Python: Auto Function Calling - Passing Arguments that are not part of the prompt

[vslepakov]

Hey folks,

When using Semantic Kernel (Chat Completion, Assistant Agent, or AI Agent) with auto function calling, is there a way to pass additional arguments to the functions - such as user context information - that are not initialized based on the prompt?

Thanks!

[moonbox3]

Hi @vslepakov, would this help your use-case? When you define a kernel function with the arguments parameter, they KernelArguments will get included as part of function invocation.

[moonbox3]

Hi @vslepakov, you can provide arguments per function invocation with a single AzureAIAgent:

• get_response(...) supports invoke-level argument overrides:

  semantic-kernel/python/semantic_kernel/agents/azure_ai/azure_ai_agent.py

  Line 187 in f80e080

  arguments: KernelArguments | None = None,

• invoke(...) also support invoke-level argument overrides:

  semantic-kernel/python/semantic_kernel/agents/azure_ai/azure_ai_agent.py

  Line 252 in f80e080

  arguments: KernelArguments | None = None,

• invoke_stream(...) supports the same

The invoke-level arguments take precedence over constructor provided arguments.

[vslepakov]

Thanks @moonbox3 , that's a good one, I missed that. I tried like this:

 arguments = KernelArguments(request_context=request_context)
 response: ChatMessageContent = await ai_agent.get_response(thread_id=thread.id, arguments=arguments)

but the passed in KernelArguments only contain the arguments from the prompt (just test data) and the arguments object is empty:

request_context is a complex object. Do I need to pass in a serialized json string and then deserialize?

[moonbox3]

Hi @vslepakov, there's a bug in AgentThreadActions. We aren't passing in the arguments. Will fix right now.

[moonbox3]

@vslepakov, I will also solve the other issue you raise related to not being able to properly handle the AutoFunctionInvocationContext with this bug fix. That'll take me a bit longer, but I will get a PR out with both soon.

[vslepakov]

fantastic, thanks @moonbox3 !

[moonbox3]

Fix is in #11250