Azure OpenAI

Provider azure. Adapter over Azure OpenAI Service — the same OpenAI models (GPT-4o, GPT-4.1, o-series, embeddings…), but hosted on Azure, with Microsoft's SLA, data residency and network isolation. Since the dialect is OpenAI's chat.completions, it inherits from _OpenAICompatible; only the authentication and the fact that the "model" is the deployment name change.

pip install "jangada-ai[openai]"

• provider=: "azure"
• Environment variables: AZURE_OPENAI_API_KEY, AZURE_OPENAI_ENDPOINT (e.g. https://my-resource.openai.azure.com) and AZURE_OPENAI_API_VERSION (e.g. 2024-10-21).
• Model: the deployment name you created in the Azure portal (not the raw model name). E.g. a deployment called gpt-4o.

from jangada_ai import LLM

# uses AZURE_OPENAI_API_KEY / AZURE_OPENAI_ENDPOINT / AZURE_OPENAI_API_VERSION
llm = LLM("azure", "gpt-4o")   # "gpt-4o" = your deployment name
resp = llm.complete("Summarize {{ topic }} in one sentence.", topic="rafts")
print(resp.text, resp.cost)

Configuration via environment

AZURE_OPENAI_API_KEY=...
AZURE_OPENAI_ENDPOINT=https://my-resource.openai.azure.com
AZURE_OPENAI_API_VERSION=2024-10-21

The endpoint and api_version can be passed explicitly in the constructor (via extra=) when you need to point at another resource without touching the environment.

What it does

Everything the openai provider does, since they share the _OpenAICompatible base:

• Text, streaming and structured output (parse) with chat.completions.parse.
• Vision (images=) and documents (files=).
• Audio transcription and embeddings when you have matching deployments (e.g. whisper, text-embedding-3-large).

Note: each capability depends on a deployment of that model existing in your Azure resource. Without an embeddings deployment, there is no embed.

When to choose Azure OpenAI

When the company is already on Azure and needs compliance, data residency and a corporate contract with Microsoft, while keeping OpenAI's models. The code is identical to direct OpenAI — swap LLM("openai", ...) for LLM("azure", ...). See OpenAI and Providers.

What changed in 1.9.0

• profile_model=: on Azure model is the deployment name, so the per-model rules (gpt-5 without temperature, max_completion_tokens…) could not match. Give the base model: LLM("azure", "my-deploy", profile_model="gpt-5").
• Responses API through API v1: remote MCP and native tools (web search, file search, code interpreter, image generation) use <endpoint>/openai/v1/, without api_version. chat.completions keeps api_version (default 2024-10-21, overridable).
• Content-filter chunks (without choices) no longer break the stream.