Autogen

AutoGen은 Microsoft에서 개발한 AI 에이전트 및 애플리케이션을 구축하기 위한 프레임워크입니다. 복잡한 다중 에이전트 시스템 생성을 단순화하고, 대화형 AI(AgentChat), 핵심 다중 에이전트 기능(Core), 외부 서비스와의 통합(Extensions)을 제공합니다. AutoGen은 또한 Studio 코드 없는 에이전트 프로토타이핑을 위한 도구를 제공합니다. 자세한 내용은 공식 AutoGen 문서를 참조하세요.

이 가이드는 AutoGen에 대한 기본적인 이해가 있다고 가정합니다.

Weave는 AutoGen과 통합되어 다중 에이전트 애플리케이션의 실행을 추적하고 시각화하는 데 도움을 줍니다. Weave를 초기화하는 것만으로 autogen_agentchat, autogen_core, 그리고 autogen_ext 내의 상호작용을 자동으로 추적할 수 있습니다. 이 가이드는 Weave를 AutoGen과 함께 사용하는 다양한 예시를 안내합니다.

사전 요구사항

시작하기 전에 AutoGen과 Weave가 설치되어 있는지 확인하세요. 또한 사용하려는 LLM 제공업체(예: OpenAI, Anthropic)의 SDK도 필요합니다.

pip install autogen_agentchat "autogen_ext[openai,anthropic]" weave 

API 키를 환경 변수로 설정하세요:

import os

os.environ["OPENAI_API_KEY"] = "<your-openai-api-key>"
os.environ["ANTHROPIC_API_KEY"] = "<your-anthropic-api-key>"

기본 설정

스크립트 시작 부분에서 Weave를 초기화하여 추적을 시작하세요.

# highlight-next-line
import weave
# highlight-next-line
weave.init("autogen-demo")

간단한 모델 클라이언트 추적

Weave는 AutoGen 내에서 모델 클라이언트에 직접 이루어진 호출을 추적할 수 있습니다.

클라이언트 생성 호출 추적

이 예제는 OpenAIChatCompletionClient에 대한 단일 호출을 추적하는 방법을 보여줍니다.

import asyncio
from autogen_core.models import UserMessage
from autogen_ext.models.openai import OpenAIChatCompletionClient
# from autogen_ext.models.anthropic import AnthropicChatCompletionClient

async def simple_client_call(model_name = "gpt-4o"):
    model_client = OpenAIChatCompletionClient(
        model=model_name,
    )
    # Alternatively, you can use Anthropic or other model clients
    # model_client = AnthropicChatCompletionClient(
        # model="claude-3-haiku-20240307"
    # )
    response = await model_client.create(
        [UserMessage(content="Hello, how are you?", source="user")]
    )
    print(response)

asyncio.run(simple_client_call())

스트리밍을 사용한 클라이언트 생성 호출 추적

Weave는 스트리밍 응답 추적도 지원합니다.

async def simple_client_call_stream(model_name = "gpt-4o"):
    openai_model_client = OpenAIChatCompletionClient(model=model_name)
    async for item in openai_model_client.create_stream(
        [UserMessage(content="Hello, how are you?", source="user")]
    ):
      print(item, flush=True, end="")

asyncio.run(simple_client_call_stream())

Weave는 캐시된 호출을 기록합니다

AutoGen의 ChatCompletionCache를 사용할 수 있으며, Weave는 이러한 상호작용을 추적하여 응답이 캐시에서 왔는지 새로운 호출에서 왔는지 보여줍니다.

from autogen_ext.models.cache import ChatCompletionCache

async def run_cache_client(model_name = "gpt-4o"):
      openai_model_client = OpenAIChatCompletionClient(model=model_name)
      cache_client = ChatCompletionCache(openai_model_client,)

      response = await cache_client.create(
          [UserMessage(content="Hello, how are you?", source="user")]
      )
      print(response)  # Should print response from OpenAI
      response = await cache_client.create(
          [UserMessage(content="Hello, how are you?", source="user")]
      )
      print(response)  # Should print cached response

asyncio.run(run_cache_client())

