Here's my theory - sorry if it's a little long-winded. As a starting point, the AzureOpenAI vectorizer config interface is different from the OpenAI vectorizer config interface. The Azure class has no model while the OpenAI class does have model, modelVersion, type_, and dimensions. This is what's in the Python client:

class _Text2VecAzureOpenAIConfig(_VectorizerConfigCreate):
    vectorizer: Union[Vectorizers, _EnumLikeStr] = Field(
        default=Vectorizers.TEXT2VEC_OPENAI, frozen=True, exclude=True
    )
    baseURL: Optional[AnyHttpUrl]
    resourceName: str
    deploymentId: str
    vectorizeClassName: bool

class _Text2VecOpenAIConfig(_VectorizerConfigCreate):
    vectorizer: Union[Vectorizers, _EnumLikeStr] = Field(
        default=Vectorizers.TEXT2VEC_OPENAI, frozen=True, exclude=True
    )
    baseURL: Optional[AnyHttpUrl]
    dimensions: Optional[int]
    model: Optional[str]
    modelVersion: Optional[str]
    type_: Optional[OpenAIType]
    vectorizeClassName: bool

In the vectorizer module, a default model is assigned, which is a v3 model, and doesn't take into account any of the deploymentId information from the Azure interface:

The Dimensions is then determined based on the model:

And this dimension ultimately winds up getting passed unconditionally in the request body:

I set up a dev weaviate/weaviate-python-client environment and I'm able to work around the issue client-side by expanding _Text2VecAzureOpenAIConfig to include model, modelVersion, type_ and dimensions. The workaround looks like:

    vectorizer_config=wvc.Configure.Vectorizer.text2vec_azure_openai(
        resource_name="myCompanyResource",
        deployment_id="myCompany-text-embedding-ada-002",
        model="ada",
        model_version="002",
        type_="text"
    )

I can submit a PR in the client repos for this change, but from a UX perspective I think the validations on the db should be updated as well, like:

• If Azure and no model was provided, attempt to use the deploymentId as the model (when deploying a model in Azure, it pre-fills the deployment name with the OpenAI model name, but allows the user to rename it. So I think it's safe to assume there are many deploymentIds which match the model name)
• If Azure - don't use a default model. Require a user-specified model (or deploymentId matching an accepted model) for everything derived from it, e.g. if dimensions is specified, require that the model was provided by the user