Data Sources and Access for Data Agent Assessments

Many data agents answer from lakehouses, warehouses, KQL databases, SQL databases and mirrored databases as well as from semantic models. You can assess these data agents in BI Pixie too, so you can see where the setup of each source doesn't follow Microsoft's published guidance for that kind of source.

Prerequisites

The Contributor role or higher on the workspace that holds the data agent.
For a warehouse, mirrored database or SQL database, or for the descriptions on a lakehouse's tables, permission for BI Pixie to read OneLake storage as you. See Grant access to read a source.

Review what is checked on a source

For each source other than a semantic model, you get findings on:

The data source instructions.
The data source description.
The example queries.
The schema object descriptions on tables and columns, for a lakehouse, warehouse, SQL database or mirrored database. A KQL database doesn't support them.
Whether the source duplicates a semantic model the data agent is also grounded on. This happens when a data agent uses both a semantic model and the Direct Lake source it reads. Only the semantic model carries the measures, relationships and descriptions, so a question sent to the copy is answered without them.

If you select Ask your AI, your AI also judges the wording of each source, including:

Whether the description says what the source contains and tells it apart from the data agent's other sources.
Whether the instructions define tables and columns one by one, which belongs in each object's own description.
Whether each example question is clear and relates to its query.

If reports over a Direct Lake semantic model built on the source have Pixies, you also see whether the columns report viewers use most have descriptions. For the full list of checks, see Score dimensions.

To fix these findings, open Fix on your own on the assessment and make each change in Microsoft Fabric, where the finding says. BI Pixie writes changes only to the data agent itself.

Note: Benchmarks and A/B tests compute their correct answers from a semantic model. You can't benchmark a data agent, or use it in an A/B test, when it is grounded only on sources other than semantic models. You can still assess it and apply the changes BI Pixie offers.

Review what is read from each kind of source

Source kind	What Microsoft Fabric uses from the data agent's setup	What BI Pixie reads
Semantic model	The selected tables and the data agent's instructions. Data source instructions, descriptions and example queries aren't supported. Fabric uses the AI Instructions in Prep data for AI when it writes a DAX query.	The full semantic model definition, and its own AI Readiness score if it has been assessed.
Lakehouse	The selected tables, views and functions, data source instructions, a data source description, example queries, and schema object descriptions on the preview runtime.	Table names, with the Fabric permission you already hold. Table and column descriptions, after you grant access.
Warehouse, mirrored database, SQL database	The same as a lakehouse.	Table names and descriptions, after you grant access. Until then, the checks that compare the setup with the source's tables don't run.
KQL database, listed as Eventhouse in Data sources	The selected objects, which can be tables or materialized views, the data source instructions, a data source description and example queries. Schema object descriptions aren't supported.	The database schema from the eventhouse item, with the Fabric permission you already hold. The schema says whether each object is a table or a materialized view. If no object names come back, the checks that need them are left out.
Mirrored Azure Databricks catalog, and other kinds	Whatever Microsoft Fabric supports for that kind of source.	Nothing. The source isn't part of the score. A kind BI Pixie doesn't recognize is listed as Not recognized yet.

If BI Pixie can't open a source, you see the reason on the source's card, and the checks that needed it are left out of the score.

On a KQL database, the source's row reads Objects used rather than Tables used when at least one selected object isn't a table, or when BI Pixie couldn't read what kind each object is. The findings that name what the data agent selected say object rather than table on a KQL database. A KQL database holds materialized views beside its tables, and a data agent's own setup doesn't tell the two apart. To read the names of what the data agent selected, grouped by kind, select Show what this data agent selected on the source's row.

Grant access to read a source

To read the tables in a warehouse, mirrored database or SQL database, or the descriptions in a lakehouse, BI Pixie needs permission to read OneLake storage as you. If data residency is turned on for your account, you have already granted it and can skip this section.

When an assessment couldn't read a source for lack of this permission, you see this sentence on the source's card and on the assessment:

BI Pixie needs your permission to read the tables this warehouse holds. It reads the names of the tables, never the data in them.

Granting access starts a new assessment of the data agent, which counts toward your plan like any other. To grant access:

Under the sentence, select Grant access.
In the Microsoft dialog, accept the permission.

After you grant access:

You follow the new assessment on the data agent's page.
The permission is yours, not the account's. Each teammate grants it on their own.
BI Pixie reads table names and descriptions only. No row of your data reaches BI Pixie or your AI.

If you don't grant access:

If you close the Microsoft dialog, nothing changes and no assessment starts.
If your organization's consent policy refuses the request, you get text to send your administrator and a link to the list of every permission BI Pixie asks for.

Who can assess and change a data agent

To assess or change a data agent, you need the Contributor role or higher on its workspace. BI Pixie acts as you and has no access of its own to any Fabric workspace.
Anyone on your BI Pixie account can read a stored assessment, including someone who has since lost write access.
Without write access, you can't reassess or apply a change. If you try, you see which workspace and which access you need.