> ## Documentation Index
> Fetch the complete documentation index at: https://docs.zeroruntime.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Azure OpenAI

> Use the Azure OpenAI LLM plugin in a Zero Runtime pipeline. Setup and usage in Python.

Azure OpenAI is an **LLM** plugin. It generates the agent's text responses from the caller's transcribed speech, with support for tool calling and streaming.

## Setup

Set your Azure OpenAI API key in the worker environment. Generate a key from the [Azure AI Foundry portal](https://learn.microsoft.com/en-us/azure/ai-foundry/openai/how-to/create-resource?pivots=web-portal):

```bash theme={null}
export AZURE_OPENAI_API_KEY=<key>
```

## Usage

Import the plugin and pass it to the pipeline's `llm` slot.

```python theme={null}
from zrt.plugins import AzureOpenAILLM

llm = AzureOpenAILLM(
    azure_endpoint="https://<resource>.openai.azure.com/",
    deployment="gpt-4o",
)

# Pipeline(llm=llm, ...)
```

## Parameters

*Constructor parameters for the Python SDK (`AzureOpenAILLM`).*

| Parameter             | Type                                 | Default        | Description                                                                                                                                                                                          |
| --------------------- | ------------------------------------ | -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `api_key`             | `str \| None`                        | `None`         | Azure OpenAI API key. Falls back to the `AZURE_OPENAI_API_KEY` environment variable when unset.                                                                                                      |
| `azure_endpoint`      | `str \| None`                        | `None`         | Full Azure OpenAI resource endpoint URL, e.g. `"https://<resource>.openai.azure.com/"`. Falls back to the `AZURE_OPENAI_ENDPOINT` environment variable when unset.                                   |
| `deployment`          | `AzureOpenAILLMModel \| str \| None` | `None`         | Name of the Azure OpenAI deployment (not the underlying model name) - the custom name chosen when deploying a model in the Azure portal or AI Foundry, e.g. `"gpt-4o-deployment"`. Must be provided. |
| `api_version`         | `str`                                | `"2024-10-21"` | Azure OpenAI REST API version date string.                                                                                                                                                           |
| `temperature`         | `float`                              | `0.7`          | Sampling temperature in `[0.0, 2.0]`. Lower values produce more deterministic output.                                                                                                                |
| `max_output_tokens`   | `int`                                | `1024`         | Maximum tokens the model may generate per response.                                                                                                                                                  |
| `top_p`               | `float \| None`                      | `None`         | Nucleus sampling probability mass in `(0.0, 1.0]`. Mutually exclusive with `temperature`.                                                                                                            |
| `frequency_penalty`   | `float \| None`                      | `None`         | Float in `[-2.0, 2.0]`. Penalizes token repetition based on cumulative frequency so far.                                                                                                             |
| `presence_penalty`    | `float \| None`                      | `None`         | Float in `[-2.0, 2.0]`. Penalizes tokens that have already appeared, encouraging topic diversity.                                                                                                    |
| `seed`                | `int \| None`                        | `None`         | Integer seed for deterministic sampling.                                                                                                                                                             |
| `stop`                | `str \| None`                        | `None`         | String (or up to four strings) at which the model stops generating further tokens.                                                                                                                   |
| `user`                | `str \| None`                        | `None`         | Opaque end-user identifier forwarded to Azure for abuse monitoring.                                                                                                                                  |
| `tool_choice`         | `str \| None`                        | `None`         | Controls how the model selects tools: `"none"`, `"auto"`, `"required"`, or a named tool dict. Defaults to `"auto"` when tools are present.                                                           |
| `parallel_tool_calls` | `bool \| None`                       | `None`         | When `True`, the model may emit multiple tool calls in a single response turn.                                                                                                                       |
| `response_format`     | `dict \| None`                       | `None`         | Enforce a structured output schema - `{"type": "json_object"}` for JSON mode, or a JSON Schema dict for strict structured outputs.                                                                   |
| `reasoning_effort`    | `str \| None`                        | `None`         | Controls reasoning token budget for o-series (reasoning) models: `"low"`, `"medium"`, or `"high"`. Ignored for non-reasoning models.                                                                 |

The LLM's response is synthesized to speech by the [TTS](/plugins/overview) plugin configured in the pipeline.