도구 호출이 있는 에이전트 추적

Weave는 에이전트와 그들의 도구 사용을 추적하여 에이전트가 도구를 선택하고 실행하는 방법에 대한 가시성을 제공합니다.

from autogen_agentchat.agents import AssistantAgent

async def get_weather(city: str) -> str:
    return f"The weather in {city} is 73 degrees and Sunny."

async def run_agent_with_tools(model_name = "gpt-4o"):
    model_client = OpenAIChatCompletionClient(model=model_name)

    agent = AssistantAgent(
        name="weather_agent",
        model_client=model_client,
        tools=[get_weather],
        system_message="You are a helpful assistant.",
        reflect_on_tool_use=True,
    )
    # For streaming output to console:
    # await Console(agent.run_stream(task="What is the weather in New York?"))
    res = await agent.run(task="What is the weather in New York?")
    print(res)
    await model_client.close()

asyncio.run(run_agent_with_tools())

GroupChat 추적 - RoundRobin

그룹 채팅 내의 상호작용(예: RoundRobinGroupChat)은 Weave에 의해 추적되어 에이전트 간의 대화 흐름을 따라갈 수 있습니다.

from autogen_agentchat.conditions import TextMentionTermination
from autogen_agentchat.teams import RoundRobinGroupChat

# we add this weave op here because we want to trace the entire group chat
# it's completely optional but highly recommended to use it

# highlight-next-line
@weave.op
async def run_round_robin_group_chat(model_name="gpt-4o"):
    model_client = OpenAIChatCompletionClient(model=model_name)

    primary_agent = AssistantAgent(
        "primary",
        model_client=model_client,
        system_message="You are a helpful AI assistant.",
    )

    critic_agent = AssistantAgent(
        "critic",
        model_client=model_client,
        system_message="Provide constructive feedback. Respond with 'APPROVE' to when your feedbacks are addressed.",
    )

    text_termination = TextMentionTermination("APPROVE")

    team = RoundRobinGroupChat(
        [primary_agent, critic_agent], termination_condition=text_termination
    )
    await team.reset()
    # For streaming output to console:
    # await Console(team.run_stream(task="Write a short poem about the fall season."))
    result = await team.run(task="Write a short poem about the fall season.")
    print(result)
    await model_client.close()


asyncio.run(run_round_robin_group_chat())

메모리 추적

AutoGen의 메모리 구성 요소는 Weave로 추적할 수 있습니다. @weave.op()를 사용하여 더 나은 가독성을 위해 메모리 작업을 단일 추적으로 그룹화할 수 있습니다.

from autogen_core.memory import ListMemory, MemoryContent, MemoryMimeType

# We add this weave op here because we want to trace 
# the memory add calls along with the memory get calls under a single trace
# it's completely optional but highly recommended to use it

# highlight-next-line
@weave.op
async def run_memory_agent(model_name="gpt-4o"):
    user_memory = ListMemory()

    await user_memory.add(
        MemoryContent(
            content="The weather should be in metric units",
            mime_type=MemoryMimeType.TEXT,
        )
    )

    await user_memory.add(
        MemoryContent(
            content="Meal recipe must be vegan", mime_type=MemoryMimeType.TEXT
        )
    )

    async def get_weather(city: str, units: str = "imperial") -> str:
        if units == "imperial":
            return f"The weather in {city} is 73 °F and Sunny."
        elif units == "metric":
            return f"The weather in {city} is 23 °C and Sunny."
        else:
            return f"Sorry, I don't know the weather in {city}."

    model_client = OpenAIChatCompletionClient(model=model_name)
    assistant_agent = AssistantAgent(
        name="assistant_agent",
        model_client=model_client,
        tools=[get_weather],
        memory=[user_memory],
    )

    # For streaming output to console:
    # stream = assistant_agent.run_stream(task="What is the weather in New York?")
    # await Console(stream)
    result = await assistant_agent.run(task="What is the weather in New York?")
    print(result)
    await model_client.close()


asyncio.run(run_memory_agent())

RAG 워크플로우 추적

검색 증강 생성(RAG) 워크플로우는 문서 인덱싱 및 ChromaDBVectorMemory와 같은 메모리 시스템을 사용한 검색을 포함하여 추적 가능합니다. RAG 프로세스를 @weave.op()로 장식하면 전체 흐름을 시각화하는 데 도움이 됩니다.

