| 1 | """Azure OpenAI provider using the OpenAI SDK Responses API. |
| 2 | |
| 3 | Uses ``AsyncOpenAI`` pointed at ``https://{endpoint}/openai/v1/`` which |
| 4 | routes to the Responses API (``/responses``). Reuses shared conversion |
| 5 | helpers from :mod:`nanobot.providers.openai_responses`. |
| 6 | """ |
| 7 | |
| 8 | from __future__ import annotations |
| 9 | |
| 10 | import uuid |
| 11 | from collections.abc import Awaitable, Callable |
| 12 | from typing import Any |
| 13 | |
| 14 | from openai import AsyncOpenAI |
| 15 | |
| 16 | from nanobot.providers.base import LLMProvider, LLMResponse |
| 17 | from nanobot.providers.openai_responses import ( |
| 18 | consume_sdk_stream, |
| 19 | convert_messages, |
| 20 | convert_tools, |
| 21 | parse_response_output, |
| 22 | ) |
| 23 | |
| 24 | |
| 25 | class AzureOpenAIProvider(LLMProvider): |
| 26 | """Azure OpenAI provider backed by the Responses API. |
| 27 | |
| 28 | Features: |
| 29 | - Uses the OpenAI Python SDK (``AsyncOpenAI``) with |
| 30 | ``base_url = {endpoint}/openai/v1/`` |
| 31 | - Calls ``client.responses.create()`` (Responses API) |
| 32 | - Reuses shared message/tool/SSE conversion from |
| 33 | ``openai_responses`` |
| 34 | """ |
| 35 | |
| 36 | def __init__( |
| 37 | self, |
| 38 | api_key: str = "", |
| 39 | api_base: str = "", |
| 40 | default_model: str = "gpt-5.2-chat", |
| 41 | ): |
| 42 | super().__init__(api_key, api_base) |
| 43 | self.default_model = default_model |
| 44 | |
| 45 | if not api_key: |
| 46 | raise ValueError("Azure OpenAI api_key is required") |
| 47 | if not api_base: |
| 48 | raise ValueError("Azure OpenAI api_base is required") |
| 49 | |
| 50 | # Normalise: ensure trailing slash |
| 51 | if not api_base.endswith("/"): |
| 52 | api_base += "/" |
| 53 | self.api_base = api_base |
| 54 | |
| 55 | # SDK client targeting the Azure Responses API endpoint |
| 56 | base_url = f"{api_base.rstrip('/')}/openai/v1/" |
| 57 | self._client = AsyncOpenAI( |
| 58 | api_key=api_key, |
| 59 | base_url=base_url, |
| 60 | default_headers={"x-session-affinity": uuid.uuid4().hex}, |
| 61 | max_retries=0, |
| 62 | ) |
| 63 | |
| 64 | # ------------------------------------------------------------------ |
| 65 | # Helpers |
| 66 | # ------------------------------------------------------------------ |
| 67 | |
| 68 | @staticmethod |
| 69 | def _supports_temperature( |
| 70 | deployment_name: str, |
| 71 | reasoning_effort: str | None = None, |
| 72 | ) -> bool: |
| 73 | """Return True when temperature is likely supported for this deployment.""" |
| 74 | if reasoning_effort: |
| 75 | return False |
| 76 | name = deployment_name.lower() |
| 77 | return not any(token in name for token in ("gpt-5", "o1", "o3", "o4")) |
| 78 | |
| 79 | def _build_body( |
| 80 | self, |
| 81 | messages: list[dict[str, Any]], |
| 82 | tools: list[dict[str, Any]] | None, |
| 83 | model: str | None, |
| 84 | max_tokens: int, |
| 85 | temperature: float, |
| 86 | reasoning_effort: str | None, |
| 87 | tool_choice: str | dict[str, Any] | None, |
| 88 | ) -> dict[str, Any]: |
| 89 | """Build the Responses API request body from Chat-Completions-style args.""" |
| 90 | deployment = model or self.default_model |
| 91 | instructions, input_items = convert_messages(self._sanitize_empty_content(messages)) |
| 92 | |
| 93 | body: dict[str, Any] = { |
| 94 | "model": deployment, |
| 95 | "instructions": instructions or None, |
| 96 | "input": input_items, |
