mem0ai--mem0/mem0/llms/azure_openai_structured.py

import copy
import json
import os
from typing import Dict, List, Optional

from azure.identity import DefaultAzureCredential, get_bearer_token_provider
from openai import AzureOpenAI

from mem0.configs.llms.base import BaseLlmConfig
from mem0.llms.base import LLMBase
from mem0.memory.utils import extract_json

SCOPE = "https://cognitiveservices.azure.com/.default"


class AzureOpenAIStructuredLLM(LLMBase):
    def __init__(self, config: Optional[BaseLlmConfig] = None):
        super().__init__(config)

        # Model name should match the custom deployment name chosen for it.
        if not self.config.model:
            self.config.model = "gpt-5-mini"

        api_key = self.config.azure_kwargs.api_key or os.getenv("LLM_AZURE_OPENAI_API_KEY")
        azure_deployment = self.config.azure_kwargs.azure_deployment or os.getenv("LLM_AZURE_DEPLOYMENT")
        azure_endpoint = self.config.azure_kwargs.azure_endpoint or os.getenv("LLM_AZURE_ENDPOINT")
        api_version = self.config.azure_kwargs.api_version or os.getenv("LLM_AZURE_API_VERSION")
        default_headers = self.config.azure_kwargs.default_headers

        # If the API key is not provided or is a placeholder, use DefaultAzureCredential.
        if api_key is None or api_key == "" or api_key == "your-api-key":
            self.credential = DefaultAzureCredential()
            azure_ad_token_provider = get_bearer_token_provider(
                self.credential,
                SCOPE,
            )
            api_key = None
        else:
            azure_ad_token_provider = None

        # Can display a warning if API version is of model and api-version
        self.client = AzureOpenAI(
            azure_deployment=azure_deployment,
            azure_endpoint=azure_endpoint,
            azure_ad_token_provider=azure_ad_token_provider,
            api_version=api_version,
            api_key=api_key,
            http_client=self.config.http_client,
            default_headers=default_headers,
        )

    def generate_response(
        self,
        messages: List[Dict[str, str]],
        response_format: Optional[str] = None,
        tools: Optional[List[Dict]] = None,
        tool_choice: str = "auto",
    ) -> str:
        """
        Generate a response based on the given messages using Azure OpenAI.

        Args:
            messages (List[Dict[str, str]]): A list of dictionaries, each containing a 'role' and 'content' key.
            response_format (Optional[str]): The desired format of the response. Defaults to None.

        Returns:
            str: The generated response.
        """

        # Azure's "Indirect Attacks" content filter can flag the literal word
        # "assistant" in the prompt, so it is rewritten to "ai" before the request.
        # Work on a copy so the caller's messages are left untouched and string-only
        # content is handled without breaking multimodal (list) content.
        messages = self._rewrite_assistant_keyword(messages)

        is_reasoning = self._is_reasoning_model(self.config.model)
        params = {
            "model": self.config.model,
            "messages": messages,
        }
        # Reasoning models (o1/o3/GPT-5 series) reject temperature/top_p; only
        # forward the sampling params for non-reasoning models. Mirrors the
        # reasoning-aware handling of OpenAIStructuredLLM (#5458).
        if not is_reasoning:
            params["temperature"] = self.config.temperature
            params["top_p"] = self.config.top_p
        # Reasoning models require max_completion_tokens rather than max_tokens.
        if is_reasoning or self._uses_max_completion_tokens(self.config.model):
            params["max_completion_tokens"] = self.config.max_tokens
        else:
            params["max_tokens"] = self.config.max_tokens
        if is_reasoning:
            reasoning_effort = getattr(self.config, "reasoning_effort", None)
            if reasoning_effort:
                params["reasoning_effort"] = reasoning_effort
        if response_format:
            params["response_format"] = response_format
        if tools:
            params["tools"] = tools
            params["tool_choice"] = tool_choice

        response = self.client.chat.completions.create(**params)
        return self._parse_response(response, tools)

    @staticmethod
    def _rewrite_assistant_keyword(messages):
        """
        Return a copy of ``messages`` with the word "assistant" replaced by "ai"
        in the last message's textual content.

        Azure's content management policy can flag the literal word "assistant",
        which makes ``add`` fail (see issue #2636). The rewrite targets that
        trigger without mutating the caller's messages and without assuming the
        content is a string, so multimodal (list) content passes through untouched.
        """
        if not messages:
            return messages

        messages = copy.deepcopy(messages)
        last_content = messages[-1].get("content")
        if isinstance(last_content, str):
            messages[-1]["content"] = last_content.replace("assistant", "ai")
        return messages

    def _parse_response(self, response, tools):
        """
        Process the response based on whether tools are used or not.

        Args:
            response: The raw response from API.
            tools: The list of tools provided in the request.

        Returns:
            str or dict: The processed response.
        """
        if tools:
            processed_response = {
                "content": response.choices[0].message.content,
                "tool_calls": [],
            }

            if response.choices[0].message.tool_calls:
                for tool_call in response.choices[0].message.tool_calls:
                    processed_response["tool_calls"].append(
                        {
                            "name": tool_call.function.name,
                            "arguments": json.loads(extract_json(tool_call.function.arguments)),
                        }
                    )

            return processed_response
        else:
            return response.choices[0].message.content