Skip to content

A reactive notebook for Python — run reproducible experiments, query with SQL, execute as a script, deploy as an app, and version with git. Stored as pure Python. All in a modern, AI-native editor.

add docs for azure and properly handle endpoint (#6587)

Shahmir Varqhaaffaa95cfaae
diff --git a/docs/guides/configuration/llm_providers.md b/docs/guides/configuration/llm_providers.md
index 34f849b7f5..581f31b723 100644
--- a/docs/guides/configuration/llm_providers.md
+++ b/docs/guides/configuration/llm_providers.md
@@ -179,6 +179,36 @@ export GOOGLE_CLOUD_LOCATION='us-central1'
For details and advanced configuration, see the `google-genai` Python client docs: `https://googleapis.github.io/python-genai/#create-a-client`.
+### Azure
+
+There are two offerings for serving LLMs on Azure
+
+**Azure OpenAI**
+
+```toml title="marimo.toml"
+[ai.models]
+chat_model = "azure/gpt-4.1-mini"
+
+[ai.azure]
+api_key = "sk-proj-..."
+base_url = "https://<your-resource-name>.openai.azure.com/openai/deployments/<deployment_name>?api-version=<api-version>"
+```
+
+The deployment name is typically the model name.
+
+**Azure AI Foundry**
+
+AI Foundry uses OpenAI-compatible models, so you can use the same configuration as OpenAI-compatible providers.
+
+```toml title="marimo.toml"
+[ai.models]
+custom_models = ["custom-azure/mistral-medium"]
+
+[ai.open_ai_compatible]
+api_key = "sk-proj-..."
+base_url = "https://<your-resource-name>.services.ai.azure.com/openai/v1"
+```
+
### GitHub Copilot
Use Copilot for code refactoring or the chat panel (Copilot subscription required).
diff --git a/frontend/src/components/app-config/ai-config.tsx b/frontend/src/components/app-config/ai-config.tsx
index 3017cabbae..296967c8b6 100644
--- a/frontend/src/components/app-config/ai-config.tsx
+++ b/frontend/src/components/app-config/ai-config.tsx
@@ -791,7 +791,8 @@ export const AiProvidersConfig: React.FC<AiConfigProps> = ({
form={form}
config={config}
name="ai.azure.base_url"
- placeholder="https://<your-resource-name>.openai.azure.com"
+ placeholder="https://<your-resource-name>.openai.azure.com/openai/deployments/<deployment-name>?api-version=<api-version>"
+ defaultValue="https://<your-resource-name>.openai.azure.com/openai/deployments/<deployment-name>?api-version=<api-version>"
testId="ai-azure-base-url-input"
/>
</AccordionFormItem>
diff --git a/marimo/_server/ai/providers.py b/marimo/_server/ai/providers.py
index a7d3d57a61..934b1820b3 100644
--- a/marimo/_server/ai/providers.py
+++ b/marimo/_server/ai/providers.py
@@ -17,6 +17,7 @@ from typing import (
Union,
cast,
)
+from urllib.parse import parse_qs, urlparse
from starlette.exceptions import HTTPException
@@ -579,24 +580,68 @@ class OpenAIProvider(
class AzureOpenAIProvider(OpenAIProvider):
- def get_client(self, config: AnyProviderConfig) -> AsyncOpenAI:
- from urllib.parse import parse_qs, urlparse
+ def _is_reasoning_model(self, model: str) -> bool:
+ # https://learn.microsoft.com/en-us/answers/questions/5519548/does-gpt-5-via-azure-support-reasoning-effort-and
+ # Only custom models support reasoning effort, we can expose this as a parameter in the future
+ del model
+ return False
+
+ def _handle_azure_openai(self, base_url: str) -> tuple[str, str, str]:
+ """Handle Azure OpenAI.
