Performance drop for Gemini on Vertex AI

[domp5]

Hello,

I'm working on a structured extraction task and initially used Gemini via the Generative Language API with great results. However, after switching to Vertex AI for production, the output quality dropped significantly. I'm using gemini-2.0-flash and I tried different libraries such as python-genai, instructor and pydantic-ai but with none of these I was able to get comparable results.

Has someone experienced the same? Could this be due to configuration differences, or is there another potential cause?

Thanks!

[TomerH]

Yes. I'm observing the same issues. (Primarily with 2.0-flash).
(Just created a post about it).

I requested a quotas increase, but I doubt that's the issue.
I suspect the issue is with the service.