Assistant x Chatgpt API Quality (PDF reading)

I was setting up training prompts in Chatgpt to run the Assistant APIs for extracting data from PDF, but I had the unfortunate surprise that the algorithm that works perfectly in Chatgpt does not have the correct answer in Assistant.
Does anyone have any suggestions for changing the configuration? I ran the assistant on gpt-4-turbo and chatgpt, on gpt-4. tks