LangGraph构建Ai智能体-7-智能体人机交互-HITL

在钱塘江2025-08-11 17:01

前言

人机交互 Human in the loop HITL，简称HITL

涉及人工智能工作流程中的人工监督，使代理能够在关键点暂停、寻求反馈并相应地调整其行为。

这在可能出现错误、敏感作或不确定结果的情况下特别有用。 例如，代理在执行金融交易、完成报告或与敏感数据交互之前可能需要人工批准

核心概念

  • 断点
  • 检查点
  • 状态编辑

示例1 断点的示例

python 复制代码 
import os
import random
import subprocess
import sys
from typing import TypedDict

from dotenv import load_dotenv
from langchain_core.runnables.graph import MermaidDrawMethod
from langchain_openai import ChatOpenAI
from langgraph.checkpoint.memory import MemorySaver
from langgraph.graph import StateGraph, START, END

load_dotenv()
model = ChatOpenAI(model="qwen-max",
                   base_url=os.getenv("BASE_URL"),
                   api_key=os.getenv("OPENAI_API_KEY"),
                   streaming=True)


def display_graph(graph, output_folder="output", ):
    # Code to visualise the graph, we will use this in all lessons
    mermaid_png = graph.get_graph(xray=1).draw_mermaid_png(
        draw_method=MermaidDrawMethod.API
    )

    # Create output folder if it doesn't exist
    output_folder = "."
    os.makedirs(output_folder, exist_ok=True)

    filename = os.path.join(output_folder, f"graph_{random.randint(1, 100000)}.png")
    with open(filename, 'wb') as f:
        f.write(mermaid_png)

    if sys.platform.startswith('darwin'):
        subprocess.call(('open', filename))
    elif sys.platform.startswith('linux'):
        subprocess.call(('xdg-open', filename))
    elif sys.platform.startswith('win'):
        os.startfile(filename)


#
class State(TypedDict):
    input: str


def step_1(state: State):
    print('--- Step 1 ---')
    return state


def step_2(state: State):
    print('--- Step 2 ---')
    return state


builder = StateGraph(State)
builder.add_node('step_1', step_1)
builder.add_node('step_2', step_2)

builder.add_edge(START, 'step_1')
builder.add_edge('step_1', 'step_2')
builder.add_edge('step_2', END)

# Set up memory and breakpoints
memory = MemorySaver()
graph = builder.compile(checkpointer=memory, interrupt_before=['step_2'])
display_graph(graph)

# 执行
config = {'configurable': {'thread_id': 'thread-1'}}
initial_input = {'input': 'Hello, LangGraph!'}
thread = {'configurable': {'thread_id': '1'}}
for event in graph.stream(initial_input, thread, stream_mode='values'):
    print(event)

#     接收用户输入并继续
user_approval = input('Do you approve to continue to Step 2? (yes/no): ')
if user_approval.lower() == 'yes':
    for event in graph.stream(None, thread, stream_mode='values'):
        print(event)
    else:
        print('Execution halted by user.')

输出图为

输入yes的

css 复制代码 
{'input': 'Hello, LangGraph!'}
--- Step 1 ---
{'input': 'Hello, LangGraph!'}
Do you approve to continue to Step 2? (yes/no): yes
{'input': 'Hello, LangGraph!'}
--- Step 2 ---
{'input': 'Hello, LangGraph!'}
Execution halted by user.

输入no的

css 复制代码 
{'input': 'Hello, LangGraph!'}
--- Step 1 ---
{'input': 'Hello, LangGraph!'}
Do you approve to continue to Step 2? (yes/no): no

可以看到输入yes的继续执行第二步了。

示例 2：内容审核的实际示例

python 复制代码 
import os
import random
import subprocess
import sys
from typing import TypedDict

from dotenv import load_dotenv
from langchain_core.runnables.graph import MermaidDrawMethod
from langchain_openai import ChatOpenAI
from langgraph.checkpoint.memory import MemorySaver
from langgraph.graph import StateGraph, START, END

load_dotenv()
llm = ChatOpenAI(model="qwen-max",
                 base_url=os.getenv("BASE_URL"),
                 api_key=os.getenv("OPENAI_API_KEY"),
                 streaming=True)


def display_graph(graph, output_folder="output", ):
    # Code to visualise the graph, we will use this in all lessons
    mermaid_png = graph.get_graph(xray=1).draw_mermaid_png(
        draw_method=MermaidDrawMethod.API
    )

    # Create output folder if it doesn't exist
    output_folder = "."
    os.makedirs(output_folder, exist_ok=True)

    filename = os.path.join(output_folder, f"graph_{random.randint(1, 100000)}.png")
    with open(filename, 'wb') as f:
        f.write(mermaid_png)

    if sys.platform.startswith('darwin'):
        subprocess.call(('open', filename))
    elif sys.platform.startswith('linux'):
        subprocess.call(('xdg-open', filename))
    elif sys.platform.startswith('win'):
        os.startfile(filename)


class State(TypedDict):
    input: str
    draft_content: str


def create_draft(state: State):
    print("--- 生成草稿 ---")
    prompt = f"写一个主题是{state['input']}的博客"
    response = llm.invoke([{"role": "user", "content": prompt}])
    state["draft_content"] = response.content
    print(f"草稿内容:{response.content}")
    return state


def review_draft(state: State):
    print("--- 审核草稿 ---")
    return state


def publish_content(state: State):
    print("--- 发布内容 ---")
    return state


builder = StateGraph(State)
builder.add_node("create_draft", create_draft)
builder.add_node("review_draft", review_draft)
builder.add_node("publish_content", publish_content)

builder.add_edge(START, "create_draft")
builder.add_edge("create_draft", "review_draft")
builder.add_edge("review_draft", "publish_content")
builder.add_edge("publish_content", END)

memory = MemorySaver()
graph = builder.compile(checkpointer=memory, interrupt_before=["publish_content"])
display_graph(graph)

config = {"configurable": {"thread_id": "thread-1"}}
initial_input = {"input": "AI在现代创作中的重要性"}
thread = {"configurable": {"thread_id": "1"}}
for event in graph.stream(initial_input, thread, stream_mode="values"):
    print(event)
user_approval = input("需要发布草稿吗? (yes/no/mod): ")
if user_approval.lower() == "yes":
    # 继续调用
    for event in graph.stream(None, thread, stream_mode="values"):
        print(event)
elif user_approval.lower() == "mod":
    updated_draft = input("请修改草稿内容:\n")
    #memory.update({"draft_content": updated_draft})  # Update memory with new content
    print("修改完毕")
    for event in graph.stream(None, thread, stream_mode="values"):
        print(event)
else:
    print("用户终止")

