用LangGraph搭建智能体—AI Agents in LangGraph（四、持久化和流式输出）

1. 学习背景

一般在构建代理时，会经过较长时间的任务处理流程，对于这些任务，持久化和流式输出是非常重要的两个概念。
持久性：允许您在特定时间点保留代理的状态，这可以让您返回到该状态，并在将来的交互中恢复该状态。
流式传输：您可以发出正在进行的工作的信号列表，对于长时间运行的应用程序，您可以准确知道代理正在执行的操作。

2. 示例

2.1 准备环境

python 复制代码

from dotenv import load_dotenv

_ = load_dotenv()

python 复制代码

from langgraph.graph import StateGraph, END
from typing import TypedDict, Annotated
import operator
from langchain_core.messages import AnyMessage, SystemMessage, HumanMessage, ToolMessage
from langchain_openai import ChatOpenAI
from langchain_community.tools.tavily_search import TavilySearchResults

python 复制代码

# 初始化工具
tool = TavilySearchResults(max_results=2)

2.2 编写代理

python 复制代码

class AgentState(TypedDict):
    messages: Annotated[list[AnyMessage], operator.add]

为了处理持久化，在langgraph中添加了checkpoint，检查点基本上在每个节点之后和之间对状态进行检查

python 复制代码

# 准备持久化，使用SqliteSaver，内存级数据库，重启则丢失内容

from langgraph.checkpoint.sqlite import SqliteSaver

memory = SqliteSaver.from_conn_string(":memory:")

编写Agent，具体代码介绍见：https://blog.csdn.net/l8947943/article/details/140592234?spm=1001.2014.3001.5502

python 复制代码

class Agent:
	# 可以看到，初始化阶段提供了checkpointer
    def __init__(self, model, tools, checkpointer, system=""):
        self.system = system
        graph = StateGraph(AgentState)
        graph.add_node("llm", self.call_openai)
        graph.add_node("action", self.take_action)
        graph.add_conditional_edges("llm", self.exists_action, {True: "action", False: END})
        graph.add_edge("action", "llm")
        graph.set_entry_point("llm")
        self.graph = graph.compile(checkpointer=checkpointer) # 在图编译阶段，将checkpointer内置
        self.tools = {t.name: t for t in tools}
        self.model = model.bind_tools(tools)

    def call_openai(self, state: AgentState):
        messages = state['messages']
        if self.system:
            messages = [SystemMessage(content=self.system)] + messages
        message = self.model.invoke(messages)
        return {'messages': [message]}

    def exists_action(self, state: AgentState):
        result = state['messages'][-1]
        return len(result.tool_calls) > 0

    def take_action(self, state: AgentState):
        tool_calls = state['messages'][-1].tool_calls
        results = []
        for t in tool_calls:
            print(f"Calling: {t}")
            result = self.tools[t['name']].invoke(t['args'])
            results.append(ToolMessage(tool_call_id=t['id'], name=t['name'], content=str(result)))
        print("Back to the model!")
        return {'messages': results}

2.3 构建prompt，初始化Agent

python 复制代码

prompt = """You are a smart research assistant. Use the search engine to look up information. \
You are allowed to make multiple calls (either together or in sequence). \
Only look up information when you are sure of what you want. \
If you need to look up some information before asking a follow up question, you are allowed to do that!
"""
model = ChatOpenAI(model="gpt-4o")
abot = Agent(model, [tool], system=prompt, checkpointer=memory) # 注意初始化存储

python 复制代码

# 构建HumanMessage
messages = [HumanMessage(content="What is the weather in sf?")]
# 设置线程字典，其实就是设置个id，用于执行过程的内容区分，可以理解为令牌
# 线程非常重要，特别是多个用户进行操作，可以用作隔离的标志
thread = {"configurable": {"thread_id": "1"}}

python 复制代码

# 对graph的内容进行
for event in abot.graph.stream({"messages": messages}, thread):
    for v in event.values():
        print(v['messages'])

