Python: New Feature: More granular spans for AutoFunctionInvocationLoop

[HuskyDanny]

We are debugging the performance issue for the agent calling which uses the functional calling that uses the AutoFunctionInvocationLoop.

From the trace, we see the span at AutoFunctionInvocationLoop level which took a quite lot of time and below it are some tool calls.

Can we add more granular spans on the AutoFunctionInvocationLoop? Specifically, I want to identify which part is taking the majority of time, I know there is functional calling to decide which tool to call, can we have this span in as well?

Overall, I feel this is important to diagnose agent performance issue.

The SDK I am using is semantic-kernel==1.28.0 in python

[markwallace-microsoft]

@TaoChenOSU what are your thoughts on this enhancement request?

[HuskyDanny]

@TaoChenOSU Any thoughts on this?

[TaoChenOSU]

Hi @HuskyDanny,

Thank you for raising the question!

It looks like you are not capturing the model invocation spans which probably take up the most time. If you have the model invocation spans captured, you will see something like this:

To capture the model invocation spans, please refer to this documentation.