How to evaluate an LLM application

This guide shows you how to run an evaluation on an LLM application using the LangSmith SDK.

In this guide we’ll go over how to evaluate an application using the evaluate() method in the LangSmith SDK.

For larger evaluation jobs in Python we recommend using aevaluate(), the asynchronous version of evaluate(). It is still worthwhile to read this guide first, as the two have identical interfaces, before reading the how-to guide on running an evaluation asynchronously.In JS/TS evaluate() is already asynchronous so no separate method is needed.It is also important to configure the max_concurrency/maxConcurrency arg when running large jobs. This parallelizes evaluation by effectively splitting the dataset across threads.

Define an application

First we need an application to evaluate. Let’s create a simple toxicity classifier for this example.

from langsmith import traceable, wrappers
from openai import OpenAI

# Optionally wrap the OpenAI client to trace all model calls.
oai_client = wrappers.wrap_openai(OpenAI())

# Optionally add the 'traceable' decorator to trace the inputs/outputs of this function.
@traceable
def toxicity_classifier(inputs: dict) -> dict:
    instructions = (
      "Please review the user query below and determine if it contains any form of toxic behavior, "
      "such as insults, threats, or highly negative comments. Respond with 'Toxic' if it does "
      "and 'Not toxic' if it doesn't."
    )
    messages = [
        {"role": "system", "content": instructions},
        {"role": "user", "content": inputs["text"]},
    ]
    result = oai_client.chat.completions.create(
        messages=messages, model="gpt-4o-mini", temperature=0
    )
    return {"class": result.choices[0].message.content}

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

// Optionally wrap the OpenAI client to trace all model calls.
const oaiClient = wrapOpenAI(new OpenAI());

// Optionally add the 'traceable' wrapper to trace the inputs/outputs of this function.
const toxicityClassifier = traceable(
  async (text: string) => {
    const result = await oaiClient.chat.completions.create({
      messages: [
        {
           role: "system",
          content: "Please review the user query below and determine if it contains any form of toxic behavior, such as insults, threats, or highly negative comments. Respond with 'Toxic' if it does, and 'Not toxic' if it doesn't.",
        },
        { role: "user", content: text },
      ],
      model: "gpt-4o-mini",
      temperature: 0,
    });

    return result.choices[0].message.content;
  },
  { name: "toxicityClassifier" }
);

We’ve optionally enabled tracing to capture the inputs and outputs of each step in the pipeline. To understand how to annotate your code for tracing, please refer to this guide.

Create or select a dataset

We need a Dataset to evaluate our application on. Our dataset will contain labeled examples of toxic and non-toxic text. Requires langsmith>=0.3.13

from langsmith import Client
ls_client = Client()

examples = [
  {
    "inputs": {"text": "Shut up, idiot"},
    "outputs": {"label": "Toxic"},
  },
  {
    "inputs": {"text": "You're a wonderful person"},
    "outputs": {"label": "Not toxic"},
  },
  {
    "inputs": {"text": "This is the worst thing ever"},
    "outputs": {"label": "Toxic"},
  },
  {
    "inputs": {"text": "I had a great day today"},
    "outputs": {"label": "Not toxic"},
  },
  {
    "inputs": {"text": "Nobody likes you"},
    "outputs": {"label": "Toxic"},
  },
  {
    "inputs": {"text": "This is unacceptable. I want to speak to the manager."},
    "outputs": {"label": "Not toxic"},
  },
]

dataset = ls_client.create_dataset(dataset_name="Toxic Queries")
ls_client.create_examples(
  dataset_id=dataset.id,
  examples=examples,
)

import { Client } from "langsmith";

const langsmith = new Client();

// create a dataset
const labeledTexts = [
  ["Shut up, idiot", "Toxic"],
  ["You're a wonderful person", "Not toxic"],
  ["This is the worst thing ever", "Toxic"],
  ["I had a great day today", "Not toxic"],
  ["Nobody likes you", "Toxic"],
  ["This is unacceptable. I want to speak to the manager.", "Not toxic"],
];

const [inputs, outputs] = labeledTexts.reduce<
  [Array<{ input: string }>, Array<{ outputs: string }>]
>(
  ([inputs, outputs], item) => [
    [...inputs, { input: item[0] }],
    [...outputs, { outputs: item[1] }],
  ],
  [[], []]
);

const datasetName = "Toxic Queries";
const toxicDataset = await langsmith.createDataset(datasetName);
await langsmith.createExamples({ inputs, outputs, datasetId: toxicDataset.id });

For more details on datasets, refer to the Manage datasets page.

