Forum Discussion
Fabric Data Agent CU Calculation based on Query
Your estimate only includes LLM token usage, but Fabric Data Agent CU includes full execution cost.
The gap comes from:
-
multiple LLM calls per query (not just one)
-
schema + prompt expansion on every call
-
SQL/Spark execution on data
-
retries and query corrections
-
metadata scans and grounding
So 10 queries ≠ 10 executions. It is usually 3–7x more operations, and data compute dominates CU, which is why you see ~54k instead of ~1k.
Recently I discussed about this in detail in my blog. Hope this will help.
-
Your estimate only includes LLM token usage, but Fabric Data Agent CU includes full execution cost.
The gap comes from:
-
multiple LLM calls per query (not just one)
-
schema + prompt expansion on every call
-
SQL/Spark execution on data
-
retries and query corrections
-
metadata scans and grounding
So 10 queries ≠ 10 executions. It is usually 3–7x more operations, and data compute dominates CU, which is why you see ~54k instead of ~1k.
Recently I discussed about this in detail in my blog. Hope this will help.