输出结果

bash 复制代码 
{'input': 'AI在现代创作中的重要性'}
--- 生成草稿 ---
草稿内容:# AI在现代创作中的重要性 随着科技的不断进步，人工智能（AI）...'}
--- 审核草稿 ---
{'input': 'AI在现代创作中的重要性', 'draft_content': '# AI在现代创作中的重要性 ...'}
需要发布草稿吗? (yes/no/mod): yes
{'input': 'AI在现代创作中的重要性', 'draft_content': '# AI在现代创作中的重要性 ...'}
--- 发布内容 ---
{'input': 'AI在现代创作中的重要性', 'draft_content': '# AI在现代创作中的重要性 ...'}

ReAct 代理示例与金融股票购买用例

使用人机交互 （HITL） 架构实现 ReAct 代理，以实现股票购买决策工作流程。代理将使用 Finnhub API 查询实时股票价格，推理是否购买股票，并在继续之前暂停以请求人工批准。

python 复制代码 
import os
import random
import subprocess
import sys
from typing import TypedDict

from dotenv import load_dotenv
from langchain_core.runnables.graph import MermaidDrawMethod
from langchain_openai import ChatOpenAI
from langgraph.checkpoint.memory import MemorySaver
from langgraph.graph import StateGraph, START, END, MessagesState
from langchain_core.tools import tool
from langgraph.prebuilt import ToolNode

load_dotenv()
llm = ChatOpenAI(model="qwen-max",
                 base_url=os.getenv("BASE_URL"),
                 api_key=os.getenv("OPENAI_API_KEY"),
                 streaming=True)


def display_graph(graph, output_folder="output", ):
    # Code to visualise the graph, we will use this in all lessons
    mermaid_png = graph.get_graph(xray=1).draw_mermaid_png(
        draw_method=MermaidDrawMethod.API
    )

    # Create output folder if it doesn't exist
    output_folder = "."
    os.makedirs(output_folder, exist_ok=True)

    filename = os.path.join(output_folder, f"graph_{random.randint(1, 100000)}.png")
    with open(filename, 'wb') as f:
        f.write(mermaid_png)

    if sys.platform.startswith('darwin'):
        subprocess.call(('open', filename))
    elif sys.platform.startswith('linux'):
        subprocess.call(('xdg-open', filename))
    elif sys.platform.startswith('win'):
        os.startfile(filename)


@tool
def get_stock_price(symbol: str):
    """Retrieve the latest stock price for the given symbol."""
    # quote = finnhub_client.quote(symbol)
    price = random.randint(500, 800)
    return f"The current price for {symbol} is ${price}."


tools = [get_stock_price]
tool_node = ToolNode(tools)
model = llm.bind_tools(tools)


def agent_reasoning(state):
    messages = state["messages"]
    response = model.invoke(messages)
    return {"messages": [response]}


def should_continue(state):
    messages = state["messages"]
    last_message = messages[-1]
    # 没有工具调用就结束
    if not last_message.tool_calls:
        return "end"
    return "continue"


workflow = StateGraph(MessagesState)
workflow.add_node("agent_reasoning", agent_reasoning)
workflow.add_node("call_tool", tool_node)
workflow.add_edge(START, "agent_reasoning")
workflow.add_conditional_edges(
    "agent_reasoning", should_continue,
    {
        "continue": "call_tool",
        "end": END
    }
)
# Normal edge: after invoking the tool, return to agent reasoning
workflow.add_edge("call_tool", "agent_reasoning")

memory = MemorySaver()
app = workflow.compile(checkpointer=memory, interrupt_before=["call_tool"])
display_graph(app)

initial_input = {"messages": [{"role": "user", "content": "今天应该买AAPL的股票吗?"}]}
thread = {"configurable": {"thread_id": "1"}}

# Run the agent reasoning step first
for event in app.stream(initial_input, thread, stream_mode="values"):
    print(event)

user_approval = input("是否查询AAPL的股价? (yes/no): ")
if user_approval.lower() == "yes":
    # Continue with tool invocation to get stock price
    for event in app.stream(None, thread, stream_mode="values"):
        print(event)
else:
    print("用户终止")

