Fix optional args toolcalling

[aymeric-roucher]

No description provided.

[albertvillanova]

I think there might be some confusion around the term "nullable" in this context.

We should clarify terminology: "nullable" can be interpreted in two ways:

• either that the argument can explicitly be set to None,
• or that it's optional in the sense that it has a default value and doesn't need to be passed at all.

I think that distinction is important when reasoning about how LLMs should construct tool arguments.

[aymeric-roucher]

"nullable": True should be equivalent to "required": False, I agree with refactoring in a follow up PR to switch to using required instead of nullable. Do you see any inconsistency with this? Else we should go ahead and merge this IMO, it's a required fix else validate_tool_arguments is a breaking change.

[albertvillanova]

Thanks. I just added a complete test suite (some of them to be addressed in a subsequent PR).