Added Azure AI Chat Completion Client

[rohanthacker]

Related issue number

#4683 Adds initial support for Azure AI Chat Completion Client

Checks

• I've included any doc changes needed for https://microsoft.github.io/autogen/. See https://microsoft.github.io/autogen/docs/Contribute#documentation to build and test documentation locally.
• I've added tests (if relevant) corresponding to the changes introduced in this PR.
• I've made sure all auto checks have passed.

[ekzhu]

@yanivvak can you review this?

[yanivvak]

@ekzhu @rohanthacker
Great work, I tried to deploy with 3 different options offered by Azure AI inference SDK

1. Azure open AI - works good, I tried it with magnetic one, the websurfer got stucked, can you take a look?
2. Serverless - it works, but I didn't got the full answer it was phi 3.5
3. Managed compute - it didn't run for me, I assume it's an issue with the endpoint and it is not related to your code

[yanivvak]

I used this code
https://github.com/Azure-Samples/dream-team/blob/main/Magenticone_example.py

[ekzhu]

@lspinheiro could you help reviewing this PR?

[lspinheiro]

I think this part could benefit from some better separation of concerns between config validation and instantiation. e.g.

class AzureAIChatCompletionClient(ChatCompletionClient):
    def __init__(self, **kwargs: Unpack[AzureAIClientConfiguration]):
        config = self._validate_config(kwargs)
        self._client = self._create_client(config)
        self._create_args = self._prepare_create_args(config)
        # ...

    @staticmethod
    def _validate_config(config: Mapping[str, Any]) -> AzureAIClientConfiguration:
        # Validation logic here
        return config

[lspinheiro]

@lspinheiro could you help reviewing this PR?

Looks quite good and is consistent with the openai client. I have a minor comment about the config validation. @jackgerrits may have more options since a lot of the design decisions here are driven by his original implementation of the openai client. If anything doesn't make since in this context he would be the best person to evaluate.

[ekzhu]

Let's use the latest update to canonicalize the finish_reason. See OpenAIChatCompletionClient

[ekzhu]

@rohanthacker Is this PR ready to review? It looks good from code perspective. @srjoglekar246 could you help to use it in an assistant agent and running some teams on it to test it out?

[ekzhu]

Model Capabilities are deprecated, use ModelInfo. See OpenAIChatCompletionClient: https://github.com/microsoft/autogen/blob/main/python/packages/autogen-ext/src/autogen_ext/models/openai/_openai_client.py#L912-L913

[ekzhu]

Use model_info as model_capabilities is deprecated.

[ekzhu]

@rohanthacker I made this PR ready for review.

[lspinheiro]

@ekzhu For context, this works for Azure AI Inference. I tested on a Phi-4 deployment I created.

from semantic_kernel import Kernel
from semantic_kernel.memory.null_memory import NullMemory
from semantic_kernel.connectors.ai.azure_ai_inference import AzureAIInferenceChatCompletion
from semantic_kernel.connectors.ai.azure_ai_inference import AzureAIInferenceChatPromptExecutionSettings
from autogen_core.models import SystemMessage, UserMessage, LLMMessage
from autogen_ext.models.semantic_kernel import SKChatCompletionAdapter


kernel = Kernel(memory=NullMemory())

execution_settings = AzureAIInferenceChatPromptExecutionSettings(
    max_tokens=100,
    temperature=0.5,
    top_p=0.9,
)

chat_completion_service = AzureAIInferenceChatCompletion(ai_model_id="Phi-4")
model_adapter = SKChatCompletionAdapter(sk_client=chat_completion_service)

messages: list[LLMMessage] = [
    SystemMessage(content="You are a helpful assistant."),
    UserMessage(content="What is 2 + 2?", source="user"),
]

azure_result = await model_adapter.create(
    messages=messages,
    extra_create_args={"kernel": kernel, "prompt_execution_settings": execution_settings},
)
print("Azure result:", azure_result.content)

[ekzhu]

This should be merged with the azure extra, because if you want to use this you need azure-identity anyway.