RAG 예제는 chromadb가 필요합니다. pip install chromadb로 설치하세요.

# highlight-next-line
# !pip install -q chromadb 
# Ensure chromadb is installed in your environment: `pip install chromadb`

import re
from typing import List
import os
from pathlib import Path

import aiofiles
import aiohttp

from autogen_core.memory import Memory, MemoryContent, MemoryMimeType
from autogen_ext.memory.chromadb import (
    ChromaDBVectorMemory,
    PersistentChromaDBVectorMemoryConfig,
)

class SimpleDocumentIndexer:
    def __init__(self, memory: Memory, chunk_size: int = 1500) -> None:
        self.memory = memory
        self.chunk_size = chunk_size

    async def _fetch_content(self, source: str) -> str:
        if source.startswith(("http://", "https://")):
            async with aiohttp.ClientSession() as session:
                async with session.get(source) as response:
                    return await response.text()
        else:
            async with aiofiles.open(source, "r", encoding="utf-8") as f:
                return await f.read()

    def _strip_html(self, text: str) -> str:
        text = re.sub(r"<[^>]*>", " ", text)
        text = re.sub(r"\\s+", " ", text)
        return text.strip()

    def _split_text(self, text: str) -> List[str]:
        chunks: list[str] = []
        for i in range(0, len(text), self.chunk_size):
            chunk = text[i : i + self.chunk_size]
            chunks.append(chunk.strip())
        return chunks

    async def index_documents(self, sources: List[str]) -> int:
        total_chunks = 0
        for source in sources:
            try:
                content = await self._fetch_content(source)
                if "<" in content and ">" in content:
                    content = self._strip_html(content)
                chunks = self._split_text(content)
                for i, chunk in enumerate(chunks):
                    await self.memory.add(
                        MemoryContent(
                            content=chunk,
                            mime_type=MemoryMimeType.TEXT,
                            metadata={"source": source, "chunk_index": i},
                        )
                    )
                total_chunks += len(chunks)
            except Exception as e:
                print(f"Error indexing {source}: {str(e)}")
        return total_chunks

# highlight-next-line
@weave.op
async def run_rag_agent(model_name="gpt-4o"):
    rag_memory = ChromaDBVectorMemory(
        config=PersistentChromaDBVectorMemoryConfig(
            collection_name="autogen_docs",
            persistence_path=os.path.join(str(Path.home()), ".chromadb_autogen_weave"),
            k=3,
            score_threshold=0.4,
        )
    )
    # await rag_memory.clear() # Uncomment to clear existing memory if needed

    async def index_autogen_docs() -> None:
        indexer = SimpleDocumentIndexer(memory=rag_memory)
        sources = [
            "https://raw.githubusercontent.com/microsoft/autogen/main/README.md",
            "https://microsoft.github.io/autogen/dev/user-guide/agentchat-user-guide/tutorial/agents.html",
        ]
        chunks: int = await indexer.index_documents(sources)
        print(f"Indexed {chunks} chunks from {len(sources)} AutoGen documents")
    
    # Only index if the collection is empty or you want to re-index
    # For demo purposes, we might index it each time or check if already indexed.
    # This example will try to index each run. Consider adding a check.
    await index_autogen_docs()

    model_client = OpenAIChatCompletionClient(model=model_name)
    rag_assistant = AssistantAgent(
        name="rag_assistant",
        model_client=model_client,
        memory=[rag_memory],
    )
    
    # For streaming output to console:
    # stream = rag_assistant.run_stream(task="What is AgentChat?")
    # await Console(stream)
    result = await rag_assistant.run(task="What is AgentChat?")
    print(result)

    await rag_memory.close()
    await model_client.close()

asyncio.run(run_rag_agent())

에이전트 런타임 추적

Weave는 SingleThreadedAgentRuntime와 같은 AutoGen의 에이전트 런타임 내의 작업을 추적할 수 있습니다. @weave.op()를 런타임 실행 함수 주위에 사용하면 관련 추적을 그룹화할 수 있습니다.

