation, Databricks reads each semantic model's definition, including DAX measure formulas, calculated columns and Power Query (M) expressions, and turns the measures, calculated columns and relationships into Ontology Snippets that Genie can draw on. The metadata refreshes every 7 days and stays in the workspace's region.

There is a governance catch that BI administrators should read before switching it on. Databricks states that Ontology Snippets created from Power BI metadata are available to all users in the Databricks account, regardless of their access in Power BI, and that they cannot be restricted to specific users or groups. Table data stays protected by Unity Catalog permissions, but the definitions themselves become account-wide.

··········

WHAT DATABRICKS ANNOUNCED FOR GENIE AND POWER BI.

The release note entry says that metadata can be pulled from Power BI semantic models into Databricks and that Genie can use Power BI measure definitions when it answers questions. The feature belongs to Genie Ontology, which Databricks describes as the unified context layer for its Genie experiences. Genie One and Genie Code search the same ontology, so context curated once applies to both. Genie Ontology itself is in Public Preview, while the external knowledge sources part is in Beta.

Databricks splits Genie's context into two kinds. Modeled context is what a team defines, governs and certifies with Unity Catalog semantics, such as metric views, domains and Pages. Inferred context is made of snippets that Genie extracts and maintains automatically from existing assets and usage, including metric views, dashboards, SQL queries and Genie Agents. Power BI metadata joins the inferred side. It does not become a certified metric view, and it is not visible as a Unity Catalog table.

Each snippet gets an authority score that Databricks bases on where the snippet was generated from, how often it is used and how fresh it is. Genie ranks the most relevant snippets, resolves conflicts and answers using only sources the user can access. In a Genie response, users can click the Snippet icon to see which knowledge sources were used, which is the first place to look when a number needs to be traced back to a Power BI measure.

The same release notes also let context-based ingress policies allow traffic from Power BI, Tableau Cloud, dbt Cloud and the Genie apps for Slack and Microsoft Teams, which helps teams that publish to Power BI from locked-down workspaces. On the modeled side, Databricks added relationships between Unity Catalog metric views on October 30.

··········

WHAT POWER BI METADATA GENIE IMPORTS.

The Power BI page of the documentation lists exactly what Databricks pulls and what it leaves out. Databricks reads each semantic model's definition with the Fabric Get Semantic Model Definition REST API, through the Unity Catalog connection. While the preview is on, Databricks pulls from every OAuth Machine to Machine connection used by a Power BI task, existing ones included, and for each connection it reads the workspaces that the service principal can access.

........

Category	Imported by Databricks
Semantic models	Tables, columns, data types, storage mode
Measure and query logic	DAX measure formulas, calculated columns, Power Query (M) expressions
Model configuration	Semantic model diagram layout
Excluded	Row-level security role definitions and members
Excluded	Power BI asset owner email addresses
Not captured	Usage activity such as report views

........

Two limits narrow the scope considerably. Databricks creates Ontology Snippets only for Power BI semantic models that read data from Unity Catalog objects, so a model built on SQL Server, Snowflake or flat files will not produce snippets even if the service principal can reach it. The feature is also unavailable in workspaces with the compliance security profile enabled, which rules out some regulated deployments for now.

The documentation does not say how Genie translates DAX logic into the SQL it runs on Databricks, how it treats measures that depend on DAX-only constructs such as time intelligence functions, or whether there are limits on model size or the number of models. Those gaps are reasons to test the feature on a small set of well-understood measures before letting business users depend on it.

··········

SETUP: PREVIEWS, CONNECTION AND THE POWER BI TASK.

Getting Power BI metadata into Genie takes switches at three levels. An account admin turns on the External knowledge preview from the account console Previews page. A workspace admin turns on the Genie Ontology snippets preview in each workspace where Genie should use Ontology Snippets. The workspace must have access to serverless compute.

On the connection side, the Unity Catalog connection must be of type Power BI with Auth type set to OAuth Machine to Machine, storing the client ID, client secret and tenant ID of a Microsoft Entra service principal. Creating that connection requires the CREATE CONNECTION privilege, and anyone using it in a task needs USE CONNECTION. The service principal must be added directly to each Power BI workspace that Databricks should read, with at least the Contributor role, and the separate connection guide also asks administrators to enable service principals in the Power BI admin portal.

The metadata pull is tied to the Power BI task for Lakeflow Jobs, which is itself in Public Preview and exists to publish and refresh Power BI semantic models from Databricks. The first time the task runs after the preview is on, Databricks starts pulling metadata. If a team already runs a Power BI task with an OAuth Machine to Machine connection, no new connection is needed, and the pull begins at that task's next run.

........

Requirement	Who handles it	Documented detail
External knowledge preview	Account admin	Account console Previews page
Genie Ontology snippets preview	Workspace admin	Each workspace that should use snippets
Serverless compute	Workspace	Must be accessible
Power BI connection	User with CREATE CONNECTION	OAuth Machine to Machine, Entra service principal
Workspace access in Power BI	Power BI admin	Service principal at least Contributor
Power BI task in a job	Job owner	Runs with a SQL warehouse; Run as identity needs privileges
Metadata refresh	Databricks	Every 7 days

........

