langsmith

LangSmith Client SDK

This package contains the TypeScript client for interacting with the LangSmith platform.

To install:

yarn add langsmith

LangSmith helps you and your team develop and evaluate language models and intelligent agents. It is compatible with any LLM Application and provides seamless integration with LangChain, a widely recognized open-source framework that simplifies the process for developers to create powerful language model applications.

Note: You can enjoy the benefits of LangSmith without using the LangChain open-source packages! To get started with your own proprietary framework, set up your account and then skip to Logging Traces Outside LangChain.

Cookbook: For tutorials on how to get more value out of LangSmith, check out the Langsmith Cookbook repo.

A typical workflow looks like:

1. Set up an account with LangSmith.
2. Log traces.
3. Debug, Create Datasets, and Evaluate Runs.

We'll walk through these steps in more detail below.

1. Connect to LangSmith

Sign up for LangSmith using your GitHub, Discord accounts, or an email address and password. If you sign up with an email, make sure to verify your email address before logging in.

Then, create a unique API key on the Settings Page.

Note: Save the API Key in a secure location. It will not be shown again.

2. Log Traces

You can log traces natively in your LangChain application or using a LangSmith RunTree.

Logging Traces with LangChain

LangSmith seamlessly integrates with the JavaScript LangChain library to record traces from your LLM applications.

yarn add langchain

1. Copy the environment variables from the Settings Page and add them to your application.

Tracing can be activated by setting the following environment variables or by manually specifying the LangChainTracer.

process.env["LANGCHAIN_TRACING_V2"] = "true";
process.env["LANGCHAIN_ENDPOINT"] = "https://api.smith.langchain.com";
process.env["LANGCHAIN_API_KEY"] = "<YOUR-LANGSMITH-API-KEY>";
// process.env["LANGCHAIN_PROJECT"] = "My Project Name"; // Optional: "default" is used if not set

Tip: Projects are groups of traces. All runs are logged to a project. If not specified, the project is set to default.

1. Run an Agent, Chain, or Language Model in LangChain

If the environment variables are correctly set, your application will automatically connect to the LangSmith platform.

import { ChatOpenAI } from "langchain/chat_models/openai";

const chat = new ChatOpenAI({ temperature: 0 });
const response = await chat.predict(
  "Translate this sentence from English to French. I love programming."
);
console.log(response);

Logging Traces Outside LangChain

You can still use the LangSmith development platform without depending on any LangChain code. You can connect either by setting the appropriate environment variables, or by directly specifying the connection information in the RunTree.

1. Copy the environment variables from the Settings Page and add them to your application.

export LANGCHAIN_API_KEY=<YOUR-LANGSMITH-API-KEY>
# export LANGCHAIN_PROJECT="My Project Name" #  Optional: "default" is used if not set
# export LANGCHAIN_ENDPOINT=https://api.smith.langchain.com # or your own server

Integrations

Langsmith's traceable wrapper function makes it easy to trace any function or LLM call in your own favorite framework. Below are some examples.

OpenAI SDK

The easiest way to trace calls from the OpenAI SDK with LangSmith is using the wrapOpenAI wrapper function available in LangSmith 0.1.3 and up.

In order to use, you first need to set your LangSmith API key:

export LANGCHAIN_API_KEY=<your-api-key>

Next, you will need to install the LangSmith SDK and the OpenAI SDK:

npm install langsmith openai

After that, initialize your OpenAI client and wrap the client with wrapOpenAI method to enable tracing for the completions and chat completions methods:

import { OpenAI } from "openai";
import { wrapOpenAI } from "langsmith/wrappers";

const openai = wrapOpenAI(new OpenAI());

await openai.chat.completions.create({
  model: "gpt-3.5-turbo",
  messages: [{ content: "Hi there!", role: "user" }],
});

Alternatively, you can use the traceable function to wrap the client methods you want to use:

import { traceable } from "langsmith/traceable";

const openai = new OpenAI();

const createCompletion = traceable(
  openai.chat.completions.create.bind(openai.chat.completions),
  { name: "OpenAI Chat Completion", run_type: "llm" }
);

await createCompletion({
  model: "gpt-3.5-turbo",
  messages: [{ content: "Hi there!", role: "user" }],
});

Note the use of .bind to preserve the function's context. The run_type field in the extra config object marks the function as an LLM call, and enables token usage tracking for OpenAI.

