Microsoft Fabric Updates Blog

Creator Improvements in the Data Agent

• AI
• Apache Spark
• Data Engineering
• Data Science
• Lakehouse
• Microsoft Fabric
• Uncategorized

• marraskuuta 4, 2025 tekijä Misha Desai 42 292 Näkymät

We’re introducing a set of new enhancements for Data Agent creators — designed to make it easier to debug, improve, and express your agent’s logic. Whether you’re tuning example queries, refining instructions, or validating performance, these updates make it faster to iterate and deliver high-quality experiences to your users.

New Debugging Tools

View referenced example queries

You can now inspect which example queries were retrieved and applied when a user asks a question. The Run Steps view shows you exactly which examples influenced the final query — making it easy to confirm that the right examples were used or diagnose when unexpected results occur.

If you notice the wrong examples being referenced, try refining your few-shot questions or adding more targeted examples for improved accuracy.

View diagnostic details

A new Diagnostic Summary lets you download a detailed trace of the agent’s reasoning steps — beyond what’s visible in the run step view. Use it to analyze internal logic, clean up data for review, or share with the Fabric support team when reporting issues.

SDK Enhancements for Example Query Validation

The Fabric Data Agent SDK now includes tools to evaluate and improve the quality of your few-shot examples. Using the new evaluate_few_shot_examples() function, you can validate each natural language/SQL pair and receive a detailed summary of which examples passed or failed.

result = evaluate_few_shot_examples( examples, 
                                     llm_client=llm_client, 
                                     model_name=model_name, 
                                     batch_size=20, 
                                     use_fabric_llm=True)

After running validation:

• Success Cases – strong examples where the SQL matched the expected result.
• Failure Cases – examples that didn’t align with the schema or question intent.

You can easily convert these into DataFrames for review:

success_df = cases_to_dataframe(result.success_cases)

failure_df = cases_to_dataframe(result.failure_cases)

display(success_df)
display(failure_df)

Iteratively improving weaker examples helps your Data Agent generate more accurate SQL and deliver better results for new questions.

To explore a full working example, you can check out the Data Agent Example Queries.

Markdown Editor for Instructions

Agent and data source instructions can now be authored using Markdown, making them easier to read, structure, and maintain. This helps both creators and users clearly understand the context and intent behind your Data Agent’s behavior.

Use Markdown to:

• Define priorities for data sources.
• Add structured lists, headings, or tables.
• Clarify query logic or special handling rules.
• Provide context on tables, columns, and relationships.

This improvement encourages better documentation practices and clearer communication between creators, users, and the Data Agent itself.

For templates and best practices, refer to the Data Agent Configurations documentation.

Multi-Tasking Flow for Easier Creation

We’ve introduced a new multi-tasking flow that allows creators to seamlessly switch between chatting with the Data Agent and configuring its settings — without losing context or progress.

This workflow was built to address key creator pain points:

• Context switching made simple — Move fluidly between testing questions in chat and editing instructions or examples.
• Faster iteration — Instantly test configuration changes in chat without reopening or reloading the workspace or losing context about the data source schema.

The result is a smoother, more efficient authoring experience that keeps you focused on improving your agent’s intelligence — not managing tabs or reloading configurations.

Why this matters

Together, these updates make it easier than ever to create, debug, and iterate on Data Agents. You’ll spend less time switching contexts or troubleshooting, and more time building intelligent, reliable agents that deliver accurate insights from your data.

To learn more about configuring your Data Agent, visit the documentation: Configure a Data Agent.