ResponsesAgent 简介

什么是 ResponsesAgent？

ResponsesAgent 是 PythonModel 的一个子类，它提供了一种与框架无关的方式来创建代理模型。使用 ResponsesAgent 编写代理可以带来以下好处：

支持返回多个输出消息，包括来自工具调用的中间输出
支持多代理场景
确保与 MLflow 日志记录、跟踪和模型服务兼容
确保您的模型与 OpenAI Responses API 兼容，使其与 OpenAI 的 responses 客户端和其他下游 UI/应用程序兼容

我们推荐使用 ResponsesAgent 而不是 ChatModel 和 ChatAgent，因为它具有 ChatAgent 的所有优点，并且支持其他功能，例如注释。

编写 ResponsesAgent

入门

要创建您自己的代理，请继承 mlflow.pyfunc.ResponsesAgent 并在 predict 方法中实现您的代理逻辑。该实现与框架无关，允许您使用任何代理编写框架。请注意，使用 ResponsesAgent 需要 pydantic>=2。有关示例实现，请参见下面的简单聊天代理和工具调用代理。

创建代理输出

在实现您的代理时，您将使用两种主要的输出类型：ResponsesAgentResponse 和 ResponsesAgentStreamEvent。这些是您应该直接创建的唯一 pydantic 对象。 mlflow.types.responses_helpers 中的其余类仅用于验证字典。

如果您想返回不符合标准接口的输出，可以使用 custom_outputs 字段。

以下是一些辅助方法，可用于在 ResponsesAgent 接口中创建通用输出

这是一个使用 ResponsesAgentResponse 和自定义输出的完整工具调用序列的示例

from mlflow.pyfunc import ResponsesAgent
from mlflow.types.responses import ResponsesAgentRequest, ResponsesAgentResponse


class SimpleResponsesAgent(ResponsesAgent):
    @mlflow.trace(span_type=SpanType.AGENT)
    def predict(self, request: ResponsesAgentRequest) -> ResponsesAgentResponse:
        return ResponsesAgentResponse(
            output=[
                self.create_function_call_item(
                    id="fc_1",
                    call_id="call_1",
                    name="python_exec",
                    arguments='{"code":"result = 4 * 3\\nprint(result)"}',
                ),
                self.create_function_call_output_item(
                    call_id="call_1",
                    output="12\n",
                ),
                self.create_text_output_item(
                    text="The result of 4 * 3 in Python is 12.",
                    id="msg_1",
                ),
            ],
            custom_outputs={"key1": "custom-value1"},
        )

流式传输代理输出

对于实时处理，您可以使用流式传输事件而不是返回完整的响应。流式传输允许您在部分结果可用时发送它们，这对于长时间运行的操作或当您想向用户显示进度时非常有用。

基本文本流式传输

要在 ResponsesAgent 接口中流式传输文本，您应该

使用块（chunks）变为可用状态时，yield response.output_text.delta 事件
- 它必须具有与单个输出项的相关事件相对应的 item_id
yield 一个 response.output_item.done 事件以聚合所有块

from mlflow.types.responses import ResponsesAgentStreamEvent


class SimpleResponsesAgent(ResponsesAgent):
    # ... continuing from above
    @mlflow.trace(span_type=SpanType.AGENT)
    def predict_stream(
        self, request: ResponsesAgentRequest
    ) -> Generator[ResponsesAgentStreamEvent, None, None]:
        # stream text, all with the same item_id
        yield ResponsesAgentStreamEvent(
            **self.create_text_delta(delta="Hello", item_id="msg_1"),
        )
        yield ResponsesAgentStreamEvent(
            **self.create_text_delta(delta="world", item_id="msg_1"),
        )
        yield ResponsesAgentStreamEvent(
            **self.create_text_delta(delta="!", item_id="msg_1"),
        )

        # the text output item id should be the same
        # item_id as the streamed text deltas
        yield ResponsesAgentStreamEvent(
            type="response.output_item.done",
            item=self.create_text_output_item(
                text="Hello world!",
                id="msg_1",
            ),
        )

