Assistant APIs provide a well thought architectural mechanism to interact with LLMs. They also provide powerful capabilities through the underlying metadata structures. That said, the current implementations of Assistant API , in beta, leaves much to be desired. In certain cases, the Assistant API is just episodically buggy. In certain other cases, the Assistant API does not provide the necessary verbs (such as temperature) to enable certain use cases.
In a nutshell, it is a love-hate relationship. We see the promise of AssistantAPI; but we are brought down to ground by it’s bugs and incomplete feature set.
openairetro (GitHub - icdev2dev/openairetro) is a set of demonstrations of the betaassi (GitHub - icdev2dev/betaassi: "betaassi: An innovative Python package for seamless integration and management of OpenAI sessions, designed to enhance developer experience with asynchronous server support and extendable classes) framework; that makes extensive use of the metadata feature embedded in the AssistantAPI framework to be compatible with the AssistantAPI at the data plane; but use ChatComplation for execution plane.
What this means is that certain features which are currently not feasible within the AssistantAPI becomes available through betaassi. For example, the temparature example include temparature as a actionable field on the Assistant structure. This means that you can set the temparature in such a manner to make the response as deterministic or random as required.