Tool calls: recommendations to improve NON deterministic arguments when using `gpt-3.5-turbo-1106` model

I’m testing the “tool/function calls”.

The declared tools/functions in my requirements can have many parameters (e.g. 5-7 parameters).

With the model gpt-3.5-turbo-1106 I get very non determinsitic arguments. It is like every 3-4 API calls they provid totally different outputs.

When defining the arguments in the tools field, I specify a description and type only.

Do you have any recommendation to improve the accuracy and determinism of the arguments in the API response?

What is your experience with it?

The fix seems to move to gpt-4-1106-preview (I get much better reliable, consistent and determinstic results from it) however it is not ready yet for production use.

Welcome to the forum, Giovanni!

The quality of LLM outputs (especially when we are talking about apps where a stable predictable output is required) are very much dependent on the prompt engineering skills and sometimes are limited by the current capabilities of the model.

Having said that, if you want to get a better answer, you’ll need to share your prompt.

1 Like

Hey @giovanni2,

did you find your way to a satisfying solution for deterministic output via tool calling? Any experiences to share?

Thanks!