LangSmith 与 LangGraph（Python 和 JS）顺利集成，帮助您跟踪智能体，无论您使用 LangChain 模块还是其他 SDK。

​
使用 LangChain

如果您在 LangGraph 中使用 LangChain 模块，您只需设置几个环境变量即可启用跟踪。 本指南将介绍一个基本示例。有关配置的更多详细信息，请参阅使用 LangChain 跟踪指南。

​
1. 安装

为 Python 和 JS 安装 LangGraph 库和 OpenAI 集成（我们在下面的代码片段中使用 OpenAI 集成）。 有关可用包的完整列表，请参阅 LangChain Python 文档和 LangChain JS 文档。 
pip install langchain_openai langgraph

​
2. 配置您的环境

export LANGSMITH_TRACING=true
export LANGSMITH_API_KEY=<your-api-key>
# 此示例使用 OpenAI，但您可以使用您选择的任何 LLM 提供商
export OPENAI_API_KEY=<your-openai-api-key>
# 对于链接到多个工作区的 LangSmith API 密钥，设置 LANGSMITH_WORKSPACE_ID 环境变量以指定要使用的工作区。
export LANGSMITH_WORKSPACE_ID=<your-workspace-id>
如果您在非无服务器环境中使用 LangChain.js 和 LangSmith，我们还建议显式设置以下内容以减少延迟：export LANGCHAIN_CALLBACKS_BACKGROUND=true如果您在无服务器环境中，我们建议设置相反的设置以允许跟踪在函数结束之前完成：export LANGCHAIN_CALLBACKS_BACKGROUND=false有关更多信息，请参阅此 LangChain.js 指南。

​
3. 记录跟踪

一旦您设置了环境，您就可以像往常一样调用 LangChain 可运行项。LangSmith 将推断正确的跟踪配置： 
from typing import Literal
from langchain.messages import HumanMessage
from langchain_openai import ChatOpenAI
from langchain.tools import tool
from langgraph.prebuilt import ToolNode
from langgraph.graph import StateGraph, MessagesState

@tool
def search(query: str):
    """Call to surf the web."""
    if "sf" in query.lower() or "san francisco" in query.lower():
        return "It's 60 degrees and foggy."
    return "It's 90 degrees and sunny."

tools = [search]
tool_node = ToolNode(tools)

model = ChatOpenAI(model="gpt-4o", temperature=0).bind_tools(tools)

def should_continue(state: MessagesState) -> Literal["tools", "__end__"]:
    messages = state['messages']
    last_message = messages[-1]
    if last_message.tool_calls:
        return "tools"
    return "__end__"

def call_model(state: MessagesState):
    messages = state['messages']
    # Invoking `model` will automatically infer the correct tracing context
    response = model.invoke(messages)
    return {"messages": [response]}

workflow = StateGraph(MessagesState)
workflow.add_node("agent", call_model)
workflow.add_node("tools", tool_node)
workflow.add_edge("__start__", "agent")
workflow.add_conditional_edges(
    "agent",
    should_continue,
)
workflow.add_edge("tools", 'agent')

app = workflow.compile()

final_state = app.invoke(
    {"messages": [HumanMessage(content="what is the weather in sf")]},
    config={"configurable": {"thread_id": 42}}
)

final_state["messages"][-1].content
An example trace from running the above code looks like this: Trace tree for a LangGraph run with LangChain

​
Without LangChain

If you are using other SDKs or custom functions within LangGraph, you will need to wrap or decorate them appropriately (with the @traceable decorator in Python or the traceable function in JS, or something like e.g. wrap_openai for SDKs). If you do so, LangSmith will automatically nest traces from those wrapped methods. Here’s an example. You can also see this page for more information.

​
1. Installation

Install the LangGraph library and the OpenAI SDK for Python and JS (we use the OpenAI integration for the code snippets below). 
pip install openai langsmith langgraph

​
2. Configure your environment

export LANGSMITH_TRACING=true
export LANGSMITH_API_KEY=<your-api-key>
# This example uses OpenAI, but you can use any LLM provider of choice
export OPENAI_API_KEY=<your-openai-api-key>
If you are using LangChain.js with LangSmith and are not in a serverless environment, we also recommend setting the following explicitly to reduce latency:export LANGCHAIN_CALLBACKS_BACKGROUND=trueIf you are in a serverless environment, we recommend setting the reverse to allow tracing to finish before your function ends:export LANGCHAIN_CALLBACKS_BACKGROUND=falseSee this LangChain.js guide for more information.

​
3. Log a trace

Once you’ve set up your environment, wrap or decorate the custom functions/SDKs you want to trace. LangSmith will then infer the proper tracing config: 
import json
import openai
import operator
from langsmith import traceable
from langsmith.wrappers import wrap_openai
from typing import Annotated, Literal, TypedDict
from langgraph.graph import StateGraph

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

tool_schema = {
    "type": "function",
    "function": {
        "name": "search",
        "description": "Call to surf the web.",
        "parameters": {
            "type": "object",
            "properties": {"query": {"type": "string"}},
            "required": ["query"],
        },
    },
}

# Decorating the tool function will automatically trace it with the correct context
@traceable(run_type="tool", name="Search Tool")
def search(query: str):
    """Call to surf the web."""
    if "sf" in query.lower() or "san francisco" in query.lower():
        return "It's 60 degrees and foggy."
    return "It's 90 degrees and sunny."

tools = [search]

def call_tools(state):
    function_name_to_function = {"search": search}
    messages = state["messages"]
    tool_call = messages[-1]["tool_calls"][0]
    function_name = tool_call["function"]["name"]
    function_arguments = tool_call["function"]["arguments"]
    arguments = json.loads(function_arguments)
    function_response = function_name_to_function[function_name](**arguments)
    tool_message = {
        "tool_call_id": tool_call["id"],
        "role": "tool",
        "name": function_name,
        "content": function_response,
    }
    return {"messages": [tool_message]}

wrapped_client = wrap_openai(openai.Client())

def should_continue(state: State) -> Literal["tools", "__end__"]:
    messages = state["messages"]
    last_message = messages[-1]
    if last_message["tool_calls"]:
        return "tools"
    return "__end__"

def call_model(state: State):
    messages = state["messages"]
    # Calling the wrapped client will automatically infer the correct tracing context
    response = wrapped_client.chat.completions.create(
        messages=messages, model="gpt-4o-mini", tools=[tool_schema]
    )
    raw_tool_calls = response.choices[0].message.tool_calls
    tool_calls = [tool_call.to_dict() for tool_call in raw_tool_calls] if raw_tool_calls else []
    response_message = {
        "role": "assistant",
        "content": response.choices[0].message.content,
        "tool_calls": tool_calls,
    }
    return {"messages": [response_message]}

workflow = StateGraph(State)
workflow.add_node("agent", call_model)
workflow.add_node("tools", call_tools)
workflow.add_edge("__start__", "agent")
workflow.add_conditional_edges(
    "agent",
    should_continue,
)
workflow.add_edge("tools", 'agent')

app = workflow.compile()

final_state = app.invoke(
    {"messages": [{"role": "user", "content": "what is the weather in sf"}]}
)

final_state["messages"][-1]["content"]
An example trace from running the above code looks like this: Trace tree for a LangGraph run without LangChain
Edit the source of this page on GitHub.
Connect these docs programmatically to Claude, VSCode, and more via MCP for real-time answers.