How to Train the Data agent based on particular dataset

[udhaya208]

Hi Everyone,

Hi Everyone,

I’m currently working with the Data Agent using an F64 capacity. We have over 100 tables and more than 10 views in the Lakehouse. However, when we ask specific questions related to the selected data, the Data Agent does not provide accurate answers—even though we've already configured it with multiple AI instructions and query examples.

My concerns are:
What kind of specific training or AI instructions should we configure to improve the accuracy of the Data Agent’s responses?

Are there any tips or best practices to ensure we get more accurate and relevant answers from the Data Agent?

@Fabric Data Agent

[v-nmadadi-msft]

Hi @udhaya208 

I wanted to check if you had the opportunity to review the information provided. Please feel free to contact us if you have any further questions. If our responses has addressed your query, please accept it as a solution and give a 'Kudos' so other members can easily find it.

Thank you.

[v-nmadadi-msft]

Hi @udhaya208 

May I ask if you have resolved this issue? If so, please mark the helpful reply and accept it as the solution. This will be helpful for other community members who have similar problems to solve it faster.

Thank you.

[udhaya208]

Hi @v-nmadadi-msft,

Thanks for your time to respond and provide the solution. But still I already did everything based on your suggestion. The answers are not accurate, evethough I provided right prompt and feeded lot of AI instruction. for ex: if we ask Actual 2024 Vs Budget 2024 it is providing the Actual Vs Budget for 2025 values. I believe some more way we need to train the data agent.

[DataGuru412]

Hi,

As per my understanding there are 2 -3 steps you need to follow as there is no 100% perfect result for AI tool (this is my understanding)

1. Check the columnNames it should be align for mapping mostly self sufficient to describe to AI what it created for. 

2. Similar to column name table Name should be self sufficent to describe

3. provide all the possible  relevant query  and questions

don't expose all the tables until unless it is not required.

I hope this may help.

[burakkaragoz]

Hi @udhaya208 ,

I’ve worked with Data Agent in Fabric a bit, so I get where you’re coming from. Training it to give relevant answers—especially with so many tables and views—can definitely be tricky.

For better accuracy, here’s what usually helps:

• When you set up AI instructions, try to be as specific as possible. Instead of broad directions, give the agent clear, targeted examples of the kind of queries and answers that are correct for your dataset.
• It helps a lot if you limit the scope of the agent to only the tables/views needed for a particular set of questions. If it’s referencing too many objects at once, accuracy drops.
• Try to give a few “bad” examples too, so the agent learns what NOT to answer (sometimes negative examples are just as helpful).
• Naming conventions matter: if your tables/columns have confusing or inconsistent names, updating those to be more descriptive can sometimes improve response relevance.
• For best practices, retrain or update your instructions after you notice which type of questions it’s missing. It’s not always a one-shot setup—usually needs a bit of trial and error.

If you’ve already tried some of these and still getting off answers, let me know a sample scenario, I can maybe share how I would phrase the training examples.

If my response resolved your query, kindly mark it as the Accepted Solution to assist others. Additionally, I would be grateful for a 'Kudos' if you found my response helpful.