Llama 2 function calling Llama 2 with grammar-based decoding (provided by llama.cpp) for constraining the output to a json schema.