如何在LangGraph工具内部访问LangGraph状态

问题背景 #

我有一个调用多工具的LangGraph代理，需要在特定工具（如taxes_states）内部获取LangGraph的状态数据，但官方文档仅说明如何在节点内访问状态，未提及工具内的实现方式。当前核心代码和工具代码如下：

代理核心代码（thread.py） #

# thread.py
# Create chatbot node
tools = [taxes_federal_internal_revenue_code, taxes_federal_court_cases, taxes_federal_treasury_regulations, taxes_international, taxes_federal_forms, taxes_states]
llm = ChatOpenAI(model='gpt-4o')
llm_with_tools = llm.bind_tools(tools)
def chatbot(state: MessagesState):
    toolResp = llm_with_tools.invoke(state["messages"])
    return MessagesState(
        messages=[toolResp], 
        log_stream_name=state["log_stream_name"],  # Corrected access
        next_step=state["next_step"],  # Ensure next_step is passed
)
graph_builder.add_node("chatbot", chatbot)

taxes_states工具代码 #

@tool()
def taxes_states(query: str) -> Tuple[List[str]]:
"""
Purpose: Retrieve relevant U.S. state tax law documents, including specific state tax codes, 
regulations, and guidance (e.g., California tax laws)
"""

print("-----------------taxes_states-----------")
print(query)
print("-----------------taxes_states-----------")

# Perform RAG here and get retrieved_docs and return it
return retrieved_docs

解决方案 #

LangGraph的工具本身无法直接访问状态，需通过显式传递参数的方式将状态数据传入工具，具体步骤如下：

1. 修改工具定义，添加状态参数 #

更新taxes_states工具，新增state参数并完善文档说明：

from langchain_core.tools import tool
from typing import Tuple, List, Dict

@tool()
def taxes_states(query: str, state: Dict) -> Tuple[List[str]]:
    """
    Purpose: Retrieve relevant U.S. state tax law documentsgate have七 specific path###"享esc 更快的account辩解""ore specific state tax codes, 
    regulations, and guidance (e.g., California tax laws)
    
    Args:
        query: 用户的查询问题
        state: LangGraph的状态数据，包含log_stream_name等字段
    """
    print("-----------------taxes_states-----------")
    print(query)
    # 访问状态中的数据
    print("当前log_stream_name:", state["log_stream_name"])
    print("-----------------taxes_states-----------")

    # 结合状态数据执行RAG逻辑
    # retrieved_docs = ...
    retrieved_docs = []
    return retrieved_docs

2. 在chatbot节点中传递状态参数 #

调整节点逻辑，在LLM生成工具调用后，为taxes_states工具手动添加状态参数：

# thread.py
# Create chatbot node
tools = [taxes_federal_internal_revenue_code, taxes_federal_court_cases, taxes_federal_treasury_regulations, taxes_international, taxes_federal_forms, taxes_states]
llm = ChatOpenAI(model='gpt-4o')
llm_with_tools = llm.bind_tools(tools)

def chatbot(state: MessagesState):
    toolResp = llm_with_tools.invoke(state["messages"])
    # 遍历工具调用，给指定工具注入状态参数
    if hasattr(toolResp, 'tool_calls'):
        for tool_call in toolResp.tool_calls:
            if tool_call["name"] == "taxes_states":
                # 将Pydantic状态转为字典传入
                tool_call["args"]["state"] = state.dict()
    return MessagesState(
        messages=[toolResp], 
        log_stream_name=state["log_stream_name"],
        next_step=state["next_step"],
    )

graph_builder.add_node("chatbot", chatbot)

3. 确保工具执行节点支持参数传递 #

如果使用LangGraph默认的tool_executor，无需额外修改；如果是自定义工具执行逻辑，需保证能将state参数正确传递给工具函数。

注意：若MessagesState是Pydantic模型，需用.dict()方法转为字典传递，避免序列化问题；若状态包含敏感数据，需先过滤再传入工具。

内容的提问来源于stack exchange，提问作者Kiran