I’m curious if anyone knows if the AI accessed through the API considers all API calls to be part of the same chat session.
The AI doesn’t remember anything of the exchanges; it needs to be reminded every time of what was before that. It sounded counter-intuitive to me at first, but knowing that GPT is all about completion, it makes sense. The ChatBot of AI Engine sends the whole chat, to make it sounds more like a discussion.
In the case of fined-tuned models, I actually don’t send the whole discussion anymore, as in most cases I don’t think we need it; the visitor wants to ask some questions and the context of the conversation is not as important (and also, fine-tuned models cost more, so better to avoid using so many tokens).
Secondly, it seems like asking it to use markup is inconsistent. I just asked it to write a press release. The first time it used much larger headings and the second time they were smaller.
I think I improved the prompts a lot on the very latest version, have you tried that one? Same with the number of paragraphs, it should be much better. However, that always depends on the type of title, and the digital mood of GPT 😉
On the right of the screen, you can check the prompts; you can play with them, and try to find if you can get a better one. If you do, absolutely let me know! But I am sure that little by little we’ll find out the best prompts.
I hope it helps 🙂