The answer here is very much “it depends”
-
Different LLMs perform better or worse… for example you have GPT 3.5 (cheapest / worse) … GPT 4o (more expensive / better). I would recommend you reduce variable by going to your LLM page and disabling GPT-4 and 3.5 turbo and only having 4o enabled. https://forum.georgegammon.com/admin/plugins/discourse-ai/ai-llms
-
System prompts matter, if you want the LLM to behave in a certain way, include some examples of “good behavior” in the system prompt.
-
There is an aspect of randomness as well which is hard to control.
Please retry with GPT 4o and see how you go.