Azure OpenAI¶

The Azure provider runs OpenAI's models from an Azure OpenAI resource. Use it when your organisation buys OpenAI capacity through Azure: quota, networking and billing live there, and requests must go to your resource, not api.openai.com.

Azure serves the same models, but not at the same address and not with the same auth, which is why it is its own provider rather than a row in the OpenAI-compatible table. Three things differ:

The URL names a deployment, not a model. Whoever created the deployment chose its name, so gpt4o-prod is an ordinary name for gpt-4o.
Auth is an api-key header, or an Entra ID token, not Authorization: Bearer.
Every request carries an api-version. Azure has no rolling "latest".

Under the hood this is the OpenAI provider with Azure's client behind it, so message conversion, tool calls, streaming and structured output behave exactly as they do there.

Getting started¶

1. Point at your resource¶

export AZURE_OPENAI_ENDPOINT=https://my-resource.openai.azure.com
export AZURE_OPENAI_API_KEY=...

It uses the same openai extra as the OpenAI provider; there is no new dependency.

2. Address a deployment¶

from tulip.agent import Agent

agent = Agent(model="azure:gpt4o-prod", system_prompt="You answer questions about orders.")

azure: selects the provider; gpt4o-prod is the deployment name from your resource, not the underlying model id.

3. Or configure it explicitly¶

from tulip.agent import Agent
from tulip.models.native.azure import AzureOpenAIModel

model = AzureOpenAIModel(
    model="gpt4o-prod",
    endpoint="https://my-resource.openai.azure.com",
    api_version="2024-10-21",
)
agent = Agent(model=model)

Configuration¶

Setting	Argument	Environment	Default
Deployment name	`model`	—	required
Resource endpoint	`endpoint`	`AZURE_OPENAI_ENDPOINT`	required
API key	`api_key`	`AZURE_OPENAI_API_KEY`	one of key or token required
Entra ID token	`azure_ad_token`	—	used instead of a key
API version	`api_version`	`AZURE_OPENAI_API_VERSION`	`2024-10-21`

An argument wins over the environment. 2024-10-21 is a GA version with tool calling and streaming; set a newer one when a deployment needs a newer feature.

Common gotchas¶

Symptom	Likely cause
`Azure OpenAI needs a resource endpoint`	Neither `endpoint=` nor `AZURE_OPENAI_ENDPOINT` is set.
`No credentials for Azure OpenAI`	Neither an API key nor `azure_ad_token` is set.
404 `DeploymentNotFound`	`model` is the model id (`gpt-4o`) instead of the deployment name.
A parameter is rejected as unsupported	The deployment needs a newer `api_version`.

Source¶

tulip/models/native/azure.py