Skip to main content

Trace with LangChain (Python and JS/TS)

LangSmith integrates seamlessly with LangChain (Python and JS), the popular open-source framework for building LLM applications.

Installation​

Install the core library and the OpenAI integration for Python and JS (we use the OpenAI integration for the code snippets below).

For a full list of packages available, see the LangChain Python docs and LangChain JS docs.

pip install langchain_openai langchain_core

Quick start​

1. Configure your environment​

export LANGCHAIN_TRACING_V2=true
export LANGCHAIN_API_KEY=<your-api-key>
# The below examples use the OpenAI API, though it's not necessary in general
export OPENAI_API_KEY=<your-openai-api-key>

2. Log a trace​

No extra code is needed to log a trace to LangSmith. Just run your LangChain code as you normally would.

from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser

prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a helpful assistant. Please respond to the user's request only based on the given context."),
    ("user", "Question: {question}\nContext: {context}")
])
model = ChatOpenAI(model="gpt-3.5-turbo")
output_parser = StrOutputParser()

chain = prompt | model | output_parser

question = "Can you summarize this morning's meetings?"
context = "During this morning's meeting, we solved all world conflict."
chain.invoke({"question": question, "context": context})

3. View your trace​

By default, the trace will be logged to the project with the name default. An example of a trace logged using the above code is made public and can be viewed here.

Trace selectively​

The previous section showed how to trace all invocations of a LangChain runnables within your applications by setting a single environment variable. While this is a convenient way to get started, you may want to trace only specific invocations or parts of your application.

There are two ways to do this in Python: by manually passing in a LangChainTracer (reference docs) instance as a callback, or by using the tracing_v2_enabled context manager (reference docs).

In JS/TS, you can pass a LangChainTracer (reference docs) instance as a callback.

# You can configure a LangChainTracer instance to trace a specific invocation.
from langchain.callbacks.tracers import LangChainTracer

tracer = LangChainTracer()
chain.invoke({"question": "Am I using a callback?", "context": "I'm using a callback"}, config={"callbacks": [tracer]})

# LangChain Python also supports a context manager for tracing a specific block of code.
from langchain_core.tracers.context import tracing_v2_enabled
with tracing_v2_enabled():
    chain.invoke({"question": "Am I using a context manager?", "context": "I'm using a context manager"})

# This will NOT be traced (assuming LANGCHAIN_TRACING_V2 is not set)
chain.invoke({"question": "Am I being traced?", "context": "I'm not being traced"})

Log to a specific project​

Statically​

As mentioned in the tracing conceptual guide LangSmith uses the concept of a Project to group traces. If left unspecified, the tracer project is set to default. You can set the LANGCHAIN_PROJECT environment variable to configure a custom project name for an entire application run. This should be done before executing your application.

export LANGCHAIN_PROJECT=my-project

Dynamically​

This largely builds off of the previous section and allows you to set the project name for a specific LangChainTracer instance or as parameters to the tracing_v2_enabled context manager in Python.

# You can set the project name for a specific tracer instance:
from langchain.callbacks.tracers import LangChainTracer

tracer = LangChainTracer(project_name="My Project")
chain.invoke({"question": "Am I using a callback?", "context": "I'm using a callback"}, config={"callbacks": [tracer]})

# You can set the project name using the project_name parameter.
from langchain_core.tracers.context import tracing_v2_enabled
with tracing_v2_enabled(project_name="My Project"):
    chain.invoke({"question": "Am I using a context manager?", "context": "I'm using a context manager"})

Add metadata and tags to traces​

You can send annotate your traces with arbitrary metadata and tags by providing them in the Config. This is useful for associating additional information with a trace, such as the environment in which it was executed, or the user who initiated it. For information on how to query traces and runs by metadata and tags, see this guide

note

When you attach metadata or tags to a runnable (either through the RunnableConfig or at runtime with invocation params), they are inherited by all child runnables of that runnable.

from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser

prompt = ChatPromptTemplate.from_messages([
  ("system", "You are a helpful AI."),
  ("user", "{input}")
])
# The tag "model-tag" and metadata {"model-key": "model-value"} will be attached to the ChatOpenAI run only
chat_model = ChatOpenAI().with_config({"tags": ["model-tag"], "metadata": {"model-key": "model-value"}})
output_parser = StrOutputParser()

# Tags and metadata can be configured with RunnableConfig
chain = (prompt | chat_model | output_parser).with_config({"tags": ["config-tag"], "metadata": {"config-key": "config-value"}})

# Tags and metadata can also be passed at runtime
chain.invoke({"input": "What is the meaning of life?"}, {"tags": ["invoke-tag"], "metadata": {"invoke-key": "invoke-value"}})

Customize run name​

You can customize the name of a given run when invoking or streaming your LangChain code by providing it in the Config. This name is used to identify the run in LangSmith and can be used to filter and group runs. The name is also used as the title of the run in the LangSmith UI. This can be done by setting a run_name in the RunnableConfig object at construction or by passing a run_name in the invocation parameters in JS/TS.

note

This feature is not currently supported directly for LLM objects.

