[주의] 병렬처리한 node로 empty state가 전달되는 케이스가 발견되었습니다. 추후 버그 패치될때까지 사용을 추천하지 않습니다.
How to create map-reduce branches for parallel execution에 따라 Map Reduce 형태로 workflow를 생성할 수 있습니다. 상세한 내용은 LangGraph - Controllability with Map Reduce (Youtube)을 참조합니다.
상세한 코드는 map-reduce.ipynb을 참조합니다.
generate_topics 노드를 정의합니다.
class Subjects(BaseModel):
"""List of subjects as a json format"""
subjects: List[str] = Field(
description="different subjects to follow"
)
def generate_topics(state: OverallState):
topic = state['topic']
human = (
"Generate list of between 2 and 5 examples related to: {topic}."
"Provide the only final answer"
)
subjects_prompt = ChatPromptTemplate.from_messages(
[
("human", human),
]
)
subject = subjects_prompt | chat
output = subject.invoke({"topic": topic})
structured_llm = chat.with_structured_output(Subjects, include_raw=True)
response = structured_llm.invoke(output.content)
return {"subjects": response['parsed'].subjects}Joke 노드를 정의합니다.
class Joke(BaseModel):
"""List of jokes as a json format"""
joke: List[str] = Field(
description="a list of jokes to follow"
)
class JokeState(TypedDict):
subject: str
def generate_joke(state: JokeState):
subject = state['subject']
human = (
"Generate a joke about {subject}."
"Provide the only final answer"
)
jok_prompt = ChatPromptTemplate.from_messages(
[
("human", human),
]
)
subject_chain = jok_prompt | chat
output = subject_chain.invoke({"subject": subject})
print('output: ', output)
structured_llm = chat.with_structured_output(Joke, include_raw=True)
response = structured_llm.invoke(output.content)
return {"jokes": response['parsed'].joke}Conditional edge을 정의합니다.
def continue_to_jokes(state: OverallState):
return [Send("generate_joke", {"subject": s}) for s in state["subjects"]]best_joke 노드를 정의합니다.
class BestJoke(BaseModel):
id: int = Field(description="Index of the best joke, starting with 0", ge=0)
def best_joke(state: OverallState):
topic = state['topic']
jokes = "\n\n".join(state["jokes"])
human = (
"Below are a bunch of jokes about {topic}. Select the best one! Return the ID of the best one."
"Provide the only final answer."
)
best_joke_prompt = ChatPromptTemplate.from_messages(
[
("human", human),
("human", "{jokes}")
]
)
best_joke_chain = best_joke_prompt | chat
output = best_joke_chain.invoke({
"topic": topic,
"jokes": jokes
})
print('output: ', output)
structured_llm = chat.with_structured_output(BestJoke, include_raw=True)
response = structured_llm.invoke(output.content)
print('response: ', response)
best_selected_joke = state["jokes"][response['parsed'].id]
return {"best_selected_joke": best_selected_joke}이제 State Diagram을 준비합니다.
from langgraph.constants import Send
from langgraph.graph import END, StateGraph, START
from pydantic import BaseModel, Field
from typing import Annotated, List, Tuple, TypedDict, Literal, Sequence, Union
import operator
class OverallState(TypedDict):
topic: str
subjects: list
jokes: Annotated[list, operator.add]
best_selected_joke: str
graph = StateGraph(OverallState)
graph.add_node("generate_topics", generate_topics)
graph.add_node("generate_joke", generate_joke)
graph.add_node("best_joke", best_joke)
graph.add_edge(START, "generate_topics")
graph.add_conditional_edges("generate_topics", continue_to_jokes, ["generate_joke"])
graph.add_edge("generate_joke", "best_joke")
graph.add_edge("best_joke", END)
app = graph.compile()완성된 diagram은 아래와 같습니다.
from IPython.display import Image
Image(app.get_graph().draw_mermaid_png())
아래와 같이 실행합니다.
for s in app.stream({"topic": "animals"}):
print(s)이때의 결과는 아래와 같습니다.