Define an evaluator

You can also check out LangChain’s open source evaluation package openevals for common pre-built evaluators.

Evaluators are functions for scoring your application’s outputs. They take in the example inputs, actual outputs, and, when present, the reference outputs. Since we have labels for this task, our evaluator can directly check if the actual outputs match the reference outputs.

Python: Requires langsmith>=0.3.13
TypeScript: Requires langsmith>=0.2.9

def correct(inputs: dict, outputs: dict, reference_outputs: dict) -> bool:
    return outputs["class"] == reference_outputs["label"]

import type { EvaluationResult } from "langsmith/evaluation";

function correct({
  outputs,
  referenceOutputs,
}: {
  outputs: Record<string, any>;
  referenceOutputs?: Record<string, any>;
}): EvaluationResult {
  const score = outputs.output === referenceOutputs?.outputs;
  return { key: "correct", score };
}

Run the evaluation

We’ll use the evaluate() / aevaluate() methods to run the evaluation. The key arguments are:

a target function that takes an input dictionary and returns an output dictionary. The example.inputs field of each Example is what gets passed to the target function. In this case our toxicity_classifier is already set up to take in example inputs so we can use it directly.
data - the name OR UUID of the LangSmith dataset to evaluate on, or an iterator of examples
evaluators - a list of evaluators to score the outputs of the function

Python: Requires langsmith>=0.3.13

# Can equivalently use the 'evaluate' function directly:
# from langsmith import evaluate; evaluate(...)
results = ls_client.evaluate(
    toxicity_classifier,
    data=dataset.name,
    evaluators=[correct],
    experiment_prefix="gpt-4o-mini, baseline",  # optional, experiment name prefix
    description="Testing the baseline system.",  # optional, experiment description
    max_concurrency=4, # optional, add concurrency
)

import { evaluate } from "langsmith/evaluation";

await evaluate((inputs) => toxicityClassifier(inputs["input"]), {
  data: datasetName,
  evaluators: [correct],
  experimentPrefix: "gpt-4o-mini, baseline",  // optional, experiment name prefix
  maxConcurrency: 4, // optional, add concurrency
});

Explore the results

Each invocation of evaluate() creates an Experiment which can be viewed in the LangSmith UI or queried via the SDK. Evaluation scores are stored against each actual output as feedback. If you’ve annotated your code for tracing, you can open the trace of each row in a side panel view.

Reference code

Click to see a consolidated code snippet

from langsmith import Client, traceable, wrappers
from openai import OpenAI

# Step 1. Define an application
oai_client = wrappers.wrap_openai(OpenAI())

@traceable
def toxicity_classifier(inputs: dict) -> str:
    system = (
      "Please review the user query below and determine if it contains any form of toxic behavior, "
      "such as insults, threats, or highly negative comments. Respond with 'Toxic' if it does "
      "and 'Not toxic' if it doesn't."
    )
    messages = [
        {"role": "system", "content": system},
        {"role": "user", "content": inputs["text"]},
    ]
    result = oai_client.chat.completions.create(
        messages=messages, model="gpt-4o-mini", temperature=0
    )
    return result.choices[0].message.content

