WIP: fix: AzureOpenAI authentication to support API Key auth

[younes-io]

Based on the notebook and the AzureOpenAI class code, authentication with API Key authentication is not supported.

The azure_deployment field is required to differentiate between the model and its Azure deployment, which is crucial when multiple deployments of the same model exist, or when the deployment is not named after the model..

from guidance import models, gen
import tiktoken

azureai_model = models.AzureOpenAI(
    model='gpt-3.5-turbo',
    azure_endpoint=azure_endpoint,
    azure_deployment='my_crazy_deployment_name',
    api_key=api_key,
)

[younes-io]

Required: to test using token_ad_provider (not yet)

[slundberg]

Thanks @younes-io ! This looks good to me, @riedgar-ms or @Harsha-Nori any feedback?

[riedgar-ms]

Isn't the deployment contained within the endpoint URI? I have always used URIs of the form:

https://< < Resource Name > >.openai.azure.com/openai/deployments/< < AOAI Deployment > >/chat/completions?api-version=2023-07-01-preview

Although I wouldn't be surprised if the information can be given in other ways.

[younes-io]

@riedgar-ms : what you share is the full URL, which is constructed by the underlying library (langchain, etc).
This https://< < Resource Name > >.openai.azure.com is the base_url or (in the latest version) the azure_endpoint

[riedgar-ms]

I wouldn't object to passing through all the arguments (I assume that the AzureOpenAI object can sort everything out itself), although providing the 'full URL' as the endpoint does seem to work.

[younes-io]

@riedgar-ms never tried that xD but that's a good thing :D

[slundberg]

Sounds like this is a nice to have add-on, so I will merge :) Thanks @younes-io !