Problems with Assistant and GPT-4 Turbo

Generally, amazing :clap:t2:

But I have encountered a few problems with GPT-4 Turbo, particularly when using the Assistant API:

  • I have not been able to guide it to follow the instructions in the Assistant API (“produce the response in HTML”, for example). This was not a problem with GPT-4. Even with just a call to the completion endpoint, it struggles to obey instructions
  • this is likely due to demand, but the Assistant API run creations get cancelled fairly regularly
  • although the file size limit it 256MB for the Assistant, I have found that when the files are over 3MB or so, the messages take a long time to return. I would imagine the embeddings are created when the Assistant is created, so this is odd, but fairly reproducible


1 Like

Do you have a specific conversation / playground link showing the “produce the response in HTML” issue?

Not to hand, I will try to reproduce it.

But that is not my main concern. The API appears to ignore the “mood” instructions and the instruction to mirror the language used in the documents.

Is there something I can do about that?