# When tracing within LangChain, run names default to the class name of the traced object (e.g., 'ChatOpenAI').
configured_chain = chain.with_config({"run_name": "MyCustomChain"})
configured_chain.invoke({"input": "What is the meaning of life?"})

# You can also configure the run name at invocation time, like below
chain.invoke({"input": "What is the meaning of life?"}, {"run_name": "MyCustomChain"})

Customize run ID​

You can customize the ID of a given run when invoking or streaming your LangChain code by providing it in the Config. This ID is used to uniquely identify the run in LangSmith and can be used to query specific runs. The ID can be useful for linking runs across different systems or for implementing custom tracking logic. This can be done by setting a run_id in the RunnableConfig object at construction or by passing a run_id in the invocation parameters in JS/TS.

note

This feature is not currently supported directly for LLM objects.

import uuid

my_uuid = uuid.uuid4()
# You can configure the run ID at invocation time:
chain.invoke({"input": "What is the meaning of life?"}, {"run_id": my_uuid})

Note that if you do this at the root of a trace (i.e., the top-level run, that run ID will be used as the trace_id).

Ensure all traces are submitted before exiting​

In LangChain Python, LangSmith's tracing is done in a background thread to avoid obstructing your production application. This means that your process may end before all traces are successfully posted to LangSmith. This is especially prevalent in a serverless environment, where your VM may be terminated immediately once your chain or agent completes.

In LangChain JS/TS, the default is to block for a short period of time for the trace to finish due to the greater popularity of serverless environments. You can make callbacks asynchronous by setting the LANGCHAIN_CALLBACKS_BACKGROUND environment variable to "true".

For both languages, LangChain exposes methods to wait for traces to be submitted before exiting your application. Below is an example:

from langchain_openai import ChatOpenAI
from langchain_core.tracers.langchain import wait_for_all_tracers

llm = ChatOpenAI()
try:
    llm.invoke("Hello, World!")
finally:
    wait_for_all_tracers()

Trace without setting environment variables​

As mentioned in other guides, the following environment variables allow you to configure tracing enabled, the api endpoint, the api key, and the tracing project:

  • LANGCHAIN_TRACING_V2
  • LANGCHAIN_API_KEY
  • LANGCHAIN_ENDPOINT
  • LANGCHAIN_PROJECT

However, in some environments, it is not possible to set environment variables. In these cases, you can set the tracing configuration programmatically.

This largely builds off of the previous section.

from langchain.callbacks.tracers import LangChainTracer
import langsmith as ls

# You can create a client instance with an api key and api url
client = ls.Client(
    api_key="YOUR_API_KEY",  # This can be retrieved from a secrets manager
    api_url="https://api.smith.langchain.com",  # Update appropriately for self-hosted installations or the EU region
)

# You can pass the client and project_name to the LangChainTracer instance
tracer = LangChainTracer(client=client, project_name="test-no-env")
chain.invoke({"question": "Am I using a callback?", "context": "I'm using a callback"}, config={"callbacks": [tracer]})

# LangSmith python also supports a context manager which allows passing the client and project_name
import langsmith as ls
with ls.tracing_context(client=client, project_name="test-no-env", enabled=True):
    chain.invoke({"question": "Am I using a context manager?", "context": "I'm using a context manager"})

Distributed tracing with LangChain (Python)​

LangSmith supports distributed tracing with LangChain Python. This allows you to link runs (spans) across different services and applications. The principles are similar to the distributed tracing guide for the LangSmith SDK.

import langsmith
from langchain_core.runnables import chain
from langsmith.run_helpers import get_current_run_tree

# -- This code should be in a separate file or service --
@chain
def child_chain(inputs):
    return inputs["test"] + 1

def child_wrapper(x, headers):
    with langsmith.tracing_context(parent=headers):
        child_chain.invoke({"test": x})

# -- This code should be in a separate file or service --
@chain
def parent_chain(inputs):

    rt = get_current_run_tree()
    headers = rt.to_headers()
    # ... make a request to another service with the headers
    # The headers should be passed to the other service, eventually to the child_wrapper function

parent_chain.invoke({"test": 1})

Interoperability between LangChain (Python) and LangSmith SDK​

If you are using LangChain for part of your application and the LangSmith SDK (see this guide) for other parts, you can still trace the entire application seamlessly.

