Multi-account configuration for app on OpenAI assistant

API
1

We’re building an app with an OpenAI assistant and the frontend in Bubble. My question is: If multiple account holders (user/clients) are using the same OpenAI assistant through the Bubble API, can we safely serve them all through one shared assistant without issues like context drift, memory confusion, or output bleed? Or does each account need its own dedicated assistant?

We’re handling user authentication and session data on our end, but we want to make sure the assistant won’t retain or cross wires between accounts.

I appreciate any insights and advice you can share. Thanks ahead of time!

2

Welcome to the dev forum, @Nao.

The conversation context for an assistant is retained in Threads.

As long as the thread_id remains unique per user, there won’t be any context bleed or leaked conversations.

Also, the Assistants API will be deprecated next year and replaced by the Responses API.

3

Thank you, that’s good to know. So if I understand correctly, we can already start building with the Responses API? Or do we have to wait until deprecation of the Assistants API starts?

4

You can start building with the Responses API today. Docs cover examples for both Responses API and Chat Completions API.

You’ll find the drop down switch in the top right of the docs pages.

A dark-themed sidebar menu lists topics related to text input and output, including model selection, prompt engineering, and few-shot learning. (Captioned by AI)
A dark-themed sidebar menu lists topics related to text input and output, including model selection, prompt engineering, and few-shot learning. (Captioned by AI)474×934 18.5 KB

5

Thank you. Can you tell me if Responses API supports threads yet, and if not, when that is likely to be?

6

If you need a server-side conversation state, what Responses has is “previous response ID” as a parameter. It then uses the previous call’s input and output as “chat history”.

There is no effective mechanism provided to limit the length to which a conversation and its cost can grow, up to a model’s maximum input which is now 1M tokens, and there is no way to anticipate if that single request might even exceed your organization tier rate limit.

A self-managed chat database for customers is best, sending the budgeted input on which a response should be based. When OpenAI decides to ban your account haphazardly and keep your prepaid credits, you don’t have catastrophic data loss and customer lawsuits, and can offer UI drop-downs to other AI providers to fulfill a request.