Oftentimes, you use the OpenAI client inside of other functions or as part of a longer sequence. You can automatically get nested traces by using this wrapped method within other functions wrapped with traceable.

const nestedTrace = traceable(async (text: string) => {
  const completion = await openai.chat.completions.create({
    model: "gpt-3.5-turbo",
    messages: [{ content: text, role: "user" }],
  });
  return completion;
});

await nestedTrace("Why is the sky blue?");

{
  "id": "chatcmpl-8sPToJQLLVepJvyeTfzZMOMVIKjMo",
  "object": "chat.completion",
  "created": 1707978348,
  "model": "gpt-3.5-turbo-0613",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "The sky appears blue because of a phenomenon known as Rayleigh scattering. The Earth's atmosphere is composed of tiny molecules, such as nitrogen and oxygen, which are much smaller than the wavelength of visible light. When sunlight interacts with these molecules, it gets scattered in all directions. However, shorter wavelengths of light (blue and violet) are scattered more compared to longer wavelengths (red, orange, and yellow). \n\nAs a result, when sunlight passes through the Earth's atmosphere, the blue and violet wavelengths are scattered in all directions, making the sky appear blue. This scattering of shorter wavelengths is also responsible for the vibrant colors observed during sunrise and sunset, when the sunlight has to pass through a thicker portion of the atmosphere, causing the longer wavelengths to dominate the scattered light."
      },
      "logprobs": null,
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 13,
    "completion_tokens": 154,
    "total_tokens": 167
  },
  "system_fingerprint": null
}

:::tip Click here to see an example LangSmith trace of the above. :::

Next.js

You can use the traceable wrapper function in Next.js apps to wrap arbitrary functions much like in the example above.

One neat trick you can use for Next.js and other similar server frameworks is to wrap the entire exported handler for a route to group traces for the any sub-runs. Here's an example:

import { NextRequest, NextResponse } from "next/server";

import { OpenAI } from "openai";
import { traceable } from "langsmith/traceable";
import { wrapOpenAI } from "langsmith/wrappers";

export const runtime = "edge";

const handler = traceable(
  async function () {
    const openai = wrapOpenAI(new OpenAI());

    const completion = await openai.chat.completions.create({
      model: "gpt-3.5-turbo",
      messages: [{ content: "Why is the sky blue?", role: "user" }],
    });

    const response1 = completion.choices[0].message.content;

    const completion2 = await openai.chat.completions.create({
      model: "gpt-3.5-turbo",
      messages: [
        { content: "Why is the sky blue?", role: "user" },
        { content: response1, role: "assistant" },
        { content: "Cool thank you!", role: "user" },
      ],
    });

    const response2 = completion2.choices[0].message.content;

    return {
      text: response2,
    };
  },
  {
    name: "Simple Next.js handler",
  }
);

export async function POST(req: NextRequest) {
  const result = await handler();
  return NextResponse.json(result);
}

The two OpenAI calls within the handler will be traced with appropriate inputs, outputs, and token usage information.

:::tip Click here to see an example LangSmith trace of the above. :::

Vercel AI SDK

The Vercel AI SDK contains integrations with a variety of model providers. Here's an example of how you can trace outputs in a Next.js handler:

import { traceable } from "langsmith/traceable";
import { OpenAIStream, StreamingTextResponse } from "ai";

// Note: There are no types for the Mistral API client yet.
import MistralClient from "@mistralai/mistralai";

const client = new MistralClient(process.env.MISTRAL_API_KEY || "");

export async function POST(req: Request) {
  // Extract the `messages` from the body of the request
  const { messages } = await req.json();

  const mistralChatStream = traceable(client.chatStream.bind(client), {
    name: "Mistral Stream",
    run_type: "llm",
  });

  const response = await mistralChatStream({
    model: "mistral-tiny",
    maxTokens: 1000,
    messages,
  });

  // Convert the response into a friendly text-stream. The Mistral client responses are
  // compatible with the Vercel AI SDK OpenAIStream adapter.
  const stream = OpenAIStream(response as any);

  // Respond with the stream
  return new StreamingTextResponse(stream);
}

See the AI SDK docs for more examples.

Arbitrary SDKs

You can use the generic wrapSDK method to add tracing for arbitrary SDKs.

Do note that this will trace ALL methods in the SDK, not just chat completion endpoints. If the SDK you are wrapping has other methods, we recommend using it for only LLM calls.

Here's an example using the Anthropic SDK:

import { wrapSDK } from "langsmith/wrappers";
import { Anthropic } from "@anthropic-ai/sdk";

const originalSDK = new Anthropic();
const sdkWithTracing = wrapSDK(originalSDK);

const response = await sdkWithTracing.messages.create({
  messages: [
    {
      role: "user",
      content: `What is 1 + 1? Respond only with "2" and nothing else.`,
    },
  ],
  model: "claude-3-sonnet-20240229",
  max_tokens: 1024,
});

:::tip Click here to see an example LangSmith trace of the above. :::

Alternatives: Log traces using a RunTree.

