Excessive CU usage in Fabric IQ

[joaomacp]

I’m running a tiny PoC with Fabric IQ and hitting a bottleneck with capacity consumption.

The setup is super simple:

Capacity: F4

Data: 2 tables from Lakehouse. A "Sale" fact table with 2k rows and a "Salesperson" dimension with 25 records.

Ontology: 2 nodes (Sale and Salesperson) and 1 property.

Data agent: connected exclusively to the ontology.

I asked the Data Agent exactly 10 questions, and my capacity went straight into background throttling. Looking at the metrics, 90% of the CUs were swallowed by GraphIndex (Graph General Operations).

Is it normal for such a small dataset to trigger this much usage? Has anyone else seen this or found a way to optimize it?

Attached the screenshot of the spike.

Before:

After 10 prompts:

Any tips?

[arabalca · Accepted solution]

Hi @joaomacp ,

The same thing has happened to me. We are currently testing and, while I understand that optimizations will improve over time, I think it is important to acknowledge something: AI workloads like this are not designed to run efficiently on small capacities such as F2 or F4.

In my opinion, Fabric IQ brings a lot of value and is something that you almost need to have in your roadmap. However, we also need to be aware that this value comes with a cost. This is where a key aspect comes into play: DataOps.

We are moving into more complex scenarios (real-time, agents, graphs, etc.), and this means it is no longer just about building solutions. We need to start thinking seriously about:

• The ROI of each use case
• How to optimize capacity consumption
• Which workloads truly justify this level of investment

As an additional recommendation, I believe it will be key to:

• Properly separate environments (dev/test/prod) to avoid unnecessary capacity consumption
• Monitor usage from the beginning
• Prioritize use cases with real business impact

Thanks for sharing your experience. If you notice any updates or improvements, it would be great if you could share them

[joaomacp]

Hi @arabalca , thanks for the reply! I'll definitely keep the tests running and share my future findings.

[arabalca]

Hi @joaomacp 
if this helped you, I’d appreciate it if you could mark it as the accepted solution and give it a like 👍
This helps others in the community find the answer more easily.
Thanks! 🙌

[arabalca]

Perfect @joaomacp ,if you find the answer helpful, you can consider it resolved

I remain at your disposal for any further questions.

Best regards,

[joaomacp]

Done!