We have been using function calling for quite some time now. We had multiple functions provided to the model, for it to optionally choose one within the “auto” mode. With the latest switch to “tools” from functions, it looks like the model simply ignores the description of the tool, and insists on calling it even if it is not needed to. So, even “never ever use this function/tool in these cases: …” type of descriptions cannot prevent it from calling the tool anyways. Do you have any suggestions?
Edit: It looks like the gpt-4-1106-preview model handles the description better. But gpt-3.5-turbo-1106 model simply cannot comprehend the instructions in the tool description.
I’m having the same issue where i have tool calls on auto and I explicitly mention in the description of one of the tools to call it when required and yet it never calls it unless I respond in the chat with the next message saying “why didn’t you call the function???” so odd (I’m using GPT-4 and not gpt-4-1106-preview but I’ll try gpt-4-1106-preview)
The tool description should describe what it does and what it returns, why it is useful to the AI for certain user inputs. Not with rules for the AI, which won’t get the best results.