A RunTree tracks your application. Each RunTree object is required to have a name and run_type. These and other important attributes are as follows:

• name: string - used to identify the component's purpose
• run_type: string - Currently one of "llm", "chain" or "tool"; more options will be added in the future
• inputs: Record<string, any> - the inputs to the component
• outputs: Optional<Record<string, any>> - the (optional) returned values from the component
• error: Optional<string> - Any error messages that may have arisen during the call

import { RunTree, RunTreeConfig } from "langsmith";

const parentRunConfig: RunTreeConfig = {
  name: "My Chat Bot",
  run_type: "chain",
  inputs: {
    text: "Summarize this morning's meetings.",
  },
  serialized: {}, // Serialized representation of this chain
  // project_name: "Defaults to the LANGCHAIN_PROJECT env var"
  // apiUrl: "Defaults to the LANGCHAIN_ENDPOINT env var"
  // apiKey: "Defaults to the LANGCHAIN_API_KEY env var"
};

const parentRun = new RunTree(parentRunConfig);

await parentRun.postRun();

const childLlmRun = await parentRun.createChild({
  name: "My Proprietary LLM",
  run_type: "llm",
  inputs: {
    prompts: [
      "You are an AI Assistant. The time is XYZ." +
        " Summarize this morning's meetings.",
    ],
  },
});

await childLlmRun.postRun();

await childLlmRun.end({
  outputs: {
    generations: [
      "I should use the transcript_loader tool" +
        " to fetch meeting_transcripts from XYZ",
    ],
  },
});

await childLlmRun.patchRun();

const childToolRun = await parentRun.createChild({
  name: "transcript_loader",
  run_type: "tool",
  inputs: {
    date: "XYZ",
    content_type: "meeting_transcripts",
  },
});
await childToolRun.postRun();

await childToolRun.end({
  outputs: {
    meetings: ["Meeting1 notes.."],
  },
});

await childToolRun.patchRun();

const childChainRun = await parentRun.createChild({
  name: "Unreliable Component",
  run_type: "tool",
  inputs: {
    input: "Summarize these notes...",
  },
});

await childChainRun.postRun();

try {
  // .... the component does work
  throw new Error("Something went wrong");
} catch (e) {
  await childChainRun.end({
    error: `I errored again ${e.message}`,
  });
  await childChainRun.patchRun();
  throw e;
}

await childChainRun.patchRun();

await parentRun.end({
  outputs: {
    output: ["The meeting notes are as follows:..."],
  },
});

// False directs to not exclude child runs
await parentRun.patchRun();

Evaluation

Create a Dataset from Existing Runs

Once your runs are stored in LangSmith, you can convert them into a dataset. For this example, we will do so using the Client, but you can also do this using the web interface, as explained in the LangSmith docs.

import { Client } from "langsmith/client";
const client = new Client({
  // apiUrl: "https://api.langchain.com", // Defaults to the LANGCHAIN_ENDPOINT env var
  // apiKey: "my_api_key", // Defaults to the LANGCHAIN_API_KEY env var
  /* callerOptions: {
         maxConcurrency?: Infinity; // Maximum number of concurrent requests to make
         maxRetries?: 6; // Maximum number of retries to make
    */
});
const datasetName = "Example Dataset";
// We will only use examples from the top level AgentExecutor run here,
// and exclude runs that errored.
const runs = await client.listRuns({
  projectName: "my_project",
  executionOrder: 1,
  error: false,
});

const dataset = await client.createDataset(datasetName, {
  description: "An example dataset",
});

for (const run of runs) {
  await client.createExample(run.inputs, run.outputs ?? {}, {
    datasetId: dataset.id,
  });
}

Evaluating Runs

Check out the LangSmith Testing & Evaluation dos for up-to-date workflows.

For generating automated feedback on individual runs, you can run evaluations directly using the LangSmith client.

import { StringEvaluator } from "langsmith/evaluation";

function jaccardChars(output: string, answer: string): number {
  const predictionChars = new Set(output.trim().toLowerCase());
  const answerChars = new Set(answer.trim().toLowerCase());
  const intersection = [...predictionChars].filter((x) => answerChars.has(x));
  const union = new Set([...predictionChars, ...answerChars]);
  return intersection.length / union.size;
}

async function grader(config: {
  input: string;
  prediction: string;
  answer?: string;
}): Promise<{ score: number; value: string }> {
  let value: string;
  let score: number;
  if (config.answer === null || config.answer === undefined) {
    value = "AMBIGUOUS";
    score = 0.5;
  } else {
    score = jaccardChars(config.prediction, config.answer);
    value = score > 0.9 ? "CORRECT" : "INCORRECT";
  }
  return { score: score, value: value };
}

const evaluator = new StringEvaluator({
  evaluationName: "Jaccard",
  gradingFunction: grader,
});

const runs = await client.listRuns({
  projectName: "my_project",
  executionOrder: 1,
  error: false,
});

for (const run of runs) {
  client.evaluateRun(run, evaluator);
}

Additional Documentation

To learn more about the LangSmith platform, check out the docs.