Support Azure OpenAI in transformer agents

[huajianmao]

Feature request

Support it by setting the openai properties.

openai.api_type = "azure"
openai.api_version = "2023-03-15-preview"
openai.api_base = "https://THE_NAME.openai.azure.com/"
openai.api_key = "AZURE_OPENAI_API_KEY"

# response = openai.Completion.create(engine="xxxx", ...)

Motivation

Azure OpenAI is also used by a lot of people.
It would be great to support it in the transformer agents.

Your contribution

I may contribute to test it.

[amyeroberts]

@huajianmao thanks for opening this issue,

As a user can enable azure open AI support by adding these lines to their own code, there isn't a requirement to add support in the library.

[waundme]

@huajianmao thanks for opening this issue,

As a user can enable azure open AI support by adding these lines to their own code, there isn't a requirement to add support in the library.

That's not entirely correct. You'll get an InvalidRequestError: Must provide an 'engine' or 'deployment_id' parameter to create a <class
'openai.api_resources.completion.Completion'>

Because you would have to support some way of specifying the Azure-deployment_id in your code.

[MSDNAndi]

https://huggingface.co/docs/transformers/v4.29.1/en/main_classes/agent#transformers.OpenAiAgent
agent = OpenAiAgent(model="text-davinci-003", api_key=pswd)
ideally would be changed. The Azure API is a different type and different enough to allow the relevant parameters to be passed in.
It doesn't sound "clean" to me to access the openai API parameters directly outside the OpenAiAgent.
The Azure based version is rapidly gaining in importance as people can use their corporate or personal Azure credits vs having to pay OpenAI separately.

Also the call needs to be changed to something like that currently:
result = openai.Completion.create(
#model=self.model,
deployment_id=self.model,
prompt=prompts,
temperature=0,
stop=stop,
max_tokens=200,
)

btw. something to easily stumble over that the version is api_version="2022-12-01" even though the portal shows "1". Best is to go to the playground and have it generate example code currently it seems.

[cforce]

for GPT4 you need to use
openai.api_version = "2023-03-15-preview" # used for GPT-4 - see https://learn.microsoft.com/en-gb/azure/cognitive-services/openai/how-to/chatgpt?pivots=programming-language-chat-completions

Be aware the chat message pattern changed nd as well the json for the response, which is the new mode for all new models after GPT4 as well.

[MSDNAndi]

Sure, but we were not talking about GPT-4. The examples only covered DaVinci-003

[cforce]

Anyway the api is changed - no matter what model you use - you need to adopt to the new prompt style.