Improve analysis of structured (JSON) data

Prompting
,
1

I’m using gpt 4o mini to answer questions based on some JSON data that I include in the prompt. About half the time it returns information that does not exist in the original JSON data or it returns values that are mismatched, for example a “document_id” and “text_summary” that don’t go together.

Can someone please give me any tips to improve performance so these issues occur less frequently?

I’ve tried:

  • explaining the JSON file structure in the prompt
  • flattening the JSON so that there’s no nested data
  • using clear/descriptive value labels
  • prompting it to be warned of these mistakes
  • lowering the temperature
2

Use gpt-3.5-turbo-16k if you have access, or gpt-3.5-turbo-1106. You will find these inexpensive models to be of much higher quality in performing tasks on data.

2 Likes
3

In general, the easiest way to improve performance is to pick a better model :confused:

1 Like
4

Interesting. Will give that a try!