google-vertex provider now uses google-genai but fails due to no default location

[dhimmel]

Here's my code:

from pydantic_ai import Agent

def test_pydantic_ai_agent_sync() -> None:
    agent = Agent(
        model="google-vertex:gemini-2.0-flash",
    )
    result = agent.run_sync(
        user_prompt="This is a test. Respond with only the word 'okay'."
    )
    assert result.output.strip() == "okay"

And I'm getting the error:

ImportError: Please install `google-genai` to use the Google model, you can use the `google` optional group — `pip install "pydantic-ai-slim[google]"

My environment specifies pydantic-ai-slim[vertexai].

The expected behavior here is that the google-vertex provider would be used and the google-genai package would not be needed.

I am very confused because the above code used to work, and I'm not really sure what changed... perhaps an update or some additional configuration or environment setting got set on my system?

I am seeing this in the stacktrace:

.venv/lib/python3.13/site-packages/pydantic_ai/agent.py:289: in __init__
    self.model = models.infer_model(model)
                 ^^^^^^^^^^^^^^^^^^^^^^^^^
.venv/lib/python3.13/site-packages/pydantic_ai/models/__init__.py:564: in infer_model
    from .google import GoogleModel

[Kludex]

I sneakily replaced the model used underneath. Instead of using GeminiModel, I've changed to GoogleModel - there are some models around it.

What I did forgot (and that's why this issue exists) is that vertexai extra doesn't install google-genai. If you replace pydantic-ai-slim[vertexai] by pydantic-ai-slim[google] it should work again.

Just to be clear, this is a mistake on our side - I thought it wouldn't be a breaking change (which is not for users that are installing pydantic-ai, but since you are installing pydantic-ai-slim with extras, then it is).

I don't think it makes sense for us to revert given that we intend to remove the GeminiModel, but I see that we need to do it faster to avoid confusion.

[dhimmel]

replaced the model used underneath

Do you have the commit that made the change? I'd like to see what release it occurs in.

Instead of using GeminiModel, I've changed to GoogleModel

Okay so using model="google-vertex:gemini-2.0-flash" will construct a GoogleModel, but I can still explicitly use GeminiModel from its constructor rather than by string name?

The thing is that the former GeminiModel authentication / configuration just worked but GoogleModel gives me:

ValueError: Project and location or API key must be set when using the Vertex AI API.

[Kludex]

• This is the PR that implemented the model: Add Google GenAI provider #1751
• This is the PR that set the GoogleModel as default: Use GoogleModel instead of GeminiModel on inference #1881

The thing is that the former GeminiModel authentication / configuration just worked but GoogleModel gives me:

So after installing the pydantic-ai-slim[google] you are having this issue? Can you give me the code that you have?

[dhimmel]

This is the PR that set the GoogleModel as default

Got it. 67c381e first released in v0.2.18.

So after installing the pydantic-ai-slim[google] you are having this issue? Can you give me the code that you have?

Yes. Here is what I'm doing.

Update pyproject dependencies to include:

    "pydantic-ai-slim[google]>=0.2.17",

uv locks this to:

[[package]]
name = "pydantic-ai-slim"
version = "0.3.1"

I've run gcloud auth application-default login to setup Application Default Credentials. Then I run the test_pydantic_ai_agent_sync above and get the Project and location or API key must be set error.

If I run gcloud config list, I see I do have a default project set, but no default location.

If I prefix my pytest command with GOOGLE_CLOUD_LOCATION=us-east4 to set that environment variable the test now passes.

So can we conclude that GoogleModel requires location to be set while GeminiModel was using a default?

[dhimmel]

From this page about which vertex gen ai region to pick

• global is highest availability but excludes some features
• us-central1 in Iowa has the most model support. This is what my pydantic-ai requests used by default when GeminiModel was in use.

I liked the ability to use str model specifiers for agents. It seems like the google genai package is requiring explicit region specification. I don't know if pydantic-ai can set a reasonable default. It seems odd to set a GOOGLE_CLOUD_LOCATION environment variable which is scoped beyond just vertex ai.

[Kludex]

I think we can match the behavior.

• I assume this solves the issue? Set 'us-central1' by default on GoogleProvider #2031