输出如下：

[AIMessage(content='', additional_kwargs={'tool_calls': [{'id': 'call_aqUDZtKH5HolN9ZVPANsz4ep', 'function': {'arguments': '{"query":"current weather in San Francisco"}', 'name': 'tavily_search_results_json'}, 'type': 'function'}]}, response_metadata={'token_usage': {'completion_tokens': 22, 'prompt_tokens': 151, 'total_tokens': 173}, 'model_name': 'gpt-4o', 'system_fingerprint': 'fp_c4e5b6fa31', 'finish_reason': 'tool_calls', 'logprobs': None}, id='run-7c06cd16-f52a-4d4a-95b5-868dfc91456d-0', tool_calls=[{'name': 'tavily_search_results_json', 'args': {'query': 'current weather in San Francisco'}, 'id': 'call_aqUDZtKH5HolN9ZVPANsz4ep'}])]
Calling: {'name': 'tavily_search_results_json', 'args': {'query': 'current weather in San Francisco'}, 'id': 'call_aqUDZtKH5HolN9ZVPANsz4ep'}
Back to the model!
[ToolMessage(content='[{\'url\': \'https://www.weatherapi.com/\', \'content\': "{\'location\': {\'name\': \'San Francisco\', \'region\': \'California\', \'country\': \'United States of America\', \'lat\': 37.78, \'lon\': -122.42, \'tz_id\': \'America/Los_Angeles\', \'localtime_epoch\': 1721638208, \'localtime\': \'2024-07-22 1:50\'}, \'current\': {\'last_updated_epoch\': 1721637900, \'last_updated\': \'2024-07-22 01:45\', \'temp_c\': 13.0, \'temp_f\': 55.5, \'is_day\': 0, \'condition\': {\'text\': \'Partly Cloudy\', \'icon\': \'//cdn.weatherapi.com/weather/64x64/night/116.png\', \'code\': 1003}, \'wind_mph\': 7.2, \'wind_kph\': 11.5, \'wind_degree\': 204, \'wind_dir\': \'SSW\', \'pressure_mb\': 1011.0, \'pressure_in\': 29.86, \'precip_mm\': 0.0, \'precip_in\': 0.0, \'humidity\': 91, \'cloud\': 40, \'feelslike_c\': 12.1, \'feelslike_f\': 53.8, \'windchill_c\': 12.1, \'windchill_f\': 53.8, \'heatindex_c\': 13.0, \'heatindex_f\': 55.5, \'dewpoint_c\': 11.6, \'dewpoint_f\': 52.9, \'vis_km\': 10.0, \'vis_miles\': 6.0, \'uv\': 1.0, \'gust_mph\': 11.0, \'gust_kph\': 17.7}}"}, {\'url\': \'https://www.timeanddate.com/weather/usa/san-francisco/hourly\', \'content\': \'Hour-by-Hour Forecast for San Francisco, California, USA. Weather Today Weather Hourly 14 Day Forecast Yesterday/Past Weather Climate (Averages) Currently: 64 °F. Passing clouds. (Weather station: San Francisco International Airport, USA). See more current weather.\'}]', name='tavily_search_results_json', tool_call_id='call_aqUDZtKH5HolN9ZVPANsz4ep')]
[AIMessage(content='The current weather in San Francisco is partly cloudy with a temperature of 13°C (55.5°F). The wind is coming from the south-southwest at 7.2 mph (11.5 kph), and the humidity is at 91%. The visibility is 10 kilometers (6 miles), and the UV index is 1.', response_metadata={'token_usage': {'completion_tokens': 72, 'prompt_tokens': 675, 'total_tokens': 747}, 'model_name': 'gpt-4o', 'system_fingerprint': 'fp_c4e5b6fa31', 'finish_reason': 'stop', 'logprobs': None}, id='run-7f7a1179-4d54-412f-afdc-50db86262313-0')]

可以看到，执行过程中，先是AIMessage，其次是ToolMessage，最后是响应的AIMessage。通过这种流方法，我们可以得到中间结果。

2.4 继续提问

python 复制代码

messages = [HumanMessage(content="What about in la?")]
thread = {"configurable": {"thread_id": "1"}}
for event in abot.graph.stream({"messages": messages}, thread):
    for v in event.values():
        print(v)

输出如下：

{'messages': [AIMessage(content='', additional_kwargs={'tool_calls': [{'id': 'call_FLX2DJNqEGmsb5mgdxbfNOpJ', 'function': {'arguments': '{"query":"current weather in Los Angeles"}', 'name': 'tavily_search_results_json'}, 'type': 'function'}]}, response_metadata={'token_usage': {'completion_tokens': 22, 'prompt_tokens': 759, 'total_tokens': 781}, 'model_name': 'gpt-4o', 'system_fingerprint': 'fp_c4e5b6fa31', 'finish_reason': 'tool_calls', 'logprobs': None}, id='run-f2054e4e-984c-41dc-a55a-d017e1b52211-0', tool_calls=[{'name': 'tavily_search_results_json', 'args': {'query': 'current weather in Los Angeles'}, 'id': 'call_FLX2DJNqEGmsb5mgdxbfNOpJ'}])]}
Calling: {'name': 'tavily_search_results_json', 'args': {'query': 'current weather in Los Angeles'}, 'id': 'call_FLX2DJNqEGmsb5mgdxbfNOpJ'}
Back to the model!
{'messages': [ToolMessage(content="[{'url': 'https://www.accuweather.com/en/us/los-angeles/90012/current-weather/347625', 'content': 'Current weather in Los Angeles, CA. Check current conditions in Los Angeles, CA with radar, hourly, and more.'}, {'url': 'https://forecast.weather.gov/zipcity.php?inputstring=Los Angeles,CA', 'content': 'Los Angeles CA 34.05°N 118.25°W (Elev. 377 ft) Last Update: 9:52 pm PDT Jul 19, 2024. Forecast Valid: 2am PDT Jul 20, 2024-6pm PDT Jul 26, 2024 . ... Severe Weather ; Current Outlook Maps ; Drought ; Fire Weather ; Fronts/Precipitation Maps ; Current Graphical Forecast Maps ; Rivers ; Marine ; Offshore and High Seas; Hurricanes ;'}]", name='tavily_search_results_json', tool_call_id='call_FLX2DJNqEGmsb5mgdxbfNOpJ')]}
{'messages': [AIMessage(content='For the most current weather information in Los Angeles, you can check the following sources:\n\n1. [AccuWeather - Los Angeles](https://www.accuweather.com/en/us/los-angeles/90012/current-weather/347625): Provides current conditions, radar, and hourly updates.\n\n2. [NOAA - Los Angeles](https://forecast.weather.gov/zipcity.php?inputstring=Los Angeles,CA): Offers detailed forecasts, severe weather alerts, and more.\n\nBy visiting these links, you will get the most up-to-date weather information for Los Angeles.', response_metadata={'token_usage': {'completion_tokens': 117, 'prompt_tokens': 981, 'total_tokens': 1098}, 'model_name': 'gpt-4o', 'system_fingerprint': 'fp_400f27fa1f', 'finish_reason': 'stop', 'logprobs': None}, id='run-e64b680f-f97b-451c-b7f4-e09b0d0b2d51-0')]}

