[boxcarr]

Not sure if you found this as well, but gpt-3.5-turbo-016 does a poor job following instructions other than parsing. So, to work around this, we prompt gpt-3.5-turbo with the rules we want applied say an extracted field and then go back to gpt-3.5-turbo-016 to parse with chatgpt functions.

Bottom line, every single post requires approximately 10 different prompts to refine the extraction.

[hubraumhugo]

yeah I had the same experience, divide and conquer works better than trying to do everything with a single prompt. Are you interested in comparing notes? Feel free to ping me via the email in my profile :)