Hi! I am trying to create a sales assistant which would recommend products. When I created it in the GPT builder, it was working flawlessly. Recommending good products, asking relevant questions etc. However, when I tried to achieve the same results in assistants API, my model seems so much more worse. It talks about random products, recommends wrong links… Both models have the same files and settings. Why is this so? Any tips? Any links to docs or something? I really can’t figure out what is going on here.
The “secret sauce” as it were, is in the instructions given to the models with respect to how to use those tools.
You’ll need to do some more work on your system message to bring the quality in line with OpenAI’s offering.
I am using retrieval tool. So I have to play around with the instructions until my model acts like the one in GPT builder?
I kind of tried this, but the more I add to instructrions the more expensive it becomes and the results are still not that great.
It sounds a little bit misleading.
I think, this is the real answer: Custom GPT and Assistant API does not have the same engine under the hood.