Python: New Feature: Support Foundry SDK in Python

[johnsinco]

---

Support Foundry SDK in Python

---

The recently released Azure AI Foundry Python SDK supports retrieving Azure OpenAI connections from your Foundry project. This improves project security as you are not dealing with the AOAI endpoints and secret keys. I have not been able to find an integration point where I can initialize the SK-Python Kernel object in such a way to utilize the Foundry initialized AOAI instance.

Proposed Change

Allow the AOAI instance openai.AzureOpenAI to be injected similar to the below...

from azure.identity import DefaultAzureCredential
from azure.ai.projects import AIProjectClient
from semantic_kernel import Kernel
import openai.AzureOpenAI

project_connection_string="<<AI Foundry connection string>>"

project = AIProjectClient.from_connection_string(conn_str=project_connection_string, credential=DefaultAzureCredential())  # or Managed Identity, etc.

# retrieve the AOAI instance from the Foundry project
injected_aoai_client = project.inference.get_azure_openai_client(api_version="2024-06-01")  # this is an openai.AzureOpenAI object

# Create the Kernel object and add the existing AOAI service retrieved from the Foundry Project
kernel = Kernel(services={"azure_openai": injected_aoai_clieint})

[moonbox3]

Hi @johnsinco, thanks for opening this item. I was looking for the code for the get_azure_openai_client method and it returns an AsyncAzureOpenAI client:

async def get_azure_openai_client(self, *, api_version: Optional[str] = None, **kwargs) -> "AsyncAzureOpenAI":
    ...

You could inject this custom client into our AzureChatCompletion class like:

chat = AzureChatCompletion(
    async_client=<your_custom_foundry_client>
)

[alliscode]

@johnsinco This issue appears to be resolved so I'm closing it for now. Please reopen if needed. Thanks!

[johnsinco]

Thanks @moonbox3 for the quick reply. I was able to adapt your hints above and plug AI Foundry to the SK kernel like so...

import asyncio
  # from azure.ai.projects import AIProjectClient
from azure.ai.projects.aio import AIProjectClient  # <<-- identically named async version of the client
from azure.identity import DefaultAzureCredential
from semantic_kernel import Kernel
from semantic_kernel.connectors.ai.open_ai import AzureChatCompletion

# Retrieve the connection string from environment variables
project_connection_string = os.getenv("PROJECT_CONNECTION_STRING")

project = AIProjectClient.from_connection_string(
  conn_str=project_connection_string,
  credential=DefaultAzureCredential())

aoai_chat = AzureChatCompletion(
    deployment_name="gpt-4o",
    async_client=aoai_client
)

# Create the Kernel object and add the Azure OpenAI Chat Completion service
kernel = Kernel()
service_id="agent"
kernel.add_service(AzureChatCompletion(service_id=service_id, async_client=aoai_client, deployment_name="gpt-4o"))