from dataclasses import dataclass
from typing import Callable

from autogen_core import (
    DefaultTopicId,
    MessageContext,
    RoutedAgent,
    default_subscription,
    message_handler,
    AgentId,
    SingleThreadedAgentRuntime
)

@dataclass
class Message:
    content: int

@default_subscription
class Modifier(RoutedAgent):
    def __init__(self, modify_val: Callable[[int], int]) -> None:
        super().__init__("A modifier agent.")
        self._modify_val = modify_val

    @message_handler
    async def handle_message(self, message: Message, ctx: MessageContext) -> None:
        val = self._modify_val(message.content)
        print(f"{'-'*80}\\nModifier:\\nModified {message.content} to {val}")
        await self.publish_message(Message(content=val), DefaultTopicId())

@default_subscription
class Checker(RoutedAgent):
    def __init__(self, run_until: Callable[[int], bool]) -> None:
        super().__init__("A checker agent.")
        self._run_until = run_until

    @message_handler
    async def handle_message(self, message: Message, ctx: MessageContext) -> None:
        if not self._run_until(message.content):
            print(f"{'-'*80}\\nChecker:\\n{message.content} passed the check, continue.")
            await self.publish_message(
                Message(content=message.content), DefaultTopicId()
            )
        else:
            print(f"{'-'*80}\\nChecker:\\n{message.content} failed the check, stopping.")

# we add this weave op here because we want to trace 
# the entire agent runtime call under a single trace
# it's completely optional but highly recommended to use it

# highlight-next-line
@weave.op
async def run_agent_runtime() -> None:
    runtime = SingleThreadedAgentRuntime()

    await Modifier.register(
        runtime,
        "modifier",
        lambda: Modifier(modify_val=lambda x: x - 1),
    )

    await Checker.register(
        runtime,
        "checker",
        lambda: Checker(run_until=lambda x: x <= 1),
    )

    runtime.start()
    await runtime.send_message(Message(content=3), AgentId("checker", "default"))
    await runtime.stop_when_idle()

asyncio.run(run_agent_runtime())

워크플로우 추적(순차적)

에이전트 상호작용의 시퀀스를 정의하는 복잡한 에이전트 워크플로우를 추적할 수 있습니다. @weave.op()를 사용하여 전체 워크플로우에 대한 상위 수준 추적을 제공할 수 있습니다.

from autogen_core import TopicId, type_subscription
from autogen_core.models import ChatCompletionClient, SystemMessage, UserMessage

@dataclass
class WorkflowMessage:
    content: str

concept_extractor_topic_type = "ConceptExtractorAgent"
writer_topic_type = "WriterAgent"
format_proof_topic_type = "FormatProofAgent"
user_topic_type = "User"

@type_subscription(topic_type=concept_extractor_topic_type)
class ConceptExtractorAgent(RoutedAgent):
    def __init__(self, model_client: ChatCompletionClient) -> None:
        super().__init__("A concept extractor agent.")
        self._system_message = SystemMessage(
            content=(
                "You are a marketing analyst. Given a product description, identify:\n"
                "- Key features\n"
                "- Target audience\n"
                "- Unique selling points\n\n"
            )
        )
        self._model_client = model_client

    @message_handler
    async def handle_user_description(self, message: WorkflowMessage, ctx: MessageContext) -> None:
        prompt = f"Product description: {message.content}"
        llm_result = await self._model_client.create(
            messages=[self._system_message, UserMessage(content=prompt, source=self.id.key)],
            cancellation_token=ctx.cancellation_token,
        )
        response = llm_result.content
        assert isinstance(response, str)
        print(f"{'-'*80}\\n{self.id.type}:\\n{response}")
        await self.publish_message(
            WorkflowMessage(response), topic_id=TopicId(writer_topic_type, source=self.id.key)
        )