输出结果

bash 复制代码 
{'messages': [HumanMessage(content='今天应该买AAPL的股票吗?', additional_kwargs={}, response_metadata={}, id='1506d0d0-ee04-4245-8c83-09f3e7814918')]}
{'messages': [HumanMessage(content='今天应该买AAPL的股票吗?', additional_kwargs={}, response_metadata={}, id='1506d0d0-ee04-4245-8c83-09f3e7814918'), AIMessage(content='决定是否购买AAPL（苹果公司）的股票取决于多种因素，：\n\n此外，获取最新的股票价格信息也可能有助于您的决策过程。我可以通过调用`get_stock_price`函数来帮助您获取AAPL的最新股票价格。让我们来查一下。', additional_kwargs={'tool_calls': [{'index': 0, 'id': 'call_5b9dcc6901f8422faceb06', 'function': {'arguments': '{"symbol": "AAPL"}', 'name': 'get_stock_price'}, 'type': 'function'}]}, response_metadata={'finish_reason': 'tool_calls', 'model_name': 'qwen-max'}, id='run--f732f6e9-4c5d-4426-86c4-81b1bb8d03ca-0', tool_calls=[{'name': 'get_stock_price', 'args': {'symbol': 'AAPL'}, 'id': 'call_5b9dcc6901f8422faceb06', 'type': 'tool_call'}])]}
是否查询AAPL的股价? (yes/no): yes
{'messages': [HumanMessage(content='今天应该买AAPL的股票吗?', additional_kwargs={}, response_metadata={}, id='1506d0d0-ee04-4245-8c83-09f3e7814918'), AIMessage(content='决定是否购买AAPL（苹果公司）的股票取决于多种因素，：\n\n此外，获取最新的股票价格信息也可能有助于您的决策过程。我可以通过调用`get_stock_price`函数来帮助您获取AAPL的最新股票价格。让我们来查一下。', additional_kwargs={'tool_calls': [{'index': 0, 'id': 'call_5b9dcc6901f8422faceb06', 'function': {'arguments': '{"symbol": "AAPL"}', 'name': 'get_stock_price'}, 'type': 'function'}]}, response_metadata={'finish_reason': 'tool_calls', 'model_name': 'qwen-max'}, id='run--f732f6e9-4c5d-4426-86c4-81b1bb8d03ca-0', tool_calls=[{'name': 'get_stock_price', 'args': {'symbol': 'AAPL'}, 'id': 'call_5b9dcc6901f8422faceb06', 'type': 'tool_call'}])]}
{'messages': [HumanMessage(content='今天应该买AAPL的股票吗?', additional_kwargs={}, response_metadata={}, id='1506d0d0-ee04-4245-8c83-09f3e7814918'), AIMessage(content='决定是否购买AAPL（苹果公司）的股票取决于多种因素，：\n\n此外，获取最新的股票价格信息也可能有助于您的决策过程。我可以通过调用`get_stock_price`函数来帮助您获取AAPL的最新股票价格。让我们来查一下。', additional_kwargs={'tool_calls': [{'index': 0, 'id': 'call_5b9dcc6901f8422faceb06', 'function': {'arguments': '{"symbol": "AAPL"}', 'name': 'get_stock_price'}, 'type': 'function'}]}, response_metadata={'finish_reason': 'tool_calls', 'model_name': 'qwen-max'}, id='run--f732f6e9-4c5d-4426-86c4-81b1bb8d03ca-0', tool_calls=[{'name': 'get_stock_price', 'args': {'symbol': 'AAPL'}, 'id': 'call_5b9dcc6901f8422faceb06', 'type': 'tool_call'}]), ToolMessage(content='The current price for AAPL is $517.', name='get_stock_price', id='51f1adeb-c67a-4a75-95eb-c53ff3ce6776', tool_call_id='call_5b9dcc6901f8422faceb06')]}
{'messages': [HumanMessage(content='今天应该买AAPL的股票吗?', additional_kwargs={}, response_metadata={}, id='1506d0d0-ee04-4245-8c83-09f3e7814918'), AIMessage(content='决定是否购买AAPL（苹果公司）的股票取决于多种因素，：\n\n此外，获取最新的股票价格信息也可能有助于您的决策过程。我可以通过调用`get_stock_price`函数来帮助您获取AAPL的最新股票价格。让我们来查一下。', additional_kwargs={'tool_calls': [{'index': 0, 'id': 'call_5b9dcc6901f8422faceb06', 'function': {'arguments': '{"symbol": "AAPL"}', 'name': 'get_stock_price'}, 'type': 'function'}]}, response_metadata={'finish_reason': 'tool_calls', 'model_name': 'qwen-max'}, id='run--f732f6e9-4c5d-4426-86c4-81b1bb8d03ca-0', tool_calls=[{'name': 'get_stock_price', 'args': {'symbol': 'AAPL'}, 'id': 'call_5b9dcc6901f8422faceb06', 'type': 'tool_call'}]), ToolMessage(content='The current price for AAPL is $517.', name='get_stock_price', id='51f1adeb-c67a-4a75-95eb-c53ff3ce6776', tool_call_id='call_5b9dcc6901f8422faceb06'), AIMessage(content='苹果公司（AAPL）的最新股票价格是 $517。\n\n请记住，这只代表了 AAPL 股票的当前市场价格，并不单独构成买卖建议。您应该结合自己的研究和其他信息来源来决定是否购买。如果您需要进一步的帮助来进行分析或有其他问题，请随时告诉我。', additional_kwargs={}, response_metadata={'finish_reason': 'stop', 'model_name': 'qwen-max'}, id='run--14905aec-befd-4689-a6ba-b7864d64fed7-0')]}

在执行过程中编辑图形状态

在继续之前修改图形状态

python 复制代码 
from typing_extensions import TypedDict
from langgraph.graph import StateGraph, START, END
from langgraph.checkpoint.memory import MemorySaver


# Define the state structure
class State(TypedDict):
    input: str
    modified_input: str


def step_1(state: State):
    print(f"Original input: {state['input']}")
    return state


def modify_state(state: State):
    # Allow the user to modify the state
    return state


def step_3(state: State):
    print(f"Modified input: {state['modified_input']}")
    return state


builder = StateGraph(State)
builder.add_node("step_1", step_1)
builder.add_node("modify_state", modify_state)
builder.add_node("step_3", step_3)

# Define the flow
builder.add_edge(START, "step_1")
builder.add_edge("step_1", "modify_state")
builder.add_edge("modify_state", "step_3")
builder.add_edge("step_3", END)

# Set up memory and breakpoints
memory = MemorySaver()
graph = builder.compile(checkpointer=memory, interrupt_before=["modify_state"])
initial_input = {"input": "Initial Input"}
config = {"configurable": {"thread_id": "thread-1"}}

for event in graph.stream(initial_input, config):
    print(event)

modified_value = input("Enter the modified input: ")
graph.update_state(config, {"modified_input": modified_value})

# Continue the graph execution
for event in graph.stream(None, config):
    print(event)

结果为

python 复制代码 
Original input: Initial Input
{'step_1': {'input': 'Initial Input'}}
{'__interrupt__': ()}
Enter the modified input: 12322
{'modify_state': {'input': 'Initial Input', 'modified_input': '12322'}}
Modified input: 12322
{'step_3': {'input': 'Initial Input', 'modified_input': '12322'}}

