Seed parameter with gpt-4o

[Rvey]

Hello,

I would like to ask about the seed parameter with GPT-4o with vercel SDK. Is it supported? I am building an application to onboard users, and I need the model to provide the same experience consistently across all users.

[lgrammel]

https://sdk.vercel.ai/docs/reference/ai-sdk-core/stream-text#seed

[Rvey]

thanks for sharing , anyway i want to ask if you have any information is it really working (openai - side) because i try to set it and i don't really notice a deferent .

[gpoole]

An old question, but just to share my experience: I had the same problem so to be sure I checked the source and used the debugger to make sure seed is definitely being passed down and sent to OpenAI. Definitely will be supported for GPT-4o as it was first introduced back in 2023. The API docs don't name any supported models specifically though.

I think the problem is more likely to do with what seed actually does. First, as the docs mention, it isn't guaranteed to work. Second, to even have a shot at getting the same result, everything needs to be identical between requests including the exact same prompt, temperature etc. Because of the way LLMs and seeded random number generators work, even a small difference in the input, or even a slight difference when generating the output will cause a pretty different final result.

I think overall it takes a fair bit of care to set everything up right, and even then it may still not work every time.

Some useful discussions:
https://community.openai.com/t/the-seed-option-for-gpt-does-not-increase-the-determinism-level/512892
https://community.openai.com/t/seed-param-and-reproducible-output-do-not-work/487245/23