使用流式传输进行工具调用

您还可以流式传输工具调用及其结果。每个工具调用及其输出都作为单独的 response.output_item.done 事件发送。这使 MLflow 跟踪成为可能，并使客户端更容易重建流式传输的消息历史记录。

from mlflow.types.responses import ResponsesAgentStreamEvent


class SimpleResponsesAgent(ResponsesAgent):
    # ... continuing from above
    @mlflow.trace(span_type=SpanType.AGENT)
    def predict_stream(
        self, request: ResponsesAgentRequest
    ) -> Generator[ResponsesAgentStreamEvent, None, None]:
        yield ResponsesAgentStreamEvent(
            type="response.output_item.done",
            item=self.create_function_call_item(
                id="fc_1",
                call_id="call_1",
                name="python_exec",
                arguments='{"code":"result = 4 * 3\\nprint(result)"}',
            ),
        )
        yield ResponsesAgentStreamEvent(
            type="response.output_item.done",
            item=self.create_function_call_output_item(
                call_id="call_1",
                output="12\n",
            ),
        )
        yield ResponsesAgentStreamEvent(
            type="response.output_item.done",
            item=self.create_text_output_item(
                text="The result of 4 * 3 in Python is 12.",
                id="msg_1",
            ),
        )

记录您的代理

使用来自代码的模型方法记录您的代理。这种方法与框架无关，并支持所有创作框架

with mlflow.start_run():
    logged_agent_info = mlflow.pyfunc.log_model(
        python_model="agent.py",  # replace with your relative path to agent code
        name="agent",
    )

为了方便使用，MLflow 内置了以下功能

自动模型签名推断
- 将设置符合 ResponsesAgentRequest 和 ResponsesAgentResponse 模式的输入和输出签名
元数据
- {"task": "agent/v1/responses"} 将自动附加到您在记录模型时可能传入的任何元数据
输入示例
- 提供输入示例是可选的，默认情况下将使用 mlflow.types.responses.RESPONSES_AGENT_INPUT_EXAMPLE
- 如果您确实提供了输入示例，请确保它是 ResponsesAgentRequest 模式的字典

测试您的代理

要测试 ResponsesAgent，您可以在记录之前和之后传递一个遵循 ResponsesAgentRequest 模式的单个输入字典

from mlflow.pyfunc import ResponsesAgent


class MyResponsesAgent(ResponsesAgent):
    ...


responses_agent = MyResponsesAgent()
responses_agent.predict(
    {
        "input": [{"role": "user", "content": "what is 4*3 in python"}],
        "context": {"conversation_id": "123", "user_id": "456"},
    }
)
# ... log responses_agent using code from above
# load it back from mlflow
loaded_model = mlflow.pyfunc.load_model(path)
loaded_model.predict(
    {
        "input": [{"role": "user", "content": "what is 4*3 in python"}],
        "context": {"conversation_id": "123", "user_id": "456"},
    }
)

从 `ChatAgent` 迁移

从 ChatAgent 迁移到 ResponsesAgent 时，主要任务涉及将您的消息格式从 ChatCompletion API 调整为 Responses API 模式。有关这些更改的详细信息，请参阅 OpenAI 文档

如果您使用的 LLM 提供商使用 chat completions，您可以修改下面的辅助函数，将来自您的 ResponsesAgent 的输出转换为 chat completions 消息，以支持多轮代理聊天

def convert_to_chat_completion_format(message: dict[str, Any]) -> list[dict[str, Any]]:
    """Convert from ResponsesAgent output to a ChatCompletions compatible list of messages"""
    msg_type = message.get("type", None)
    if msg_type == "function_call":
        return [
            {
                "role": "assistant",
                "content": None,
                "tool_calls": [
                    {
                        "id": message["call_id"],
                        "type": "function",
                        "function": {
                            "arguments": message["arguments"],
                            "name": message["name"],
                        },
                    }
                ],
            }
        ]
    elif msg_type == "message" and isinstance(message["content"], list):
        return [
            {"role": message["role"], "content": content["text"]}
            for content in message["content"]
        ]
    elif msg_type == "function_call_output":
        return [
            {
                "role": "tool",
                "content": message["output"],
                "tool_call_id": message["call_id"],
            }
        ]
    compatible_keys = ["role", "content", "name", "tool_calls", "tool_call_id"]
    return [{k: v for k, v in message.items() if k in compatible_keys}]

ResponsesAgent 接口扩展了先前在 ChatAgent 中可用的所有功能，同时引入了新功能。下面，我们概述了两个接口之间针对常见用例的消息表示的关键差异

标准文本响应

ResponsesAgent

{
  "type": "message",
  "id": "",
  "content": [
    {
      "annotations": [],
      "text": "",
      "type": "output_text"
    }
  ],
  "role": "assistant",
  "status": "completed"
}

