Assistant version forced to v2, no downgrade options

API
1

As mentioned in a previous thread, we are seeing that our assistants automatically flipped to V2, increasing our costs 10x. We tried using the “use v1” button" in the dashboard but it automatically flips back.

After reaching out to OpenAI support, they have informed us that it is “by design” and there is NO WAY to downgrade to v1 prior to the deprecation in 2024.

From OpenAI support:

Regarding the inability to downgrade to v1 from the dashboard and the issue with the Python SDK automatically upgrading to v2, it appears that these behaviors are by design, as the Assistants API v2 is intended to replace v1. The migration guide provides detailed instructions on how to transition from v1 to v2, but it does not include an option to revert back to v1 due to the planned deprecation of v1 by the end of 2024. Given the situation, here are a few steps you can consider to manage your costs while using Assistants API v2…

I understand that this situation is not ideal, and I appreciate your patience as we navigate these changes together. If you have specific questions about optimizing your usage or need further assistance, please don’t hesitate to reach out.

So there’s a planned deprecation EOY 2024, but we will automatically switch you over now, and there’s no way to go back? How does this make sense!

Why do you include a button in your dashboard UI, if it does absolutely nothing and contradicts your actual implementation??

We are not using any file search or retrieval. According to OpenAI, the defaults should be the same. Not true.

Any thoughts on this? Absolutely scary that OpenAI decided to make a hard-upgrade like this.

UPDATE: I received a response on a separate thread that the response from OpenAI was an auto-response. However, we’re still trying to understand why we cannot switch the version declared in the dashboard (we are using python SDK), as it certainly appears to be using V2.

2

“Your assistants” are not flipped. They have no such setting.

The playground is a demonstration, a place to experiment. It also doesn’t show logprobs for chat, and we are not complaining that we are denied access, because we are sophisticated when we employ that feature.

Using actual API calls are how a “we” would interact with the API. Interact by sending either the v1 or v2 header.

1 Like
3

@_j We are using the assistant created in the Assistant dashboard via API - these assistants are fully featured, configured via dashboard, and usable programmatically via the API (from our experience). Said another way - we created the assistant in the Assistants dashboard, instead of 1-off creating the assistant via API as well (how is this different?) In fact, any assistants we create via API show in the Assistants dashboad.

So are these assistant still considered “only playground” when we use these legitimate assistant objects via the API?

It sounds like you may be confusing working in the Playground chat as opposed to configuring legitimate Assistants in the OpenAI dashboard. That is NOT the Playground.

In other words, we aren’t interacting with it via the Playground assistants, we are using the python SDK in a service.

I understand what you mean about not being flipped - but considering the change was made on the assistant object (automatically) and we are using python SDK (which OpenAI states uses the version automatically).

Should we be passing a custom header via the python SDK then?

4

A link that answers your question about still using v1 and retrieval with assistants via Python:

https://platform.openai.com/docs/assistants/migration/with-sdks

The removal of “retrieval” from the remaining v1 documentation (and that you were never charged and there was never any mention of the internal workings of arbitrary file dump and a myfiles_browser tool) should impress that a move is wise.

2 Likes
5

Thanks for the direct link; it helps remove some confusion generated by the initial support response. Not sure why our service still seems to be calling the v2 namespace on the configured assistant, but will update if I find anything.