Python: Selective Function Invocation in Semantic Kernel

[vslepakov]

Hey folks,

I'm using a ChatCompletionAgent with multiple plugins and have set FunctionChoiceBehavior.Required() to ensure certain plugin functions are always invoked, keeping data fresh. However, this causes unintended function calls for prompts like "Who are you?"

Is there a way to use FunctionChoiceBehavior.Auto() to let the AI decide when to call functions, while still enforcing invocation of specific plugin functions when necessary? Right now, when using auto function calling the LLM sometimes decides not to invoke a function because the data (might be stale) is present in the chat history.

Thanks!

[moonbox3]

Hi @vslepakov,

As you're seeing, when you configure .Required(), you're telling the model that it must invoke a tool/plugin during every turn -- which can lead to unexpected behavior for general prompts like "Who are you?".

From your description, it sounds like you'd prefer the model to use tools when needed based on context, but still ensure fresh data when the situation demands it. That leans more toward .Auto() behavior, but the downside (as you’ve noticed) is that the model may skip calls if stale responses are already present in chat history.

To better understand your case and explore potential workarounds, a few questions:

1. What exactly defines a tool result as "stale" in your flow -- is it time-based, content-based, etc.?
2. Are there cases where previous tool call results are actually helpful (for example, to save latency or avoid redundant calls)?
3. Would performance be impacted if you forced a fresh tool call every time for certain queries (while using .Auto())? This perhaps could be accomplished either via instructions or a system message you supply along with other messages during invocation.

[vslepakov]

Hi @moonbox3, sorry for the delayed response.

1. It is primarily time-based, but we also want to handle scenarios where data might be modified via another client (like the web or mobile app) without the user being aware.

2. Yes, but as mentioned above, for certain operations it's important to ensure the data is current before proceeding.

3. Correct - there is a performance cost to calling APIs on every interaction. Ideally, it would be great to mark certain functions to always run (when using .Auto()), perhaps using a decorator to indicate which ones should be always invoked.

As a workaround for now, we're using .Required() with a No-Op plugin - a dummy function that essentially does nothing, described as something like "Use this function as default when no other available functions are appropriate in a given context." This approach works well when paired with the right system prompt.

Thanks