ChatAgent

{
  "role": "assistant",
  "content": ""
}

工具调用

ResponsesAgent

{
  "type": "function_call",
  "id": "fc_1",
  "arguments": "",
  "call_id": "call_1",
  "name": "",
  "status": "completed"
}

ChatAgent

{
  "role": "assistant",
  "content": "",
  "tool_calls": [
    {
      "id": "call_1",
      "type": "function",
      "function": {
        "name": "",
        "arguments": ""
      }
    }
  ]
}

工具调用结果

ResponsesAgent

{
  "type": "function_call_output",
  "call_id": "call_1",
  "output": ""
}

ChatAgent

{
  "role": "tool",
  "content": "12",
  "tool_call_id": "call_1"
}

工具定义

ResponsesAgent

{
  "name": "",
  "parameters": {},
  "strict": true,
  "type": "function",
  "description": ""
}

ChatAgent

{
  "type": "function",
  "function": {
    "name": "",
    "description": "",
    "parameters": {},
    "strict": true
  }
}

简单聊天示例

这是一个调用带有简单工具的 OpenAI gpt-4o 模型的代理示例

# uncomment below if running inside a jupyter notebook
# %%writefile agent.py
import os
from typing import Generator

import mlflow
from mlflow.entities.span import SpanType
from mlflow.models import set_model
from mlflow.pyfunc.model import ResponsesAgent
from mlflow.types.responses import (
    ResponsesAgentRequest,
    ResponsesAgentResponse,
    ResponsesAgentStreamEvent,
)
from openai import OpenAI


class SimpleResponsesAgent(ResponsesAgent):
    def __init__(self, model: str):
        self.client = OpenAI()
        self.model = model

    @mlflow.trace(span_type=SpanType.AGENT)
    def predict(self, request: ResponsesAgentRequest) -> ResponsesAgentResponse:
        response = self.client.responses.create(input=request.input, model=self.model)
        return ResponsesAgentResponse(**response.to_dict())

    @mlflow.trace(span_type=SpanType.AGENT)
    def predict_stream(
        self, request: ResponsesAgentRequest
    ) -> Generator[ResponsesAgentStreamEvent, None, None]:
        for event in self.client.responses.create(
            input=request.input, stream=True, model=self.model
        ):
            yield ResponsesAgentStreamEvent(**event.to_dict())


mlflow.openai.autolog()
agent = SimpleResponsesAgent(model="gpt-4o")
set_model(agent)

工具调用示例

这是一个调用带有简单工具的 OpenAI gpt-4o 模型的代理示例

# uncomment below if running inside a jupyter notebook
# %%writefile agent.py
import json
from typing import Any, Callable, Generator
import os
from uuid import uuid4

import backoff
import mlflow
import openai
from mlflow.entities import SpanType
from mlflow.pyfunc import ResponsesAgent
from mlflow.types.responses import (
    ResponsesAgentRequest,
    ResponsesAgentResponse,
    ResponsesAgentStreamEvent,
)
from openai import OpenAI
from pydantic import BaseModel


class ToolInfo(BaseModel):
    """
    Class representing a tool for the agent.
    - "name" (str): The name of the tool.
    - "spec" (dict): JSON description of the tool (matches OpenAI Responses format)
    - "exec_fn" (Callable): Function that implements the tool logic
    """

    name: str
    spec: dict
    exec_fn: Callable


class ToolCallingAgent(ResponsesAgent):
    """
    Class representing a tool-calling Agent
    """

    def __init__(self, model: str, tools: list[ToolInfo]):
        """Initializes the ToolCallingAgent with tools."""
        self.model = model
        self.client: OpenAI = OpenAI()
        self._tools_dict = {tool.name: tool for tool in tools}

    def get_tool_specs(self) -> list[dict]:
        """Returns tool specifications in the format OpenAI expects."""
        return [tool_info.spec for tool_info in self._tools_dict.values()]

    @mlflow.trace(span_type=SpanType.TOOL)
    def execute_tool(self, tool_name: str, args: dict) -> Any:
        """Executes the specified tool with the given arguments."""
        return self._tools_dict[tool_name].exec_fn(**args)

