feat: consolidate azure with openai provider

[codefromthecrypt]

note: in order to be able to run just integration -k azure and not also run openai and ollama, I needed to get rid of the openai subdir in the tests (as it was before I started)

[codefromthecrypt]

draft until #54

[codefromthecrypt]

hard to see here, but it is the same openai endpoint as normal openai, but only at chat/completions. This is why the openai python SDK needs little config to switch to azure

[codefromthecrypt]

@michaelneale this is a scrubbed real request/response

[codefromthecrypt]

@anuraaga this has two commits, but if any interest in the azurification

[anuraaga]

Just some nits, LGTM

[anuraaga]

Using the presence of openai in the URI to indicate azure is still blowing my mind, especially so since it's not obvious the uri doesn't contain the host which could otherwise have been api.openai.com.

Perhaps request.headers["host"] == os.environ.get("AZURE_CHAT_COMPLETIONS_HOST_NAME") is an option? Admittedly not great either, just a note to see if there isn't anything to improve here

[codefromthecrypt]

this was a goof on my side, I meant to put openai.azure.com which I think we can keep as a condition . The problem with using that ENV variable is that it is weird, like it isn't a hostname, rather a base URL.

[codefromthecrypt]

currently at least the azure deployment I'm using doesn't support vision even though the model seems the same as openai. For now, not adding the vision tests, but someone can try again in a subsequent PR perhaps with another deployment:

E               {
E                 "error": {
E                   "message": "Invalid content type. image_url is only supported by certain models.",
E                   "type": "invalid_request_error",
E                   "param": "messages.[4].content.[0].type",
E                   "code": null
E                 }
E               }

[codefromthecrypt]

@anuraaga can you verify I am understanding right, basically that we can't use vision yet?

https://learn.microsoft.com/en-us/azure/ai-services/openai/how-to/gpt-with-vision?tabs=rest#gpt-4-turbo-provisioned-managed-availability

I've tried gpt-4 and no dice, but above docs make sense

e.g. below fail before getting to assertions with invalid_request_error

@pytest.mark.vcr()
def test_azure_vision(default_azure_env):
    reply_message, reply_usage = vision(AzureProvider, AZURE_MODEL)

    assert reply_message.content == [Text(text='The first entry in the menu says "Ask Goose."')]
    assert reply_usage.total_tokens == 14241


@pytest.mark.integration
def test_azure_vision_integration():
    reply = vision(AzureProvider, AZURE_MODEL)

    assert "ask goose" in reply[0].text.lower()

[codefromthecrypt]

@baxen lemme know if you think this is more sensible than before

[baxen]

yeah this looks good!

[baxen]

Looks great!