Forum Discussion

Shiyao's avatar
Shiyao
Icon for Microsoft Employee rankMicrosoft Employee
1 year ago
Solved

Fabric data agent

I created a Fabric data agent and connected it as a knowledge source via Azure AI Foundry agent. I also developed a custom python application and run the Azure AI Foundry agent. But the run info of t...
one lake
  • Anonymous's avatar
    Anonymous
    1 year ago

    Hi Shiyao ,

    You're absolutely right currently, the Fabric Data Agent SDK is indeed limited to the Fabric runtime (like notebooks) and doesn’t support integration directly within a custom Python backend running outside Fabric. And unfortunately, even when running inside Fabric, it doesn’t expose the raw SQL queries the data agent generates during inference or interaction with the Azure AI Foundry agent.

    At this point, Microsoft hasn’t provided a built-in way to log or intercept the translated SQL queries within an agent run via the Azure AI Projects Python SDK. The obfuscated references (like ) are more about traceability than transparency.

    If your goal is to manually or automatically validate what SQL was generated, a workaround could be:

    • Injecting a custom critic or reasoning prompt into your Foundry agent pipeline that forces the agent to explain what SQL it would run not perfect, but it helps with transparency.
    • Raising feedback to Microsoft, since SQL traceability during inference is a critical need for auditing and enterprise-grade validation. This is still a gap in the current agent ecosystem.

    Thanks,
    Akhil.

Anonymous's avatar
Anonymous
Not applicable
1 year ago

Hi Shiyao 

Based on the latest documentation and updates, the Fabric Data Agent Python SDK does allow for programmatic interaction with data agents, including evaluation and inspection of responses. However, it does not currently expose the raw SQL queries generated by the Fabric data agent during a run via the Azure AI Projects Python SDK.

The obfuscated tags like 【92:0†source】 are part of the internal referencing system used to trace back to the knowledge source, but they don’t reveal the actual SQL logic used behind the scenes.

You can, evaluate agent responses using structured test sets. Retrieve detailed logs and evaluation summaries and customize evaluation prompts to judge answer quality.

For official guidance, check out.

If you're aiming to debug or audit the SQL logic, your best bet might be to use a custom critic prompt during evaluation that asks the agent to explain its reasoning or inferred query logic.
------------------------------------------------------------------------------------------------------------------------------
If this solution works for you, please consider marking it as accepted so others facing a similar issue can benefit too.

Regards,
Akhil.