@type_subscription(topic_type=writer_topic_type)
class WriterAgent(RoutedAgent):
    def __init__(self, model_client: ChatCompletionClient) -> None:
        super().__init__("A writer agent.")
        self._system_message = SystemMessage(
            content=(
                "You are a marketing copywriter. Given a block of text describing features, audience, and USPs, "
                "compose a compelling marketing copy (like a newsletter section) that highlights these points. "
                "Output should be short (around 150 words), output just the copy as a single text block."
            )
        )
        self._model_client = model_client
    
    @message_handler
    async def handle_intermediate_text(self, message: WorkflowMessage, ctx: MessageContext) -> None:
        prompt = f"Below is the info about the product:\\n\\n{message.content}"
        llm_result = await self._model_client.create(
            messages=[self._system_message, UserMessage(content=prompt, source=self.id.key)],
            cancellation_token=ctx.cancellation_token,
        )
        response = llm_result.content
        assert isinstance(response, str)
        print(f"{'-'*80}\\n{self.id.type}:\\n{response}")
        await self.publish_message(
            WorkflowMessage(response), topic_id=TopicId(format_proof_topic_type, source=self.id.key)
        )

@type_subscription(topic_type=format_proof_topic_type)
class FormatProofAgent(RoutedAgent):
    def __init__(self, model_client: ChatCompletionClient) -> None:
        super().__init__("A format & proof agent.")
        self._system_message = SystemMessage(
            content=(
                "You are an editor. Given the draft copy, correct grammar, improve clarity, ensure consistent tone, "
                "give format and make it polished. Output the final improved copy as a single text block."
            )
        )
        self._model_client = model_client

    @message_handler
    async def handle_intermediate_text(self, message: WorkflowMessage, ctx: MessageContext) -> None:
        prompt = f"Draft copy:\\n{message.content}."
        llm_result = await self._model_client.create(
            messages=[self._system_message, UserMessage(content=prompt, source=self.id.key)],
            cancellation_token=ctx.cancellation_token,
        )
        response = llm_result.content
        assert isinstance(response, str)
        print(f"{'-'*80}\\n{self.id.type}:\\n{response}")
        await self.publish_message(
            WorkflowMessage(response), topic_id=TopicId(user_topic_type, source=self.id.key)
        )

@type_subscription(topic_type=user_topic_type)
class UserAgent(RoutedAgent):
    def __init__(self) -> None:
        super().__init__("A user agent that outputs the final copy to the user.")

    @message_handler
    async def handle_final_copy(self, message: WorkflowMessage, ctx: MessageContext) -> None:
        print(f"\\n{'-'*80}\\n{self.id.type} received final copy:\\n{message.content}")

# we add this weave op here because we want to trace 
# the entire agent workflow under a single trace
# it's completely optional but highly recommended to use it

# highlight-next-line
@weave.op(call_display_name="Sequential Agent Workflow")
async def run_agent_workflow(model_name="gpt-4o"):
    model_client = OpenAIChatCompletionClient(model=model_name)
    runtime = SingleThreadedAgentRuntime()

    await ConceptExtractorAgent.register(runtime, type=concept_extractor_topic_type, factory=lambda: ConceptExtractorAgent(model_client=model_client))
    await WriterAgent.register(runtime, type=writer_topic_type, factory=lambda: WriterAgent(model_client=model_client))
    await FormatProofAgent.register(runtime, type=format_proof_topic_type, factory=lambda: FormatProofAgent(model_client=model_client))
    await UserAgent.register(runtime, type=user_topic_type, factory=lambda: UserAgent())

    runtime.start()
    await runtime.publish_message(
        WorkflowMessage(
            content="An eco-friendly stainless steel water bottle that keeps drinks cold for 24 hours"
        ),
        topic_id=TopicId(concept_extractor_topic_type, source="default"),
    )
    await runtime.stop_when_idle()
    await model_client.close()

asyncio.run(run_agent_workflow())

코드 실행기 추적

Docker 필요 이 예제는 Docker를 사용한 코드 실행을 포함하며 모든 환경(예: Colab 직접)에서 작동하지 않을 수 있습니다. 이를 시도하는 경우 로컬에서 Docker가 실행 중인지 확인하세요.

