I’m building a health and wellness web app where during the onboarding process i take the user data and build a personal diet plan. to do that i’m using the assistant api with an attached json file of all the food and macronutrienets ( protein, carbs, fat…etc ) . My problem : The assistants fails when using gpt-3.5 1025 to retrieve data and food from the file ( see attached screen ) despite being extremly well organized. It works fine with gpt 4 but it’s too expensive considering my use case where i need a big json output.
There’s this in addition to the results being different when hittin the assistant via api vs the playground. the playground gives much more consistent results.
Did anyone experience something similar or is it me ?
You must enable only retrieval before uploading, and then watch very carefully in the user interface to see the error (that would be much clearer when using API calls).
For example, I have a text file, with extension .txt, that I upload. It has some Discourse forum source code I want to ask about.
The Assistants endpoint does unwanted inspection of the contents and rejects it. Regardless if I stick another paragraph or two of plain text at the start of the file. Proving reason #156 why the platform should just be trashed.
“hontou” in Japanese means “true”, so we guess that the error is some weeabo juvenile file extractor message.