LangChain objects will be traced when invoked within a traceable function and be bound as a child run of the traceable function.

from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.output_parsers import StrOutputParser

from langsmith import traceable

prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a helpful assistant. Please respond to the user's request only based on the given context."),
    ("user", "Question: {question}\nContext: {context}")
])
model = ChatOpenAI(model="gpt-3.5-turbo")
output_parser = StrOutputParser()

chain = prompt | model | output_parser

# The above chain will be traced as a child run of the traceable function
@traceable(
    tags=["openai", "chat"],
    metadata={"foo": "bar"}
)
def invoke_runnnable(question, context):
    result = chain.invoke({"question": question, "context": context})
    return "The response is: " + result

invoke_runnnable("Can you summarize this morning's meetings?", "During this morning's meeting, we solved all world conflict.")

This will produce the following trace tree:

Interoperability between LangChain.JS and LangSmith SDK​

Tracing LangChain objects inside traceable (JS only)​

Starting with langchain@0.2.x, LangChain objects are traced automatically when used inside @traceable functions, inheriting the client, tags, metadata and project name of the traceable function.

For older versions of LangChain below 0.2.x, you will need to manually pass an instance LangChainTracer created from the tracing context found in @traceable.

import { ChatOpenAI } from "@langchain/openai";
import { ChatPromptTemplate } from "@langchain/core/prompts";
import { StringOutputParser } from "@langchain/core/output_parsers";
import { getLangchainCallbacks } from "langsmith/langchain";

const prompt = ChatPromptTemplate.fromMessages([
  [
    "system",
    "You are a helpful assistant. Please respond to the user's request only based on the given context.",
  ],
  ["user", "Question: {question}\nContext: {context}"],
]);
const model = new ChatOpenAI({ modelName: "gpt-3.5-turbo" });
const outputParser = new StringOutputParser();

const chain = prompt.pipe(model).pipe(outputParser);

const main = traceable(
  async (input: { question: string; context: string }) => {
    const callbacks = await getLangchainCallbacks();
    const response = await chain.invoke(input, { callbacks });
    return response;
  },
  { name: "main" }
);

Tracing LangChain child runs via traceable / RunTree API (JS only)​

note

We're working on improving the interoperability between traceable and LangChain. The following limitations are present when using combining LangChain with traceable:

  1. Mutating RunTree obtained from getCurrentRunTree() of the RunnableLambda context will result in a no-op.
  2. It's discouraged to traverse the RunTree obtained from RunnableLambda via getCurrentRunTree() as it may not contain all the RunTree nodes.
  3. Different child runs may have the same execution_order and child_execution_order value. Thus in extreme circumstances, some runs may end up in a different order, depending on the start_time.

In some uses cases, you might want to run traceable functions as part of the RunnableSequence or trace child runs of LangChain run imperatively via the RunTree API. Starting with LangSmith 0.1.39 and @langchain/core 0.2.18, you can directly invoke traceable-wrapped functions within RunnableLambda.

import { traceable } from "langsmith/traceable";
import { RunnableLambda } from "@langchain/core/runnables";
import { RunnableConfig } from "@langchain/core/runnables";

const tracedChild = traceable((input: string) => `Child Run: ${input}`, {
  name: "Child Run",
});

const parrot = new RunnableLambda({
  func: async (input: { text: string }, config?: RunnableConfig) => {
    return await tracedChild(input.text);
  },
});

Trace Tree

Alternatively, you can convert LangChain's RunnableConfig to a equivalent RunTree object by using RunTree.fromRunnableConfig or pass the RunnableConfig as the first argument of traceable-wrapped function.

import { traceable } from "langsmith/traceable";
import { RunnableLambda } from "@langchain/core/runnables";
import { RunnableConfig } from "@langchain/core/runnables";

const tracedChild = traceable((input: string) => `Child Run: ${input}`, {
  name: "Child Run",
});

const parrot = new RunnableLambda({
  func: async (input: { text: string }, config?: RunnableConfig) => {
    // Pass the config to existing traceable function
    await tracedChild(config, input.text);
    return input.text;
  },
});

Was this page helpful?


You can leave detailed feedback on GitHub.