Assistant API Retrieval Falls Short of Identical Custom GPT Retrieval

Totally agree (and hope they do that).

I did create Assistant API versions of these same custom GPTs BTW and discovered something interesting and strange: the responses are not similar at all in terms of quality in spite of having the same setup.

The custom GPT responses are much more detailed and coherent based on the knowledge we used to train them. They make the Assistant API versions seem almost 3.5 GPT-like by comparison. I tried every model option there to see if that might matter and it didn’t at all. Maybe they are previewing GPT 5 with custom GPTs behind the scenes or similar?

Also, Assistant API performance is frequently terrible. Response times of 30-60 seconds per reply are not uncommon.

That’s why I’m holding out for better member management in Teams, or biting the bullet and going for Enterprise. Believe it to not, members on that plan cannot add other members. So go figure.

What do you mean “knowledge used to train them?” You’re not able to train GPTs.

They almost certainly are not.

What is more likely is that the system message you are providing to your assistant isn’t providing the same quality of guidance for retrieval as the one used in custom GPTs.

Training: the content we upload that forms the basis for the custom GPT’s unique knowledge, plus the instructions we provide to guide its use of that content and its responses that use it. I’m not suggesting we are training a model here.

I see. That’s what I was assuming you meant, but I wanted to verify.

Training isn’t really the right term to use here and can definitely lead to confusion.

But, more to your issue (and I’m going to move these posts to a new topic as we’re wandering off-topic here), part of the system message for a custom GPT with knowledge files reads,

You have files uploaded as knowledge to pull from. Anytime you reference files, refer to them as your knowledge source rather than files uploaded by the user. You should adhere to the facts in the provided materials. Avoid speculations or information not contained in the documents. Heavily favor knowledge provided in the documents before falling back to baseline knowledge or other sources. If searching the documents didn"t yield any answer, just say that. Do not share the names of the files directly with end users and under no circumstances should you provide a download link to any of the files.

I would be interested to know if adding that (or something similar) to your assistant system message closes the gap at all in terms of the observed performance you’re seeing.

1 Like

I’m unable to include links here for some reason, but check out this post on the same topic.