Assistant API training possibility

Is it possible to train Assistant API with email templates and emais so that it will generate similar emails?

If yes how could I do this?

AFAIK, fine-tuning is not currently available for the Assistant API.

You could likely do what you’re wanting without fine-tuning, though, with a smart prompt and maybe a one-shot at the most.

More details on what you’re trying to do?

Probably better to do it on normal API and build your own RAG system on the backend to have more control - and be able to use a fine-tuned model.

1 Like

Thanks. Is it possible to train GPTs and use them?

No, they’re meant to be used via ChatGPT Plus…

If it is possible could you please explain what is smart prompt and what is one-shot.

Thanks in advance.

By “smart” I just mean well-formed and useful for an LLM. Without seeing it, I can’t tell you the quality of yours.

One-shot means you give the LLM one example in your prompt…

I wouldn’t use “assistants” as a default unless you need very specific access to a particular facet of that ecosystem.

Also, “training” as a term is overused and rife for misinterpretation, as seen here. A prompt “You are a chatbot that hates kittens” can be seen as one type of training, while building a LLM on 40 terabytes of corpus data via machine learning reinforcement algorithms for months is another “training”.

I have a solution to propose: A function that the AI model can call, with multiple example forms that can be returned. An index function that can list all the forms by name and description.

That gives you an AI that can chat without being distracted by the language of a form always being present. The AI also can be prompted to recall one of those example emails with a “I need to write a tech support reply email using the right business form, apologizing for our poor response time” from a user. Having multiple to call on demand could even allow synthesis of new emails following your format.

This all can happen just with a chat completions endpoint chatbot within your control, with no need to attempt fine-tuning an AI model.

1 Like