# Step 2. Create a dataset
ls_client = Client()
dataset = ls_client.create_dataset(dataset_name="Toxic Queries")
examples = [
  {
    "inputs": {"text": "Shut up, idiot"},
    "outputs": {"label": "Toxic"},
  },
  {
    "inputs": {"text": "You're a wonderful person"},
    "outputs": {"label": "Not toxic"},
  },
  {
    "inputs": {"text": "This is the worst thing ever"},
    "outputs": {"label": "Toxic"},
  },
  {
    "inputs": {"text": "I had a great day today"},
    "outputs": {"label": "Not toxic"},
  },
  {
    "inputs": {"text": "Nobody likes you"},
    "outputs": {"label": "Toxic"},
  },
  {
    "inputs": {"text": "This is unacceptable. I want to speak to the manager."},
    "outputs": {"label": "Not toxic"},
  },
]
ls_client.create_examples(
  dataset_id=dataset.id,
  examples=examples,
)

# Step 3. Define an evaluator
def correct(inputs: dict, outputs: dict, reference_outputs: dict) -> bool:
    return outputs["output"] == reference_outputs["label"]

# Step 4. Run the evaluation
# Client.evaluate() and evaluate() behave the same.
results = ls_client.evaluate(
    toxicity_classifier,
    data=dataset.name,
    evaluators=[correct],
    experiment_prefix="gpt-4o-mini, simple",  # optional, experiment name prefix
    description="Testing the baseline system.",  # optional, experiment description
    max_concurrency=4,  # optional, add concurrency
)

import { OpenAI } from "openai";
import { Client } from "langsmith";
import { evaluate, EvaluationResult } from "langsmith/evaluation";
import type { Run, Example } from "langsmith/schemas";
import { traceable } from "langsmith/traceable";
import { wrapOpenAI } from "langsmith/wrappers";

const oaiClient = wrapOpenAI(new OpenAI());

const toxicityClassifier = traceable(
  async (text: string) => {
    const result = await oaiClient.chat.completions.create({
      messages: [
        {
          role: "system",
          content: "Please review the user query below and determine if it contains any form of toxic behavior, such as insults, threats, or highly negative comments. Respond with 'Toxic' if it does, and 'Not toxic' if it doesn't.",
        },
        { role: "user", content: text },
      ],
      model: "gpt-4o-mini",
      temperature: 0,
    });
    return result.choices[0].message.content;
  },
  { name: "toxicityClassifier" }
);

const langsmith = new Client();

// create a dataset
const labeledTexts = [
  ["Shut up, idiot", "Toxic"],
  ["You're a wonderful person", "Not toxic"],
  ["This is the worst thing ever", "Toxic"],
  ["I had a great day today", "Not toxic"],
  ["Nobody likes you", "Toxic"],
  ["This is unacceptable. I want to speak to the manager.", "Not toxic"],
];

const [inputs, outputs] = labeledTexts.reduce<
  [Array<{ input: string }>, Array<{ outputs: string }>]
>(
  ([inputs, outputs], item) => [
    [...inputs, { input: item[0] }],
    [...outputs, { outputs: item[1] }],
  ],
  [[], []]
);

const datasetName = "Toxic Queries";
const toxicDataset = await langsmith.createDataset(datasetName);
await langsmith.createExamples({ inputs, outputs, datasetId: toxicDataset.id });

// Row-level evaluator
function correct({
  outputs,
  referenceOutputs,
}: {
  outputs: Record<string, any>;
  referenceOutputs?: Record<string, any>;
}): EvaluationResult {
  const score = outputs.output === referenceOutputs?.outputs;
  return { key: "correct", score };
}

await evaluate((inputs) => toxicityClassifier(inputs["input"]), {
  data: datasetName,
  evaluators: [correct],
  experimentPrefix: "gpt-4o-mini, simple",  // optional, experiment name prefix
  maxConcurrency: 4, // optional, add concurrency
});

Edit the source of this page on GitHub.

Connect these docs programmatically to Claude, VSCode, and more via MCP for real-time answers.

Datasets

Set up evaluations

Analyze experiment results

Annotation & human feedback

Common data types

How to evaluate an LLM application

Define an application

Create or select a dataset

Define an evaluator

Run the evaluation

Explore the results

Reference code

​Define an application

​Create or select a dataset

​Define an evaluator

​Run the evaluation

​Explore the results​

​Reference code​

​Related​

Define an application

Create or select a dataset

Define an evaluator

Run the evaluation

Explore the results

Reference code

Related