Replies: 2 comments
-
Thanks! We are aware that judicious use of examples increases accuracy. There are some examples in the TypeChat code, such as in this commented api from the music application.
Like with calibrating the amount and directness of repair instructions, we are taking a minimalist approach at this point, seeing how far we can go with minimal repair, instruction, and examples. This gives us head room as we take on larger applications, so that we can tune up these additional filters in a way that aligns with concrete scenarios. |
Beta Was this translation helpful? Give feedback.
-
on conversation history see #39 |
Beta Was this translation helpful? Give feedback.
-
These models all benefit from seeing themselves follow instruction so you really need to feedback to them their chat history with the user.
gpt-3.5-turbo
in particular, benefits from including a 1 shot example of it returning the JSON you expect back. Even if you don't include a 1 shot example you'll find that the more you show the model returning the responses you expect the more reliable its output will be. The current prompt doesn't look to have any provisions for sending the model conversation history so wondering if that's been thought about?Additionally, all of the newer OpenAI models use the Chat Completion APIs which are message based so you really want to pass that conversation history as a message array.
Beta Was this translation helpful? Give feedback.
All reactions