断点的示例

同意金融交易

python 复制代码 
from typing_extensions import TypedDict
from langgraph.graph import StateGraph, START, END
from langgraph.checkpoint.memory import MemorySaver


class State(TypedDict):
    amount: float


builder = StateGraph(State)


def define__transaction(state: State):
    print("Defining transaction...")
    return state


def verify_transaction(state: State):
    print(f"Verifying transaction amount: {state['amount']}")
    return state


builder.add_node("define_transaction", define__transaction)
builder.add_node("verify_transaction", verify_transaction)
builder.add_edge(START, "define_transaction")
builder.add_edge("define_transaction", "verify_transaction")
builder.add_edge("verify_transaction", END)

graph = builder.compile(
    interrupt_before=["verify_transaction"],
    checkpointer=MemorySaver())

initial_input = {"amount": 1000.0}
config = {"configurable": {"thread_id": "thread-1"}}
for event in graph.stream(initial_input, config):
    print(event)
approval = input("Approve this transaction? (yes/no): ")
if approval.lower() == "yes":
    for event in graph.stream(None, config):
        print(event)
else:
    print("Transaction cancelled.")

输出结果

bash 复制代码 
"""
Defining transaction...
{'define_transaction': {'amount': 1000.0}}
{'__interrupt__': ()}
Approve this transaction? (yes/no): yes
Verifying transaction amount: 1000.0
{'verify_transaction': {'amount': 1000.0}}
"""

删除数据确认

python 复制代码 
import os
from typing_extensions import TypedDict
from langgraph.graph import StateGraph, START, END
from langgraph.checkpoint.memory import MemorySaver


class State(TypedDict):
    data: str


def delete_data(state: State):
    print(f"Data to be deleted: {state['data']}")
    return state


builder = StateGraph(State)
builder.add_node("delete_data", delete_data)
builder.add_edge(START, "delete_data")
builder.add_edge("delete_data", END)

graph = builder.compile(interrupt_before=["delete_data"], checkpointer=MemorySaver())

initial_input = {"data": "Sensitive Information"}
config = {"configurable": {"thread_id": "thread-1"}}

for event in graph.stream(initial_input, config):
    print(event)

approval = input("Approve data deletion? (yes/no): ")
if approval.lower() == "yes":
    for event in graph.stream(None, config):
        print(event)
else:
    print("Data deletion cancelled.")

输出结果

bash 复制代码 
"""
{'__interrupt__': ()}
Approve data deletion? (yes/no): yes
Data to be deleted: Sensitive Information
{'delete_data': {'data': 'Sensitive Information'}}

"""

允许tool call

python 复制代码 
import os
from typing_extensions import TypedDict
from langgraph.graph import StateGraph, START, END
from langgraph.checkpoint.memory import MemorySaver
from langchain_core.tools import tool


class State(TypedDict):
    query: str


@tool
def perform_query(query: str):
    """:param query: The SQL query to be executed."""
    print(f"Performing query: {query}")
    return {"query": {query}}


def review_query(state: State):
    print(f"Reviewing query: {state['query']}")
    return state


builder = StateGraph(State)
builder.add_node("perform_query", perform_query)
builder.add_node("review_query", review_query)

builder.add_edge(START, "review_query")
builder.add_edge("review_query", "perform_query")
builder.add_edge("perform_query", END)

graph = builder.compile(interrupt_before=["perform_query"],
                        checkpointer=MemorySaver())

initial_input = {"query": "SELECT * FROM users"}
config = {"configurable": {"thread_id": "thread-1"}}

for event in graph.stream(initial_input, config):
    print(event)

approval = input("Approve query execution? (yes/no): ")
if approval.lower() == "yes":
    for event in graph.stream(None, config):
        print(event)
else:
    print("Query execution cancelled.")

输出结果

bash 复制代码 
{'__interrupt__': ()}
Approve data deletion? (yes/no): yes
Data to be deleted: Sensitive Information
{'delete_data': {'data': 'Sensitive Information'}}

动态断点

动态断点允许代理根据运行时数据或外部触发器有条件地暂停，从而增加了灵活性。动态断点不是预先定义图形的暂停位置，而是引入了根据特定条件触发暂停的功能。

基于数据阈值的条件暂停

python 复制代码 
from langgraph.types import interrupt
import os
from typing_extensions import TypedDict
from langgraph.graph import StateGraph, START, END


class State(TypedDict):
    input: str


def step_with_dynamic_interrupt(state: State):
    input_length = len(state["input"])
    if input_length > 10:
        interrupt(f"Input length {input_length} exceeds threshold of 10.")
    return state


builder = StateGraph(State)
builder.add_node("step_with_dynamic_interrupt", step_with_dynamic_interrupt)
builder.add_edge(START, "step_with_dynamic_interrupt")
builder.add_edge("step_with_dynamic_interrupt", END)

graph = builder.compile()

initial_input = {"input": "This is a long input"}
for event in graph.stream(initial_input):
    print(event)

运行结果

  • 当输入长度超过 10 个字符时，工作流会自动暂停
bash 复制代码 
{'__interrupt__': (Interrupt(value='Input length 20 exceeds threshold of 10.', id='cf3c8cd72f33e6120e00b8c619f411ce'),)}