可以看到，也经历了AIMessage、ToolMessage、AIMessage的响应。在此回话的基础上，询问哪个温度较高。

python 复制代码

messages = [HumanMessage(content="Which one is warmer?")]
thread = {"configurable": {"thread_id": "1"}}
for event in abot.graph.stream({"messages": messages}, thread):
    for v in event.values():
        print(v)

输出如下：

{'messages': [AIMessage(content='Based on the provided data, San Francisco has a current temperature of 13°C (55.5°F). For Los Angeles, specific temperature data was not retrieved, but you can view the current conditions on the [AccuWeather - Los Angeles](https://www.accuweather.com/en/us/los-angeles/90012/current-weather/347625) page.\n\nHowever, it is generally known that Los Angeles tends to be warmer than San Francisco. For precise and up-to-date temperature comparisons, please check the provided links.', response_metadata={'token_usage': {'completion_tokens': 107, 'prompt_tokens': 1110, 'total_tokens': 1217}, 'model_name': 'gpt-4o', 'system_fingerprint': 'fp_c4e5b6fa31', 'finish_reason': 'stop', 'logprobs': None}, id='run-4827c13a-d08d-4d32-b29e-3226454193cd-0')]}

可以看到，对话过程中是保留着上下文信息的，这是因为保留着检查点的持久内容。我们换个线程，看看执行结果。

python 复制代码

messages = [HumanMessage(content="Which one is warmer?")]
thread = {"configurable": {"thread_id": "2"}}
for event in abot.graph.stream({"messages": messages}, thread):
    for v in event.values():
        print(v)

输出如下：

{'messages': [AIMessage(content="Could you please provide more context or specify the items or locations you're comparing in terms of warmth? This will help me give you a precise answer.", response_metadata={'token_usage': {'completion_tokens': 30, 'prompt_tokens': 149, 'total_tokens': 179}, 'model_name': 'gpt-4o', 'system_fingerprint': 'fp_c4e5b6fa31', 'finish_reason': 'stop', 'logprobs': None}, id='run-e5f989a4-01cf-4be7-8360-717609683ba4-0')]}

可以看到，这时候的Agent竟然不知所措，这是因为我们使用了不同的线程ID。

2.5 实时Streaming tokens

python 复制代码

from langgraph.checkpoint.aiosqlite import AsyncSqliteSaver

memory = AsyncSqliteSaver.from_conn_string(":memory:")
abot = Agent(model, [tool], system=prompt, checkpointer=memory)

python 复制代码

messages = [HumanMessage(content="What is the weather in SF?")]
thread = {"configurable": {"thread_id": "4"}}
async for event in abot.graph.astream_events({"messages": messages}, thread, version="v1"):
    kind = event["event"]
    if kind == "on_chat_model_stream":
        content = event["data"]["chunk"].content
        if content:
            # Empty content in the context of OpenAI means
            # that the model is asking for a tool to be invoked.
            # So we only print non-empty content
            print(content, end="|")

输出如下：

Calling: {'name': 'tavily_search_results_json', 'args': {'query': 'current weather in San Francisco'}, 'id': 'call_8dBYQ8xGAXysN9TqtVX44adf'}
Back to the model!
The| current| weather| in| San| Francisco| is| partly| cloudy| with| a| temperature| of| |13|°C| (|55|.|5|°F|).| The| wind| is| blowing| from| the| south|-s|outh|west| at| |7|.|2| mph| (|11|.|5| k|ph|),| and| the| humidity| is| at| |91|%.| The| visibility| is| |10| km| (|6| miles|),| and| the| UV| index| is| |1|.|

最终的结果是一个个被执行出来的，这是因为使用了异步持久化，这种事件核心获取是on_chat_model_stream。最大的特点是一个个将token输出出来。

3.总结

本节学习了持久化和流式输出，最后还有个异步流式输出，按部就班即可。