    @backoff.on_exception(backoff.expo, openai.RateLimitError)
    @mlflow.trace(span_type=SpanType.LLM)
    def call_llm(self, input_messages) -> ResponsesAgentStreamEvent:
        return (
            self.client.responses.create(
                model=self.model,
                input=input_messages,
                tools=self.get_tool_specs(),
            )
            .output[0]
            .model_dump(exclude_none=True)
        )

    def handle_tool_call(self, tool_call: dict[str, Any]) -> ResponsesAgentStreamEvent:
        """
        Execute tool calls and return a ResponsesAgentStreamEvent w/ tool output
        """
        args = json.loads(tool_call["arguments"])
        result = str(self.execute_tool(tool_name=tool_call["name"], args=args))

        tool_call_output = {
            "type": "function_call_output",
            "call_id": tool_call["call_id"],
            "output": result,
        }
        return ResponsesAgentStreamEvent(
            type="response.output_item.done", item=tool_call_output
        )

    def call_and_run_tools(
        self,
        input_messages,
        max_iter: int = 10,
    ) -> Generator[ResponsesAgentStreamEvent, None, None]:
        for _ in range(max_iter):
            last_msg = input_messages[-1]
            if (
                last_msg.get("type", None) == "message"
                and last_msg.get("role", None) == "assistant"
            ):
                return
            if last_msg.get("type", None) == "function_call":
                tool_call_res = self.handle_tool_call(last_msg)
                input_messages.append(tool_call_res.item)
                yield tool_call_res
            else:
                llm_output = self.call_llm(input_messages=input_messages)
                input_messages.append(llm_output)
                yield ResponsesAgentStreamEvent(
                    type="response.output_item.done",
                    item=llm_output,
                )

        yield ResponsesAgentStreamEvent(
            type="response.output_item.done",
            item={
                "id": str(uuid4()),
                "content": [
                    {
                        "type": "output_text",
                        "text": "Max iterations reached. Stopping.",
                    }
                ],
                "role": "assistant",
                "type": "message",
            },
        )

    @mlflow.trace(span_type=SpanType.AGENT)
    def predict(self, request: ResponsesAgentRequest) -> ResponsesAgentResponse:
        outputs = [
            event.item
            for event in self.predict_stream(request)
            if event.type == "response.output_item.done"
        ]
        return ResponsesAgentResponse(
            output=outputs, custom_outputs=request.custom_inputs
        )

    @mlflow.trace(span_type=SpanType.AGENT)
    def predict_stream(
        self, request: ResponsesAgentRequest
    ) -> Generator[ResponsesAgentStreamEvent, None, None]:
        input_messages = [{"role": "system", "content": SYSTEM_PROMPT}] + [
            i.model_dump() for i in request.input
        ]
        yield from self.call_and_run_tools(input_messages=input_messages)


tools = [
    ToolInfo(
        name="get_weather",
        spec={
            "type": "function",
            "name": "get_weather",
            "description": "Get current temperature for provided coordinates in celsius.",
            "parameters": {
                "type": "object",
                "properties": {
                    "latitude": {"type": "number"},
                    "longitude": {"type": "number"},
                },
                "required": ["latitude", "longitude"],
                "additionalProperties": False,
            },
            "strict": True,
        },
        exec_fn=lambda latitude, longitude: 70,  # dummy tool implementation
    )
]

os.environ["OPENAI_API_KEY"] = "your OpenAI API key"

SYSTEM_PROMPT = "You are a helpful assistant that can call tools to get information."
mlflow.openai.autolog()
AGENT = ToolCallingAgent(model="gpt-4o", tools=tools)
mlflow.models.set_model(AGENT)

什么是 ResponsesAgent？​

编写 ResponsesAgent​

入门​

创建代理输出​

流式传输代理输出​

基本文本流式传输​

使用流式传输进行工具调用​

记录您的代理​

测试您的代理​

从 ChatAgent 迁移​

标准文本响应​

ResponsesAgent​

ChatAgent​

工具调用​

ResponsesAgent​

ChatAgent​

工具调用结果​

ResponsesAgent​

ChatAgent​

工具定义​

ResponsesAgent​

ChatAgent​

简单聊天示例​

工具调用示例​

什么是 ResponsesAgent？

编写 ResponsesAgent

入门

创建代理输出

流式传输代理输出

基本文本流式传输

使用流式传输进行工具调用

记录您的代理

测试您的代理

从 `ChatAgent` 迁移

标准文本响应

ResponsesAgent

ChatAgent

工具调用

ResponsesAgent

ChatAgent

工具调用结果

ResponsesAgent

ChatAgent

工具定义

ResponsesAgent

ChatAgent

简单聊天示例

工具调用示例