Is the LLM forced to generate a valid json?

[michelonfrancoisSUMMIT]

Hello,

I am working on an agent based project. I am using an object of the class LlamaCppAgent (here the version of the lib i use: llama-cpp-agent==0.2.35). I am using the "get_chat_response" method on the object with a message and the structured_output_settings parameters containing tools. My question is : "Is the LLM forced to generate a valid (with curly brackets, the correct arguments as a key and an argument value of the correct type) json?". I have not been able to be sure about that when looking at the documentation. I am using the "grammar' argument for other use cases and here the LLM is forced to follow a strict output format (due to token filtering), is the case also when using tools in my case? I have tested a lot the LLM and for now it seems to work fine but i want to be sure.

Best

[michelonfrancoisSUMMIT]

In turns out it is.

I have checked the code and from what i understood, the agent class return a grammar object which forces the LLM to follow it.