waken-azure-openai

An Azure OpenAI Target adapter for Waken — "nginx for AI agents." Routes waken Events to an Azure-hosted OpenAI deployment's Chat Completions API and returns the result as a waken Response.

This is a separate package from waken-openai, not a thin wrapper around it, because Azure OpenAI has a genuinely different auth/configuration shape — an endpoint URL, a deployment name, and a dated API version, instead of just an API key — and (as of this writing) only documents/supports this configuration shape for the Chat Completions API, not the Responses API waken-openai uses. See "Design notes" below.

Install

pip install waken-azure-openai

Usage

from waken import Runtime
from waken_azure_openai import AzureOpenAIAdapter

runtime = Runtime()
runtime.target("azure-openai", AzureOpenAIAdapter(deployment="my-gpt4-deployment"))
runtime.run()

Three things need to be set, either as constructor arguments or as environment variables:

deployment is the name you gave the model deployment in the Azure Portal — not an OpenAI model id. This is a distinct and commonly-confused concept worth spelling out: on Azure, you don't call a model by a name like "gpt-4.1" directly. You first create a deployment of some underlying model inside your Azure OpenAI resource and give that deployment its own name (e.g. "my-gpt4-deployment"), and every API call addresses that deployment name — Azure looks up which model actually serves it from the deployment's own configuration. That's why this adapter's constructor parameter is named deployment, not model: passing an OpenAI model id here (AzureOpenAIAdapter(deployment="gpt-4.1")) will fail unless you happen to have named your deployment exactly that.

Any other keyword the openai SDK's AsyncAzureOpenAI client accepts is forwarded straight through, e.g. azure_ad_token_provider=... to authenticate via Microsoft Entra ID instead of a static API key:

from azure.identity import DefaultAzureCredential, get_bearer_token_provider

runtime.target(
    "azure-openai",
    AzureOpenAIAdapter(
        deployment="my-gpt4-deployment",
        azure_ad_token_provider=get_bearer_token_provider(
            DefaultAzureCredential(), "https://cognitiveservices.azure.com/.default"
        ),
    ),
)

(get_bearer_token_provider returns a plain sync callable; AsyncAzureOpenAI accepts either a sync or async callable for azure_ad_token_provider and awaits it only if it returns an awaitable — confirmed from the installed SDK's _get_azure_ad_token, not assumed.)

Sessions

Like waken-mistral and waken-cohere (and unlike waken-openai or Claude, which have a server-side resumable session id), the Chat Completions API this adapter calls is stateless — it expects the full conversation history resent on every call. So AzureOpenAIAdapter maps a waken session_id to a plain, growing list of {"role": ..., "content": ...} messages, appends the new user prompt and the assistant's reply to that list on every turn, and resends the whole thing each time. History is kept client-side, in this process's memory only — there's no persistence layer, and it's lost on restart. An Event with no session_id sends a fresh one-message list each call and is never stored.

Design notes

Building this adapter meant checking the real openai SDK and Azure's own current docs rather than assuming parity with waken-openai — worth recording here since it drove the design:

• The openai package (same PyPI package as vanilla OpenAI) ships an AsyncAzureOpenAI client alongside AsyncOpenAI. Its constructor already falls back to AZURE_OPENAI_ENDPOINT / AZURE_OPENAI_API_KEY when azure_endpoint/api_key aren't passed — this adapter forwards None straight through rather than re-implementing that lookup.
• api_version is the exception: the SDK's own automatic fallback reads OPENAI_API_VERSION, not AZURE_OPENAI_API_VERSION — despite the latter being what nearly every Azure OpenAI tutorial, the Portal's own code samples, LangChain, and Semantic Kernel all use. This adapter resolves AZURE_OPENAI_API_VERSION itself and always passes an explicit api_version= to the client, so that the more common env var name actually works.
• Azure's currently-documented Python pattern for the Responses API (client.responses.create(...)) uses a plain OpenAI/AsyncOpenAI client pointed at a .../openai/v1/ base URL — not AzureOpenAI/AsyncAzureOpenAI with a dated api_version. Confirmed by reading the installed SDK's openai/lib/azure.py (its URL-rewriting only special-cases /chat/completions and similar endpoints for deployment-scoped routing, not /responses) and by cross-checking Microsoft's own current docs, where every responses.create(...) example uses the .../openai/v1/ + plain-client shape and every AzureOpenAI(azure_endpoint=..., api_version=...) example calls chat.completions.create(...). Since this task's whole reason for existing is the endpoint+deployment+api-version auth shape, this adapter is built on Chat Completions rather than forcing the Responses API's session-id pattern through a client shape that doesn't actually route it correctly on Azure.
• 2024-10-21 is used as the default api_version because it's Azure OpenAI's latest generally-available (non-preview) API release as of this writing.

Development

git clone https://github.com/WakenHQ/waken-azure-openai
cd waken-azure-openai
pip install -e ".[dev]"
pytest

Tests mock openai.AsyncAzureOpenAI entirely — no API key, endpoint, or network access needed to run the suite.

License

MIT