+ Sample base URL: https://<your-resource-name>.openai.azure.com/openai/deployments/<deployment_name>?api-version=<api-version>
+ Args:
+ base_url (str): The base URL of the Azure OpenAI.
+
+ Returns:
+ tuple[str, str, str]: The API version, deployment name, and endpoint.
+ """
+
+ parsed_url = urlparse(base_url)
+
+ deployment_name = parsed_url.path.split("/")[3]
+ api_version = parse_qs(parsed_url.query)["api-version"][0]
+
+ endpoint = f"{parsed_url.scheme}://{parsed_url.hostname}"
+ return api_version, deployment_name, endpoint
+
+ def get_client(self, config: AnyProviderConfig) -> AsyncOpenAI:
from openai import AsyncAzureOpenAI
base_url = config.base_url or None
key = config.api_key
- # Azure OpenAI clients are instantiated slightly differently
- parsed_url = urlparse(base_url)
- deployment_model = cast(str, parsed_url.path).split("/")[3]
- api_version = parse_qs(cast(str, parsed_url.query))["api-version"][0]
+ if base_url is None:
+ raise HTTPException(
+ status_code=HTTPStatus.BAD_REQUEST,
+ detail="Base URL needed to get the endpoint",
+ )
+
+ api_version = None
+ deployment_name = None
+ endpoint = None
+
+ if base_url:
+ if "services.ai.azure.com" in base_url:
+ raise HTTPException(
+ status_code=HTTPStatus.BAD_REQUEST,
+ detail="To use Azure AI Foundry, use the OpenAI-compatible provider instead.",
+ )
+ elif "openai.azure.com" in base_url:
+ api_version, deployment_name, endpoint = (
+ self._handle_azure_openai(base_url)
+ )
+ else:
+ LOGGER.warning(f"Unknown Azure OpenAI base URL: {base_url}")
+ api_version, deployment_name, endpoint = (
+ self._handle_azure_openai(base_url)
+ )
return AsyncAzureOpenAI(
api_key=key,
api_version=api_version,
- azure_deployment=deployment_model,
- azure_endpoint=f"{cast(str, parsed_url.scheme)}://{cast(str, parsed_url.hostname)}",
+ azure_deployment=deployment_name,
+ azure_endpoint=endpoint or "",
)
diff --git a/tests/_server/ai/test_providers.py b/tests/_server/ai/test_providers.py
index 1109f5eb9e..1154a4ebf9 100644
--- a/tests/_server/ai/test_providers.py
+++ b/tests/_server/ai/test_providers.py
@@ -9,6 +9,7 @@ from marimo._config.config import AiConfig
from marimo._server.ai.providers import (
AnthropicProvider,
AnyProviderConfig,
+ AzureOpenAIProvider,
BedrockProvider,
GoogleProvider,
OpenAIProvider,
@@ -148,3 +149,26 @@ async def test_openai_provider_max_tokens_parameter(
assert "max_completion_tokens" not in call_kwargs, (
"max_completion_tokens should not be present for non-reasoning models"
)
+
+
+async def test_azure_openai_provider() -> None:
+ """Test that Azure OpenAI provider uses correct parameters."""
+ config = AnyProviderConfig(
+ api_key="test-key",
+ base_url="https://test.openai.azure.com/openai/deployments/gpt-4-1?api-version=2023-05-15",
+ )
+ provider = AzureOpenAIProvider("gpt-4", config)
+
+ api_version, deployment_name, endpoint = provider._handle_azure_openai(
+ "https://test.openai.azure.com/openai/deployments/gpt-4-1?api-version=2023-05-15"
+ )
+ assert api_version == "2023-05-15"
+ assert deployment_name == "gpt-4-1"
+ assert endpoint == "https://test.openai.azure.com"
+
+ api_version, deployment_name, endpoint = provider._handle_azure_openai(
+ "https://unknown_domain.openai/openai/deployments/gpt-4-1?api-version=2023-05-15"
+ )
+ assert api_version == "2023-05-15"
+ assert deployment_name == "gpt-4-1"
+ assert endpoint == "https://unknown_domain.openai"