等待用户输入

一些流程中，AI 代理需要人工反馈才能继续。例如客户服务、医疗保健或业务审批流程。

  • 等待用户输入是指人工智能代理在其工作流程中暂停，要求人工澄清或附加信息，然后仅在收到信息后恢复的交互
  • 在 LangGraph 中，等待用户输入是使用断点实现的。在断点处，图形暂停执行并等待用户反馈，然后将反馈注入回图形的状态。工作流使用提供的输入从该点恢复。
python 复制代码 
from typing_extensions import TypedDict
from langgraph.graph import StateGraph, START, END
from langgraph.checkpoint.memory import MemorySaver


class State(TypedDict):
    input: str
    user_feedback: str


def step_1(state: State):
    print(f"Step 1: {state['input']}")
    return state


def human_feedback(state: State):
    print("--- Waiting for human feedback ---")
    feedback = input("Please provide your feedback: ")
    state['user_feedback'] = feedback
    return state


def step_3(state: State):
    print(f"Step 3: User feedback received: {state['user_feedback']}")
    return state


builder = StateGraph(State)
builder.add_node("step_1", step_1)
builder.add_node("human_feedback", human_feedback)
builder.add_node("step_3", step_3)

builder.add_edge(START, "step_1")
builder.add_edge("step_1", "human_feedback")
builder.add_edge("human_feedback", "step_3")
builder.add_edge("step_3", END)

# Set up memory and breakpoints
memory = MemorySaver()
graph = builder.compile(checkpointer=memory,
                        interrupt_before=["human_feedback"])

initial_input = {"input": "Proceed with workflow?"}
thread = {"configurable": {"thread_id": "1"}}

for event in graph.stream(initial_input, thread, stream_mode="values"):
    print(event)

user_feedback = input("User feedback: ")
graph.update_state(thread, {"user_feedback": user_feedback},
                   as_node='human_feedback')

for event in graph.stream(None, thread, stream_mode="values"):
    print(event)

运行结果

css 复制代码 
{'input': 'Proceed with workflow?'}
Step 1: Proceed with workflow?
{'input': 'Proceed with workflow?'}
User feedback: 1234
{'input': 'Proceed with workflow?', 'user_feedback': '1234'}
Step 3: User feedback received: 1234
{'input': 'Proceed with workflow?', 'user_feedback': '1234'}

复杂例子

python 复制代码 
from typing_extensions import TypedDict
from langgraph.graph import StateGraph, START, END
from langgraph.checkpoint.memory import MemorySaver
from langchain_core.tools import tool
import os
import random
import subprocess
import sys
from typing import TypedDict

from dotenv import load_dotenv
from langchain_core.runnables.graph import MermaidDrawMethod
from langchain_openai import ChatOpenAI
from langgraph.checkpoint.memory import MemorySaver
from langgraph.graph import StateGraph, START, END, MessagesState
from langchain_core.tools import tool
from langgraph.prebuilt import ToolNode


def display_graph(graph, output_folder="output", ):
    # Code to visualise the graph, we will use this in all lessons
    mermaid_png = graph.get_graph(xray=1).draw_mermaid_png(
        draw_method=MermaidDrawMethod.API
    )

    # Create output folder if it doesn't exist
    output_folder = "."
    os.makedirs(output_folder, exist_ok=True)

    filename = os.path.join(output_folder, f"graph_{random.randint(1, 100000)}.png")
    with open(filename, 'wb') as f:
        f.write(mermaid_png)

    if sys.platform.startswith('darwin'):
        subprocess.call(('open', filename))
    elif sys.platform.startswith('linux'):
        subprocess.call(('xdg-open', filename))
    elif sys.platform.startswith('win'):
        os.startfile(filename)


class State(TypedDict):
    input: str
    user_feedback: str


def agent_reasoning(state: State):
    print(f"Agent is reasoning: {state['input']}")
    # Agent decides whether to ask human based on input length
    if len(state["input"]) > 10:
        print("Agent needs clarification.")
        return state
    else:
        state["user_feedback"] = "No clarification needed"
        return state


def ask_human(state: State):
    print("--- Asking for human feedback ---")
    feedback = input("Please provide feedback on the input: ")
    state['user_feedback'] = feedback
    return state


# Define a tool action after human feedback
@tool
def perform_action(user_feedback: str):
    """
    Perform an action based on the provided user feedback.
    """
    print(f"Action taken based on feedback: {user_feedback}")
    return {"user_feedback": f"Feedback processed: {user_feedback}"}


builder = StateGraph(State)
builder.add_node("agent_reasoning", agent_reasoning)
builder.add_node("ask_human", ask_human)
builder.add_node("perform_action", perform_action)

builder.add_edge(START, "agent_reasoning")
builder.add_conditional_edges(
    "agent_reasoning",
    lambda state: "ask_human" if len(state["input"]) > 10 else "perform_action",
    {
        "ask_human": "ask_human",
        "perform_action": "perform_action"
    }
)
builder.add_edge("ask_human", "perform_action")
builder.add_edge("perform_action", END)

memory = MemorySaver()
graph = builder.compile(checkpointer=memory, interrupt_before=["ask_human"])
display_graph(graph)

# Run the graph
initial_input = {"input": "Proceed with reasoning?"}
thread = {"configurable": {"thread_id": "1"}}

# Stream the graph until the first interruption
for event in graph.stream(initial_input, thread, stream_mode="values"):
    print(event)

# Get user input and update the state
user_feedback = input("User feedback: ")
graph.update_state(thread, {"user_feedback": user_feedback},
                   as_node="ask_human")

# Resume execution after feedback
for event in graph.stream(None, thread, stream_mode="values"):
    print(event)

运行结果

css 复制代码 
{'input': 'Proceed with reasoning?'}
Agent is reasoning: Proceed with reasoning?
Agent needs clarification.
{'input': 'Proceed with reasoning?'}
User feedback: 12324
{'input': 'Proceed with reasoning?', 'user_feedback': '12324'}
Action taken based on feedback: 12324
{'input': 'Proceed with reasoning?', 'user_feedback': 'Feedback processed: 12324'}

查看并更新历史状态

  • get_state_history 获取某会话的历史状态
  • stream 方法支持从历史的某一个状态继续会话
  • update_state 支持修改历史状态记录
python 复制代码 
from langchain_openai import ChatOpenAI
from langchain_core.tools import tool
from langgraph.graph import MessagesState, START, END, StateGraph
from langgraph.checkpoint.memory import MemorySaver
from langgraph.prebuilt import ToolNode
from dotenv import load_dotenv
import os
from langchain_core.runnables.graph import MermaidDrawMethod
from langchain_core.messages import HumanMessage
import random
import subprocess
import sys


def display_graph(graph, output_folder="output", ):
    # Code to visualise the graph, we will use this in all lessons
    mermaid_png = graph.get_graph(xray=1).draw_mermaid_png(
        draw_method=MermaidDrawMethod.API
    )

