Two ways, one is to pass in the OpenAI style function defintion JSON string in the system instruction, and then prompt the LLM. Another is to use pydantic models to grammar, you can see example code here https://github.com/ggerganov/llama.cpp/blob/master/examples/pydantic_models_to_grammar_examples.py

Having gotten these to work, the best way would be to create a facade/wrapper function around the openAI chat, that will intercept the tools being passed, then utilize one of the above methods. By doing this, you can reuse any project that uses OpenAI for tool calling. Make sure the underlying model supports tool calling, most of the latest ones do. Even the tiny 1b-3b models have worked for me.