The publishing requirements behind the Power BI task still apply. Databricks documents that publishing to the Power BI service needs data in Unity Catalog, a Databricks SQL warehouse, a Premium-tier Power BI license (Premium capacity, Premium Per User or Microsoft Fabric capacity) and the XMLA endpoint set to Read Write on the capacity. For teams still working out which of those licenses fits their headcount, the comparison of Power BI Pro, Premium Per User and Fabric F64 capacity sets out the official prices and the break-even points.

··········

GOVERNANCE: SNIPPETS ARE VISIBLE ACCOUNT-WIDE.

The access model for Power BI snippets differs from the rest of Genie Ontology. For ordinary snippets, Databricks says they are gated by Unity Catalog permissions and that Genie only uses snippets extracted from assets the user has permission to see. For snippets built from external knowledge sources, the documentation states that they are available to all users in the Databricks account, regardless of their access in Power BI, and that the only way to stop serving them is for an account admin to turn off the External knowledge preview. Turning it off stops serving Ontology Snippets in all workspaces in the account.

Databricks also states that Ontology Snippets contain metadata, such as measure definitions and relationships, and not the data in tables. Users without access to the referenced tables cannot query them through Genie. In practice this means a sales analyst in one business unit could see how another unit defines a commission or a discount-adjusted margin, but could not read that unit's rows. Where measure names or formulas are themselves sensitive, the service principal's workspace list is the control point: Databricks pulls only from the workspaces the service principal can access, so adding it only to workspaces whose definitions can be shared account-wide keeps the rest out.

On cost, the Genie Ontology page says that customers do not currently incur costs for curating ontology snippets, while noting that pricing may change with notice. The pages reviewed for this feature do not list a separate charge for the Power BI task, which runs on a SQL warehouse.

··········

CHECKING A GENIE ANSWER AGAINST A POWER BI MEASURE.

One error to test for when a definition crosses from DAX to SQL is a change in aggregation order. A Power BI margin measure usually divides summed profit by summed revenue, while a naive SQL translation averages the margin of each row. The illustrative dataset below shows how far apart the two results can be.

........

Region	Month	Revenue (USD)	Cost (USD)
East	2026-09	120,000	78,000
East	2026-10	135,000	90,450
West	2026-09	80,000	60,800
West	2026-10	64,000	46,080
Central	2026-10	41,000	22,140

........

Suppose the Power BI model defines the measure this way, using the documented DIVIDE function:

Gross Margin % =
DIVIDE (
    SUM ( Sales[Revenue] ) - SUM ( Sales[Cost] ),
    SUM ( Sales[Revenue] )
)

A prompt to Genie such as the one below should return a value that matches the measure. An analyst can confirm it by running the equivalent SQL on the same Unity Catalog table, which was executed on the illustrative data before publication.

What was gross margin % by month for September and October 2026? Use the Gross Margin % definition from the Power BI sales model.

SELECT
  month,
  SUM(revenue) AS revenue,
  SUM(cost) AS cost,
  ROUND(100.0 * (SUM(revenue) - SUM(cost)) / SUM(revenue), 2) AS gross_margin_pct
FROM sales
GROUP BY month
ORDER BY month;

........

Month	Revenue (USD)	Cost (USD)	Correct margin (ratio of sums)	Wrong margin (average of rows)
2026-09	200,000	138,800	30.60%	29.50%
2026-10	240,000	158,670	33.89%	35.67%
Both months	440,000	297,470	32.39%	33.20%

........

If Genie returns 35.67% for October, it has averaged row-level margins and has not applied the Power BI definition. Teams that already validate Genie output on other sources, for example the Genie Agents that answer from PDFs and slides in Unity Catalog volumes, can add a short list of reference measures like this one to the same test routine.

··········

GENIE AND POWER BI MEASURES: WHAT DEPENDS ON YOUR SETUP.

Databricks confirms the core facts: external knowledge sources are in Beta, Power BI is the only supported source, metadata comes through an OAuth Machine to Machine Unity Catalog connection used by a Power BI task, it refreshes every 7 days and stays in the workspace's region, and snippets are created only for semantic models that read from Unity Catalog objects. RLS definitions, owner emails and usage activity are excluded.

What depends on each organization is mostly licensing and governance. The Power BI task path requires a Premium-tier Power BI license or Fabric capacity with XMLA read-write, a SQL warehouse and serverless compute access, and the external knowledge feature is not available in workspaces with the compliance security profile. The account-wide visibility of Power BI snippets means the choice of workspaces granted to the service principal is the main control, since there is no per-group restriction.

What Databricks has not disclosed is how DAX is mapped to SQL, which DAX patterns are supported, any limits on model count or size, and what the feature will cost once snippet curation is priced. Until those details are published, the safe approach is to enable it on a small set of shared, well-tested measures, compare Genie's numbers against the Power BI report and the equivalent SQL, and expand only after the results match. A similar check applies on other platforms that are adding metric logic to their semantic layers, such as Snowflake's LOD metrics in semantic views.

··········

·····

DATA STUDIOS

·····

datastudios.org