    # Create output folder if it doesn't exist
    output_folder = "."
    os.makedirs(output_folder, exist_ok=True)

    filename = os.path.join(output_folder, f"graph_{random.randint(1, 100000)}.png")
    with open(filename, 'wb') as f:
        f.write(mermaid_png)

    if sys.platform.startswith('darwin'):
        subprocess.call(('open', filename))
    elif sys.platform.startswith('linux'):
        subprocess.call(('xdg-open', filename))
    elif sys.platform.startswith('win'):
        os.startfile(filename)


load_dotenv()
model = ChatOpenAI(model="qwen-max",
                   base_url=os.getenv("BASE_URL"),
                   api_key=os.getenv("OPENAI_API_KEY"),
                   streaming=True)


# Define tools
@tool
def play_song_on_spotify(song: str):
    """Play a song on Spotify."""
    return f"Successfully played {song} on Spotify!"


@tool
def play_song_on_apple(song: str):
    """Play a song on Apple Music."""
    return f"Successfully played {song} on Apple Music!"


# List of tools
tools = [play_song_on_apple, play_song_on_spotify]
tool_node = ToolNode(tools)

# Set up model
model = model.bind_tools(tools)


# Define model-calling function
def call_model(state):
    response = model.invoke(state["messages"])
    return {"messages": [response]}


# Define continuation logic
def should_continue(state):
    last_message = state["messages"][-1]
    if last_message.tool_calls:
        return "continue"
    return "end"


# Build the graph
workflow = StateGraph(MessagesState)
workflow.add_node("agent", call_model)
workflow.add_node("action", tool_node)

# Define graph flow
workflow.add_edge(START, "agent")
workflow.add_conditional_edges(
    "agent",
    should_continue,
    {
        "continue": "action",
        "end": END
    }
)

workflow.add_edge("action", "agent")
# Set up memory for checkpointing
memory = MemorySaver()
app = workflow.compile(checkpointer=memory)
# display_graph(app)

config = {"configurable": {"thread_id": "1"}}
input_message = HumanMessage(content="你可以播放Taylor Swift的最流行的歌曲吗?")
for event in app.stream({"messages": [input_message]}, config, stream_mode="values"):
    print(event["messages"][-1].pretty_print())

# View state history
print("--- 状态历史记录 ---" + "\n\n\n")
all_states = []
state_history = app.get_state_history(config)
for state in state_history:
    all_states.append(state)
    print(state)

# 从某次状态下重复
replay_state = all_states[2]  # Replay right before tool execution
print("重复state" + " " + str(replay_state) + "\n\n\n")
print("--- Replayed State ---" + "\n\n\n")
for event in app.stream(None, replay_state.config):
    for v in event.values():
        print(v)
        print("\n\n\n")
print("--- Branching off Past States---")
# Get the last message with the tool call
last_message = replay_state.values["messages"][-1]

# Update the tool call from Apple Music to Spotify
# 修改消息
if last_message.tool_calls[0]["name"] == "play_song_on_spotify":
    last_message.tool_calls[0]["name"] = "play_song_on_apple"
else:
    last_message.tool_calls[0]["name"] = "play_song_on_spotify"

# Update the state and resume execution
branch_config = app.update_state(replay_state.config, {"messages": [last_message]})
for event in app.stream(None, branch_config):
    print(event)

运行结果

lua 复制代码 
================================ Human Message =================================

你可以播放Taylor Swift的最流行的歌曲吗?
None
================================== Ai Message ==================================
Tool Calls:
  play_song_on_spotify (call_1a95d1cd1be04fea9431c0)
 Call ID: call_1a95d1cd1be04fea9431c0
  Args:
    song: Taylor Swift - Shake It Off
None
================================= Tool Message =================================
Name: play_song_on_spotify

Successfully played Taylor Swift - Shake It Off on Spotify!
None
================================== Ai Message ==================================

成功在Spotify上播放了Taylor Swift的歌曲"Shake It Off"! 如果您想听其他歌曲，随时告诉我。
None
--- 状态历史记录 ---