Weave는 AutoGen 에이전트에 의한 코드 생성 및 실행을 추적합니다.

import tempfile
from autogen_core import DefaultTopicId
from autogen_core.code_executor import CodeBlock, CodeExecutor
from autogen_core.models import (
    AssistantMessage,
    ChatCompletionClient,
    LLMMessage,
    SystemMessage,
    UserMessage,
)
from autogen_ext.code_executors.docker import DockerCommandLineCodeExecutor


@dataclass
class CodeGenMessage:
    content: str

@default_subscription
class Assistant(RoutedAgent):
    def __init__(self, model_client: ChatCompletionClient) -> None:
        super().__init__("An assistant agent.")
        self._model_client = model_client
        self._chat_history: List[LLMMessage] = [
           SystemMessage(
                content="""Write Python script in markdown block, and it will be executed.
Always save figures to file in the current directory. Do not use plt.show(). All code required to complete this task must be contained within a single response.""",
            )
        ]

    @message_handler
    async def handle_message(self, message: CodeGenMessage, ctx: MessageContext) -> None:
        self._chat_history.append(UserMessage(content=message.content, source="user"))
        result = await self._model_client.create(self._chat_history)
        print(f"\\n{'-'*80}\\nAssistant:\\n{result.content}")
        self._chat_history.append(AssistantMessage(content=result.content, source="assistant"))
        await self.publish_message(CodeGenMessage(content=result.content), DefaultTopicId())

def extract_markdown_code_blocks(markdown_text: str) -> List[CodeBlock]:
    pattern = re.compile(r"```(?:\\s*([\\w\\+\\-]+))?\\n([\\s\\S]*?)```")
    matches = pattern.findall(markdown_text)
    code_blocks: List[CodeBlock] = []
    for match in matches:
        language = match[0].strip() if match[0] else ""
        code_content = match[1]
        code_blocks.append(CodeBlock(code=code_content, language=language))
    return code_blocks

@default_subscription
class Executor(RoutedAgent):
    def __init__(self, code_executor: CodeExecutor) -> None:
        super().__init__("An executor agent.")
        self._code_executor = code_executor

    @message_handler
    async def handle_message(self, message: CodeGenMessage, ctx: MessageContext) -> None:
        code_blocks = extract_markdown_code_blocks(message.content)
        if code_blocks:
            result = await self._code_executor.execute_code_blocks(
                code_blocks, cancellation_token=ctx.cancellation_token
            )
            print(f"\\n{'-'*80}\\nExecutor:\\n{result.output}")
            await self.publish_message(CodeGenMessage(content=result.output), DefaultTopicId())

# we add this weave op here because we want to trace 
# the entire code gen workflow under a single trace
# it's completely optional but highly recommended to use it

# highlight-next-line
@weave.op(call_display_name="CodeGen Agent Workflow")
async def run_codegen(model_name="gpt-4o"): # Updated model
    work_dir = tempfile.mkdtemp()
    runtime = SingleThreadedAgentRuntime()

    # Ensure Docker is running for this example
    try:
        async with DockerCommandLineCodeExecutor(work_dir=work_dir) as executor:
            model_client = OpenAIChatCompletionClient(model=model_name)
            await Assistant.register(runtime, "assistant", lambda: Assistant(model_client=model_client))
            await Executor.register(runtime, "executor", lambda: Executor(executor))

            runtime.start()
            await runtime.publish_message(
                CodeGenMessage(content="Create a plot of NVDA vs TSLA stock returns YTD from 2024-01-01."),
                DefaultTopicId(),
            )
            await runtime.stop_when_idle()
            await model_client.close()
    except Exception as e:
        print(f"Could not run Docker code executor example: {e}")
        print("Please ensure Docker is installed and running.")
    finally:
        import shutil
        shutil.rmtree(work_dir)


asyncio.run(run_codegen())

더 알아보기

• Weave:
  • 추적 가이드
  • Op 데코레이터
• AutoGen:
  • 공식 문서
  • AutoGen GitHub

이 가이드는 Weave와 AutoGen을 통합하기 위한 시작점을 제공합니다. Weave UI를 탐색하여 에이전트 상호작용, 모델 호출 및 도구 사용에 대한 자세한 추적을 확인하세요.