Different answers from UI to API

Using: GPT-Turbo-4.0
I’m getting drastically different results when using the UI or assistants playground than using either the assistants API or .

I’m asking the API to summarize a few uploaded files. I use file search for the Assistant’s playground and API and standard upload for the standard UI.

Overall the quality is significantly lower through the API. I’ve played with the temperature and it hasn’t had any benefit.

Is there way I can ensure I’m using the exact settings as the UI?

I’m not familiar with that model :thinking:

these are the available models: https://platform.openai.com/docs/models/gpt-4-turbo-and-gpt-4

are you sure you’re using the same models with the same parameters?

sorry gpt-4-turbo instead of gpt-turbo-4.

I’m using the same model. I’ve placed the token limit to the max and tried temperature at both 0 and .7.

gpt-4-turbo also isn’t really a model. it’s a shorthand that gets gets rewritten to some other model in the backend.

also, how are you uploading these files? if you upload the files to be indexed, then the model is gonna have a hard time summarizing them. it’s like typing “summarize the whole internet for me pls” into google.

If you want more control over the process you can consider trying the chat completions endpoint, and just include your file as text, just so you get an idea how the whole thing actually works.

Hi, thanks so much for the help.

I’m uploading them using the v1 files endpoint(https://platform.openai.com/docs/api-reference/files/create). I am not using vector store.

Unfortunately they’re in pdfs and so would prefer to use openai’s capabilities to read them directly. But if it’s the best way I can pull them into text.

I just need to mimic however the upload file works on the UI or file search works in the assistant’s playground. THat level of performanceis great.