Azure OpenAI

Status: experimental provider path with a standalone tutorial.

Azure OpenAI can emit standard GenAI OpenTelemetry attributes or use the generic callable wrapper. Haystack generator identity containing AzureOpenAI is recognized as OpenAI.

Run the tutorial

cd examples/cloud/azure
cp .env.example .env
uv sync
uv run python sdk_enriched.py

Configure:

AZURE_OPENAI_API_KEY=...
AZURE_OPENAI_ENDPOINT=https://YOUR-RESOURCE.openai.azure.com
AZURE_OPENAI_API_VERSION=2024-10-21
AZURE_OPENAI_DEPLOYMENT=YOUR-DEPLOYMENT
WITDEM_ENDPOINT=http://localhost:4318

The tutorial wraps one existing AzureOpenAI.chat.completions.create workload with generic.instrument(provider="azure.openai", model=deployment).

azure.openai, azure_openai, and azure normalize to the distinct azure_openai provider. Witdem does not reuse direct OpenAI rates. The bundled catalog includes selected Azure model IDs; custom deployment names and negotiated or region-specific rates belong in WITDEM_PRICING_FILE.

Cost limitation

Azure deployment names do not necessarily equal public OpenAI model names. Witdem cannot safely price a deployment from its name alone. Cost is measured only when the observed response model matches the catalog, the provider reports money, or you configure a catalog entry for the deployment/model evidence you emit.