그래프가 평면 체인보다 뛰어난 이유

체인의 한계

실제 에이전트는 그래프입니다

LangGraph 소개

Nodes Are Just Functions

from langgraph.graph import StateGraph

class AgentState(TypedDict):
    messages: list
    plan: str
    answer: str

def plan_step(state: AgentState) -> AgentState:
    plan = generate_plan(state['messages'])
    return {'plan': plan}

Edges Connect Nodes

from langgraph.graph import StateGraph, END

graph = StateGraph(AgentState)
graph.add_node('plan', plan_step)
graph.add_node('search', search_step)
graph.add_node('answer', answer_step)

graph.set_entry_point('plan')
graph.add_edge('plan', 'search')
graph.add_edge('search', 'answer')
graph.add_edge('answer', END)

app = graph.compile()

Running the Graph

result = app.invoke({'messages': [{'role': 'user', 'content': 'What is RAG?'}]})
print(result['answer'])

State 업데이트는 병합됩니다

조건부 분기

def needs_more_info(state):
    if state['confidence'] < 0.7:
        return 'search'
    return 'answer'

graph.add_conditional_edges('plan', needs_more_info, {
    'search': 'search',
    'answer': 'answer'
})

Loops Are Just Edges Back

graph.add_edge('search', 'plan')   # loop back to planner
# The planner decides if more search is needed.

시각화

from IPython.display import Image
Image(app.get_graph().draw_mermaid_png())

직접 작성한 반복문과 비교

LangGraph + LCEL

summarise_chain = prompt | model | parser

def summarise_node(state):
    return {'summary': summarise_chain.invoke(state['text'])}

그래프를 사용할 때

복습