What's my best option (API) Normal promting or assistant or something else I don't know about?

API
1

Hi everyone,

We are looking into creating childrens books using OpenAI.
We have quit some rules for the output. Things like:

  • 900 - 1200 words
  • 12 pages
  • If parents are in the story only use the first name once, after that refer to daddy and mummy
  • Characters should be consistent in tone for the full story
  • Output should be done in specific json scheme
  • Create a promt per page for us to create illustrations
    • Make sure only 1 character per promt for illustration
    • etc etc etc

This list will be way longer to get every output perfect.
At first I trained a custom GPT. This was quit successful but unfortunately we canā€™t run the custom GPTā€™s via API.

So I checked assistance since this would be the API version of custom GPTā€™s. I have limited succes with this option. ā€œtrainingā€ the assistant is quit minimal.
So again we went back to just prompting all our ā€œrulesā€ and than create a story. But we notice that some of the rules are not applied possible due to the fact that the promt is extremely long.

So my question to you guys, how would you tackle this issue? Normal API promting, multiple promts following each other without human interaction, assistant(s), other options we didnā€™t check yet?
Knowing we need to have the model adhere to our rules + we have to use a api call to sent the promt since it will be automated.

Thanks!

2 Likes
2

I believe there may be several ways to perform the training, but I donā€™t have the knowledge to answer precisely yet. However, I am working on developing a assistant where I selected the ā€˜File Searchā€™ option and reduced the temperature to close to zero, favoring file searches. I structured the content used by the assistant using JSON format files. Text format files allow the inclusion of various information with a small size. You can upload several files to the assistant. I will have new results soon.

I hope I have contributed, best regards