Hi OpenAI team,
Currently, in the Assistants API, you cannot control things like maximum outputted tokens, temperature, top_p etc. like you can do in the Chat Completions API. I understand that this might have been intentional as the Assistant’s API is an abstraction away from a lot of this complexity for new API devs, but these parameters are insanely useful for some applications. For example, having a deterministic temperature at 0 to replicate faults is huge.
Some might say “Just switch to the completions API then”, but unfortunately you then have to compromise on some of the great features and simplicity that Assistants brings, like prompt templates, code intepreter, easy file uploading etc. Currently, it’s a very big tradeoff that you sacrifice a lot either way.
If you could implement more control over model responses with something like an “advanced settings” in assistants, this would be super helpful to lots of devs.
Edit: Or at minimum, would it be possible to expose parameters chosen within the Assistant’s API?