I’m using the assistant API using GPT-4o to review an image of a customer purchase order and output JSON.
I had this working [almost] perfectly however whilst re-visiting the project over the past few days, i’ve noticed a considerable drop in accuracy. It’s making up information and will most [if not all] outputs have been incorrect.
It is outputting some correct information, so i know if can see the document.
Interestingly, when i give the same image with the same prompt to chatgpt, it performance almost perfectly using the GPT-4o model.
I cannot understand why this would be so dramatically different. I’ve adjusted the temperate (currently set to 0.01 temperature).
Has anyone else had similar issues? I’d really love to use the assistant API so i can take advance of the retrieval etc.
Any feedback would be great.