StateSnapshot(values={'messages': [HumanMessage(content='你可以播放Taylor Swift的最流行的歌曲吗?', additional_kwargs={}, response_metadata={}, id='ae716e3a-c657-473d-83d5-8c17cbdc85d1'), AIMessage(content='', additional_kwargs={'tool_calls': [{'index': 0, 'id': 'call_1a95d1cd1be04fea9431c0', 'function': {'arguments': '{"song": "Taylor Swift - Shake It Off"}', 'name': 'play_song_on_spotify'}, 'type': 'function'}]}, response_metadata={'finish_reason': 'tool_calls', 'model_name': 'qwen-max'}, id='run--51e4d937-1b0d-4fcc-803e-c4b5a25e50c7-0', tool_calls=[{'name': 'play_song_on_spotify', 'args': {'song': 'Taylor Swift - Shake It Off'}, 'id': 'call_1a95d1cd1be04fea9431c0', 'type': 'tool_call'}]), ToolMessage(content='Successfully played Taylor Swift - Shake It Off on Spotify!', name='play_song_on_spotify', id='5eba048c-7a79-43e8-ae7d-99deddb2e4a6', tool_call_id='call_1a95d1cd1be04fea9431c0'), AIMessage(content='成功在Spotify上播放了Taylor Swift的歌曲"Shake It Off"! 如果您想听其他歌曲，随时告诉我。', additional_kwargs={}, response_metadata={'finish_reason': 'stop', 'model_name': 'qwen-max'}, id='run--7493b409-7ad8-461f-aaab-e37e68263d39-0')]}, next=(), config={'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1f0768f8-9957-66ce-8003-0ebfe9a4a043'}}, metadata={'source': 'loop', 'step': 3, 'parents': {}}, created_at='2025-08-11T08:45:30.916014+00:00', parent_config={'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1f0768f8-8761-6788-8002-fa90d17def27'}}, tasks=(), interrupts=())
StateSnapshot(values={'messages': [HumanMessage(content='你可以播放Taylor Swift的最流行的歌曲吗?', additional_kwargs={}, response_metadata={}, id='ae716e3a-c657-473d-83d5-8c17cbdc85d1'), AIMessage(content='', additional_kwargs={'tool_calls': [{'index': 0, 'id': 'call_1a95d1cd1be04fea9431c0', 'function': {'arguments': '{"song": "Taylor Swift - Shake It Off"}', 'name': 'play_song_on_spotify'}, 'type': 'function'}]}, response_metadata={'finish_reason': 'tool_calls', 'model_name': 'qwen-max'}, id='run--51e4d937-1b0d-4fcc-803e-c4b5a25e50c7-0', tool_calls=[{'name': 'play_song_on_spotify', 'args': {'song': 'Taylor Swift - Shake It Off'}, 'id': 'call_1a95d1cd1be04fea9431c0', 'type': 'tool_call'}]), ToolMessage(content='Successfully played Taylor Swift - Shake It Off on Spotify!', name='play_song_on_spotify', id='5eba048c-7a79-43e8-ae7d-99deddb2e4a6', tool_call_id='call_1a95d1cd1be04fea9431c0')]}, next=('agent',), config={'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1f0768f8-8761-6788-8002-fa90d17def27'}}, metadata={'source': 'loop', 'step': 2, 'parents': {}}, created_at='2025-08-11T08:45:29.032692+00:00', parent_config={'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1f0768f8-875c-67e9-8001-66cc0f5b12be'}}, tasks=(PregelTask(id='4684d9cb-6fba-093d-ce28-bc57d15a99f8', name='agent', path=('__pregel_pull', 'agent'), error=None, interrupts=(), state=None, result={'messages': [AIMessage(content='成功在Spotify上播放了Taylor Swift的歌曲"Shake It Off"! 如果您想听其他歌曲，随时告诉我。', additional_kwargs={}, response_metadata={'finish_reason': 'stop', 'model_name': 'qwen-max'}, id='run--7493b409-7ad8-461f-aaab-e37e68263d39-0')]}),), interrupts=())
StateSnapshot(values={'messages': [HumanMessage(content='你可以播放Taylor Swift的最流行的歌曲吗?', additional_kwargs={}, response_metadata={}, id='ae716e3a-c657-473d-83d5-8c17cbdc85d1'), AIMessage(content='', additional_kwargs={'tool_calls': [{'index': 0, 'id': 'call_1a95d1cd1be04fea9431c0', 'function': {'arguments': '{"song": "Taylor Swift - Shake It Off"}', 'name': 'play_song_on_spotify'}, 'type': 'function'}]}, response_metadata={'finish_reason': 'tool_calls', 'model_name': 'qwen-max'}, id='run--51e4d937-1b0d-4fcc-803e-c4b5a25e50c7-0', tool_calls=[{'name': 'play_song_on_spotify', 'args': {'song': 'Taylor Swift - Shake It Off'}, 'id': 'call_1a95d1cd1be04fea9431c0', 'type': 'tool_call'}])]}, next=('action',), config={'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1f0768f8-875c-67e9-8001-66cc0f5b12be'}}, metadata={'source': 'loop', 'step': 1, 'parents': {}}, created_at='2025-08-11T08:45:29.030653+00:00', parent_config={'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1f0768f8-786f-6c85-8000-d58fce7ad3a7'}}, tasks=(PregelTask(id='a5624aa2-7d3d-fb1d-adb4-ad33f403169f', name='action', path=('__pregel_pull', 'action'), error=None, interrupts=(), state=None, result={'messages': [ToolMessage(content='Successfully played Taylor Swift - Shake It Off on Spotify!', name='play_song_on_spotify', id='5eba048c-7a79-43e8-ae7d-99deddb2e4a6', tool_call_id='call_1a95d1cd1be04fea9431c0')]}),), interrupts=())
StateSnapshot(values={'messages': [HumanMessage(content='你可以播放Taylor Swift的最流行的歌曲吗?', additional_kwargs={}, response_metadata={}, id='ae716e3a-c657-473d-83d5-8c17cbdc85d1')]}, next=('agent',), config={'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1f0768f8-786f-6c85-8000-d58fce7ad3a7'}}, metadata={'source': 'loop', 'step': 0, 'parents': {}}, created_at='2025-08-11T08:45:27.465690+00:00', parent_config={'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1f0768f8-786d-658d-bfff-47e0e417bba6'}}, tasks=(PregelTask(id='0ab5a7f4-7f15-f001-64dd-d2291b5d9c0a', name='agent', path=('__pregel_pull', 'agent'), error=None, interrupts=(), state=None, result={'messages': [AIMessage(content='', additional_kwargs={'tool_calls': [{'index': 0, 'id': 'call_1a95d1cd1be04fea9431c0', 'function': {'arguments': '{"song": "Taylor Swift - Shake It Off"}', 'name': 'play_song_on_spotify'}, 'type': 'function'}]}, response_metadata={'finish_reason': 'tool_calls', 'model_name': 'qwen-max'}, id='run--51e4d937-1b0d-4fcc-803e-c4b5a25e50c7-0', tool_calls=[{'name': 'play_song_on_spotify', 'args': {'song': 'Taylor Swift - Shake It Off'}, 'id': 'call_1a95d1cd1be04fea9431c0', 'type': 'tool_call'}])]}),), interrupts=())
StateSnapshot(values={'messages': []}, next=('__start__',), config={'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1f0768f8-786d-658d-bfff-47e0e417bba6'}}, metadata={'source': 'input', 'step': -1, 'parents': {}}, created_at='2025-08-11T08:45:27.464692+00:00', parent_config=None, tasks=(PregelTask(id='8c92498d-a5e1-bdcf-cde3-b5abcd2cfd5c', name='__start__', path=('__pregel_pull', '__start__'), error=None, interrupts=(), state=None, result={'messages': [HumanMessage(content='你可以播放Taylor Swift的最流行的歌曲吗?', additional_kwargs={}, response_metadata={}, id='ae716e3a-c657-473d-83d5-8c17cbdc85d1')]}),), interrupts=())
重复state StateSnapshot(values={'messages': [HumanMessage(content='你可以播放Taylor Swift的最流行的歌曲吗?', additional_kwargs={}, response_metadata={}, id='ae716e3a-c657-473d-83d5-8c17cbdc85d1'), AIMessage(content='', additional_kwargs={'tool_calls': [{'index': 0, 'id': 'call_1a95d1cd1be04fea9431c0', 'function': {'arguments': '{"song": "Taylor Swift - Shake It Off"}', 'name': 'play_song_on_spotify'}, 'type': 'function'}]}, response_metadata={'finish_reason': 'tool_calls', 'model_name': 'qwen-max'}, id='run--51e4d937-1b0d-4fcc-803e-c4b5a25e50c7-0', tool_calls=[{'name': 'play_song_on_spotify', 'args': {'song': 'Taylor Swift - Shake It Off'}, 'id': 'call_1a95d1cd1be04fea9431c0', 'type': 'tool_call'}])]}, next=('action',), config={'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1f0768f8-875c-67e9-8001-66cc0f5b12be'}}, metadata={'source': 'loop', 'step': 1, 'parents': {}}, created_at='2025-08-11T08:45:29.030653+00:00', parent_config={'configurable': {'thread_id': '1', 'checkpoint_ns': '', 'checkpoint_id': '1f0768f8-786f-6c85-8000-d58fce7ad3a7'}}, tasks=(PregelTask(id='a5624aa2-7d3d-fb1d-adb4-ad33f403169f', name='action', path=('__pregel_pull', 'action'), error=None, interrupts=(), state=None, result={'messages': [ToolMessage(content='Successfully played Taylor Swift - Shake It Off on Spotify!', name='play_song_on_spotify', id='5eba048c-7a79-43e8-ae7d-99deddb2e4a6', tool_call_id='call_1a95d1cd1be04fea9431c0')]}),), interrupts=())



