Assistants API - prompt requirements and OpenAI "self" knowldge

I am trying to better understand the “knowledge” limit that the API is using when creating an assistant and adding messages/runs using a thread.
For example, I build a small travel agency assistant, giving some prompt text to tell the API what I am trying to do (a few sentences).
From this point on, if I am asking a more “general” question about places and information that is not in the prompt, I do not get relevant answers.
Must I give the assistant ALL the information that he needs to answer my assistant questions? he does not use his already trained OpenAI model information?

1 Like

If you share more details we could be better helping. Just yesterday we looked into function calling for a movie database where it looked like GPT was very eager to use its vast knowledge about movies even though explicitly asked NOT to.

Try to be as explicit as you can in your instructions. It should in principle use its existing knowledge but in your instructions you can specify under which circumstances it should do that versus when it should use the proprietary information that you supply such as via file uploads.

Assistants will fail you in quality output when using retrieval.

Why?

Retrieval injects a massive amount of documentation even when not relevant to the user question.

Retrieval places a document search function that the AI must call to find out if there is anything more relevant, wasting your tokens, and returning results even when irrelevant.

The return values of a function placed into context are trained to be answered from directly but they can have zero value.

Moral: Do not build products on agent platforms out of your control.

1 Like

I did not use retrieval, or any other tool.
I just dropped some prompt instructions, expecting the OpenAI to use his “knowledge” just like asking the ChatGPT.

Can you perhaps give a specific example?

I think I figured it out - the prompt must not limit the engine, but have to be broader and then it may not “fall back” into itself and be able to provide answers.

1 Like