Function calling structure

[ankush13r]

Hello,
I'm working with function calling using my own models, but I'm encountering issues with parsing the model's output into a function tool. Currently, my model produces the following output:

<tool_call>{"arguments": {"arg1": "argument", "function_name": "function"}}</tool_call>

However, when calling the text generation interface (TGI) for chat completion, I receive an error stating that it couldn't parse the text.

Could you clarify how the model's output should be structured to correctly parse into a function tool? Additionally, can this work with streamed output as well?

Lastly, sometimes the model generates non-function output. In such cases, I would like to receive normal content instead of a function call. Could you advise how to handle both scenarios effectively?

Thank you!