Wrapper for Non Tool-Calling Models to Enable Auto Function Invocation | python

[Vizhy]

---

name: Wrapper for Non Tool-Calling Models to Enable Auto Function Invocation
about:
A wrapper to simulate tool-calling functionality for models that lack native support.

Context Preparation:
From the ongoing request, the wrapper creates a context.tool_calling object containing request.functions, request.function_call_method, and related info about formatting the response for function calling.

Request Encoding:
The wrapper injects context.tool_calling into request.messages with appropriate AuthorRole to embed tool call instructions into the prompt

Response Parsing:
The wrapper parses the combined model response—which contains both standard reply and simulated tool call data—and splits it into separate objects, matching the structure of native tool-calling model responses before forwarding to Semantic Kernel.

Purpose:
This approach extends Semantic Kernel’s usability to models beyond OpenAI and Azure OpenAI, which currently are the only ones with native tool calling support and the only ones capable of fully leveraging Semantic Kernel’s orchestration capabilities.

[moonbox3]

Hello @Vizhy, I am not sure I understand this request. We support many other AI services that allow function calling. For other AI services, aside from OpenAI/Azure OpenAI, the prompt execution settings will be configured with tools while calling the model. One can control which tools are broadcast to the model via FunctionChoiceBehavior and filters in Python: https://learn.microsoft.com/en-us/semantic-kernel/concepts/ai-services/chat-completion/function-calling/function-choice-behaviors?pivots=programming-language-python.

[Vizhy]

Hello @moonbox3
Thank you for the response.

I'm noticing that auto function invocation isn't working with non-tool-calling models.
Since the only key difference between tool-calling and non-tool-calling models is the structure of the request/response schema (i.e., tools in the request / tool_call in the response), it's likely possible to simulate this behavior.

We could create a wrapper/filter that:

• Appends the tools and related function call context directly into the messages field of the request.
• Parses the response to extract any tool call intents and presents them in a structure similar to tool-calling models (response.tool_calls).

This would effectively make a non-tool-calling model behave like a tool-calling one from Semantic Kernel’s perspective, enabling function invocation features for a wider range of models.

for example:
I am using OllamaChatCompletion as the ChatCompletionClientBase to connect to a model gemma3:4b

semantic_kernel.exceptions.function_exceptions.FunctionExecutionException: Error occurred while invoking function fpElIESyVkpgEsGQ: registry.ollama.ai/library/gemma3:4b does not support tools (status code: 400)

[moonbox3]

@Vizhy

I'm noticing that auto function invocation isn't working with non-tool-calling models.

That’s right -- auto function invocation only works with models that support tool-calling. If a model doesn’t have tool-calling, there’s no way for the system to know when to call a function. While it’s technically possible to try to simulate tool calls by parsing the model’s plain text outputs, that approach is unreliable and not something we plan to support. There are plenty of models available, including for Ollama, that natively support tool-calling and work well with auto function invocation.

[Vizhy]

@moonbox3
Thank you for the response
Understood