| 97 | "max_output_tokens": max(1, max_tokens), |
| 98 | "store": False, |
| 99 | "stream": False, |
| 100 | } |
| 101 | |
| 102 | if self._supports_temperature(deployment, reasoning_effort): |
| 103 | body["temperature"] = temperature |
| 104 | |
| 105 | if reasoning_effort: |
| 106 | body["reasoning"] = {"effort": reasoning_effort} |
| 107 | body["include"] = ["reasoning.encrypted_content"] |
| 108 | |
| 109 | if tools: |
| 110 | body["tools"] = convert_tools(tools) |
| 111 | body["tool_choice"] = tool_choice or "auto" |
| 112 | |
| 113 | return body |
| 114 | |
| 115 | @staticmethod |
| 116 | def _handle_error(e: Exception) -> LLMResponse: |
| 117 | response = getattr(e, "response", None) |
| 118 | body = getattr(e, "body", None) or getattr(response, "text", None) |
| 119 | body_text = str(body).strip() if body is not None else "" |
| 120 | msg = f"Error: {body_text[:500]}" if body_text else f"Error calling Azure OpenAI: {e}" |
| 121 | retry_after = LLMProvider._extract_retry_after_from_headers(getattr(response, "headers", None)) |
| 122 | if retry_after is None: |
| 123 | retry_after = LLMProvider._extract_retry_after(msg) |
| 124 | return LLMResponse(content=msg, finish_reason="error", retry_after=retry_after) |
| 125 | |
| 126 | # ------------------------------------------------------------------ |
| 127 | # Public API |
| 128 | # ------------------------------------------------------------------ |
| 129 | |
| 130 | async def chat( |
| 131 | self, |
| 132 | messages: list[dict[str, Any]], |
| 133 | tools: list[dict[str, Any]] | None = None, |
| 134 | model: str | None = None, |
| 135 | max_tokens: int = 4096, |
| 136 | temperature: float = 0.7, |
| 137 | reasoning_effort: str | None = None, |
| 138 | tool_choice: str | dict[str, Any] | None = None, |
| 139 | ) -> LLMResponse: |
| 140 | body = self._build_body( |
| 141 | messages, tools, model, max_tokens, temperature, |
| 142 | reasoning_effort, tool_choice, |
| 143 | ) |
| 144 | try: |
| 145 | response = await self._client.responses.create(**body) |
| 146 | return parse_response_output(response) |
| 147 | except Exception as e: |
| 148 | return self._handle_error(e) |
| 149 | |
| 150 | async def chat_stream( |
| 151 | self, |
| 152 | messages: list[dict[str, Any]], |
| 153 | tools: list[dict[str, Any]] | None = None, |
| 154 | model: str | None = None, |
| 155 | max_tokens: int = 4096, |
| 156 | temperature: float = 0.7, |
| 157 | reasoning_effort: str | None = None, |
| 158 | tool_choice: str | dict[str, Any] | None = None, |
| 159 | on_content_delta: Callable[[str], Awaitable[None]] | None = None, |
| 160 | ) -> LLMResponse: |
| 161 | body = self._build_body( |
| 162 | messages, tools, model, max_tokens, temperature, |
| 163 | reasoning_effort, tool_choice, |
| 164 | ) |
| 165 | body["stream"] = True |
| 166 | |
| 167 | try: |
| 168 | stream = await self._client.responses.create(**body) |
| 169 | content, tool_calls, finish_reason, usage, reasoning_content = ( |
| 170 | await consume_sdk_stream(stream, on_content_delta) |
| 171 | ) |
| 172 | return LLMResponse( |
| 173 | content=content or None, |
| 174 | tool_calls=tool_calls, |
| 175 | finish_reason=finish_reason, |
| 176 | usage=usage, |
| 177 | reasoning_content=reasoning_content, |
| 178 | ) |
| 179 | except Exception as e: |
| 180 | return self._handle_error(e) |
| 181 | |
| 182 | def get_default_model(self) -> str: |
| 183 | return self.default_model |
| 184 |