Hi all, I am passing 4.1 a big markdown dump that is a pdf extraction of about 16 pages. Easily enough for it to handle. I ask it to format it into a markdown table that can be easily saved as a CSV. 4o can do this, but I’m migrating.
The problem is 4.1 gets to like page 5 and then writes “the rest of the table would follow this format, let me know if you’d like me to do the rest of the table”.
I want it to do the whole thing in one go, as it’s able to! I’ve tried adding to the start and end of the prompt (I’ve heard that works)
"Please extract every page of the text I give you and do not truncate the output
{extracted_text}.
Please every page. Do NOT stop until you have extracted every page. Do not ask for the users permission"
Does anyone know How I can get it to keep going?
3 Likes
It sounds like most people migrating to gpt-4.1 should expect to re-engineer their prompts per the cookbook guide from OpenAI on gpt-4.1 prompting / instructions. It won’t let me link you to it here - but they posted it April 14.
I think you’re on the right track with your firm and clear instructions but language is important. It may not know what a “page” is necessarily depending on the markdown - maybe use another specific identifier in the structure of the markdown that identifies the end? I would play with how you define the entire scope, or how you identify the end.
If it was more than 16 pages and your output needs to be CSV, I would say you could batch into a couple API calls and compile the CSV after - but you really shouldn’t have too… 16 pages should be totally do-able with one call.
Although… maybe double check your token usage just in case you have 16 pages crammed with incredibly small text 
Hope something you try works!!
Any part of the guide you recommend for addressing this? There isn’t much discussion on markdown other than they recommend using it.