--- Replayed State ---



{'messages': [ToolMessage(content='Successfully played Taylor Swift - Shake It Off on Spotify!', name='play_song_on_spotify', id='b9f06d4d-fe20-477a-8c4c-7b6bf20e4ae5', tool_call_id='call_1a95d1cd1be04fea9431c0')]}




{'messages': [AIMessage(content='成功在Spotify上播放了Taylor Swift的歌曲"Shake It Off"！如果你想要听其他的歌曲，随时告诉我。', additional_kwargs={}, response_metadata={'finish_reason': 'stop', 'model_name': 'qwen-max'}, id='run--327816ee-b742-4c65-9931-da751f137001-0')]}




--- Branching off Past States---
{'action': {'messages': [ToolMessage(content='Successfully played Taylor Swift - Shake It Off on Apple Music!', name='play_song_on_apple', id='cc5ee933-d8ee-44b7-9004-869ea8d70b68', tool_call_id='call_1a95d1cd1be04fea9431c0')]}}
{'agent': {'messages': [AIMessage(content='成功在Apple Music上播放了Taylor Swift的歌曲"Shake It Off"！ 如果您想听其他的歌曲或者需要任何帮助，随时告诉我哦。', additional_kwargs={}, response_metadata={'finish_reason': 'stop', 'model_name': 'qwen-max'}, id='run--efc74dc2-4dfe-4ff8-b3ea-485c320e93c5-0')]}}
相关推荐
2501_9381400717 分钟前
做内容很省心，2026好用AI笔记，支持抖音小红书视频无水印保存
人工智能
sell-cloud3 小时前
Prompt Engineering要凉了？Fable 5.5能“猜“意图，但银行正在封杀AI Agent——企业AI的真正护城河是什么
人工智能·prompt
Navicat中国5 小时前
Navicat 18 内置 AI 助手让数据库开发从繁琐操作迈向智能协作
大数据·人工智能·数据库开发
4SAPI5 小时前
多模型请求路由怎么设计？成本、速度和能力如何平衡
人工智能·python·gateway·php
魔乐社区5 小时前
翼芯・智算测试与适配优化专区，让国产算力“选得准、跑得通、用得好“
人工智能·开源·大模型
XCDoris7 小时前
郴州logo设计机构
大数据·人工智能·python
Goboy7 小时前
一个人带三个Agent,我用全新TRAE从0到1搭了个资讯周刊站
人工智能·trae
俊哥V7 小时前
每日 AI 研究简报 · 2026-10-19
人工智能·ai
AI_AGENT_DEV_AI7 小时前
AI英语教培智能体的开发
人工智能
物联网软硬件开发-轨物科技7 小时前
【技术白皮书】光伏清检一体化机器人系统解决方案技术白皮书
大数据·运维·网络·人工智能·机器人
热门推荐
01国内可直接用、免费额度/永久免费的大模型API清单（含 SiliconFlow、火山、阿里、智谱、百度、Kimi、DeepSeek、DMXAPI 等）02如何新建文件夹？ 电脑新建文件夹的4种方法03历年考研数学一、数学二、数学三真题试卷及答案PDF04GitHub 镜像站点05几个好用的ip纯净度检测网站06「全语录+省流版」邵艾伦 × 孙宇晨：聊风口、聊个人IP、聊年轻人如何抓住 AI 时代的机会07苹果iPA游戏软件资源下载网站08Windows 10/11 如何退出微软账户并改用本地账户登录（含找不到选项的解决方法）09蜘蛛磁力 搜索引擎大全，如何使用蜘蛛磁力查找磁力链接10昇腾910B系列芯片完全指南：四款型号对比、服务器选型与大模型部署实战