{'generate_topics': {'subjects': ['cat', 'dog', 'lion', 'giraffe']}}
output: content="Here's a joke about cats:\n\nWhy did the cat go to the computer programming session? Because he wanted to learn some new mouse clicks!" additional_kwargs={'usage': {'prompt_tokens': 19, 'completion_tokens': 32, 'total_tokens': 51}, 'stop_reason': 'end_turn', 'model_id': 'anthropic.claude-3-sonnet-20240229-v1:0'} response_metadata={'usage': {'prompt_tokens': 19, 'completion_tokens': 32, 'total_tokens': 51}, 'stop_reason': 'end_turn', 'model_id': 'anthropic.claude-3-sonnet-20240229-v1:0'} id='run-579d536a-e960-4905-af4a-8e802c172583-0' usage_metadata={'input_tokens': 19, 'output_tokens': 32, 'total_tokens': 51}
output: content="Here's a joke about a dog:\n\nWhy did the dog cross the road twice? Because he was a double-crosser!" additional_kwargs={'usage': {'prompt_tokens': 19, 'completion_tokens': 30, 'total_tokens': 49}, 'stop_reason': 'end_turn', 'model_id': 'anthropic.claude-3-sonnet-20240229-v1:0'} response_metadata={'usage': {'prompt_tokens': 19, 'completion_tokens': 30, 'total_tokens': 49}, 'stop_reason': 'end_turn', 'model_id': 'anthropic.claude-3-sonnet-20240229-v1:0'} id='run-eee580ab-7ae9-4393-b2d8-f007e7ec5953-0' usage_metadata={'input_tokens': 19, 'output_tokens': 30, 'total_tokens': 49}
output: content="Here's a joke about a lion:\n\nWhy did the lion eat the tightrope walker? Because he wanted a well-balanced meal!" additional_kwargs={'usage': {'prompt_tokens': 19, 'completion_tokens': 33, 'total_tokens': 52}, 'stop_reason': 'end_turn', 'model_id': 'anthropic.claude-3-sonnet-20240229-v1:0'} response_metadata={'usage': {'prompt_tokens': 19, 'completion_tokens': 33, 'total_tokens': 52}, 'stop_reason': 'end_turn', 'model_id': 'anthropic.claude-3-sonnet-20240229-v1:0'} id='run-665f8d45-ca86-4539-a948-0129b59ce4ac-0' usage_metadata={'input_tokens': 19, 'output_tokens': 33, 'total_tokens': 52}
{'generate_joke': {'jokes': ['Why did the cat go to the computer programming session? Because he wanted to learn some new mouse clicks!']}}
{'generate_joke': {'jokes': ['Why did the dog cross the road twice? Because he was a double-crosser!']}}
output: content='Why did the giraffe get a neck massage? Because it was a little stiff!' additional_kwargs={'usage': {'prompt_tokens': 21, 'completion_tokens': 22, 'total_tokens': 43}, 'stop_reason': 'end_turn', 'model_id': 'anthropic.claude-3-sonnet-20240229-v1:0'} response_metadata={'usage': {'prompt_tokens': 21, 'completion_tokens': 22, 'total_tokens': 43}, 'stop_reason': 'end_turn', 'model_id': 'anthropic.claude-3-sonnet-20240229-v1:0'} id='run-8a55789c-a90d-496f-8232-b2c90d2ad55c-0' usage_metadata={'input_tokens': 21, 'output_tokens': 22, 'total_tokens': 43}
{'generate_joke': {'jokes': ['Why did the lion eat the tightrope walker? Because he wanted a well-balanced meal!']}}
{'generate_joke': {'jokes': ['Why did the giraffe get a neck massage? Because it was a little stiff!']}}
output: content='3' additional_kwargs={'usage': {'prompt_tokens': 118, 'completion_tokens': 5, 'total_tokens': 123}, 'stop_reason': 'end_turn', 'model_id': 'anthropic.claude-3-sonnet-20240229-v1:0'} response_metadata={'usage': {'prompt_tokens': 118, 'completion_tokens': 5, 'total_tokens': 123}, 'stop_reason': 'end_turn', 'model_id': 'anthropic.claude-3-sonnet-20240229-v1:0'} id='run-300a711a-4503-470d-aabd-fd68adb72d1e-0' usage_metadata={'input_tokens': 118, 'output_tokens': 5, 'total_tokens': 123}
response: {'raw': AIMessage(content='', additional_kwargs={'usage': {'prompt_tokens': 330, 'completion_tokens': 33, 'total_tokens': 363}, 'stop_reason': 'tool_use', 'model_id': 'anthropic.claude-3-sonnet-20240229-v1:0'}, response_metadata={'usage': {'prompt_tokens': 330, 'completion_tokens': 33, 'total_tokens': 363}, 'stop_reason': 'tool_use', 'model_id': 'anthropic.claude-3-sonnet-20240229-v1:0'}, id='run-0f6e9bfe-e66e-4180-9846-46e51264db00-0', tool_calls=[{'name': 'BestJoke', 'args': {'id': 3}, 'id': 'toolu_bdrk_01ScMKVP3CC4zwE4QCR6hRcn', 'type': 'tool_call'}], usage_metadata={'input_tokens': 330, 'output_tokens': 33, 'total_tokens': 363}), 'parsed': BestJoke(id=3), 'parsing_error': None}
{'best_joke': {'best_selected_joke': 'Why did the giraffe get a neck massage? Because it was a little stiff!'}}