Langfuse Core Workflow B: Evaluation, Scoring & Datasets

Overview

Implement LLM output evaluation using Langfuse scores (numeric, categorical, boolean), the experiment runner SDK for dataset-driven benchmarks, prompt management with versioned prompts, and LLM-as-a-Judge evaluation patterns.

Prerequisites

• Langfuse SDK configured with API keys
• Traces already being collected (see langfuse-core-workflow-a)
• For v4+: @langfuse/client installed

Instructions

Step 1: Score Traces via SDK

Langfuse supports three score data types: Numeric, Categorical, and Boolean.

import { LangfuseClient } from "@langfuse/client";

const langfuse = new LangfuseClient();

// Numeric score (e.g., 0-1 quality rating)
await langfuse.score.create({
  traceId: "trace-abc-123",
  name: "relevance",
  value: 0.92,
  dataType: "NUMERIC",
  comment: "Highly relevant answer with good context usage",
});

// Categorical score (e.g., pass/fail classification)
await langfuse.score.create({
  traceId: "trace-abc-123",
  observationId: "gen-xyz-456", // Optional: score a specific generation
  name: "quality-tier",
  value: "excellent",
  dataType: "CATEGORICAL",
});

// Boolean score (e.g., thumbs up/down)
await langfuse.score.create({
  traceId: "trace-abc-123",
  name: "user-approved",
  value: 1, // 1 = true, 0 = false
  dataType: "BOOLEAN",
  comment: "User clicked thumbs up",
});

Step 2: User Feedback Collection

// API endpoint for frontend feedback widget
app.post("/api/feedback", async (req, res) => {
  const { traceId, rating, comment } = req.body;

  // Thumbs up/down
  await langfuse.score.create({
    traceId,
    name: "user-feedback",
    value: rating === "positive" ? 1 : 0,
    dataType: "BOOLEAN",
    comment,
  });

  // Granular star rating (1-5)
  if (req.body.stars) {
    await langfuse.score.create({
      traceId,
      name: "star-rating",
      value: req.body.stars,
      dataType: "NUMERIC",
      comment: `${req.body.stars}/5 stars`,
    });
  }

  res.json({ success: true });
});

Step 3: Prompt Management

// Fetch a versioned prompt from Langfuse
const textPrompt = await langfuse.prompt.get("summarize-article", {
  type: "text",
  label: "production", // or "latest", "staging"
});

// Compile with variables -- replaces {{variable}} placeholders
const compiled = textPrompt.compile({
  maxLength: "100 words",
  tone: "professional",
});

// Chat prompts return message arrays
const chatPrompt = await langfuse.prompt.get("customer-support", {
  type: "chat",
});

const messages = chatPrompt.compile({
  customerName: "Alice",
  issue: "billing question",
});
// messages = [{ role: "system", content: "..." }, { role: "user", content: "..." }]

Step 4: Create and Populate Datasets

// Create a dataset for evaluation
await langfuse.api.datasets.create({
  name: "customer-support-v1",
  description: "Test cases for customer support chatbot",
  metadata: { version: "1.0", domain: "support" },
});

// Add test items
const testCases = [
  {
    input: { query: "How do I cancel my subscription?" },
    expectedOutput: { intent: "cancellation", sentiment: "neutral" },
    metadata: { category: "billing" },
  },
  {
    input: { query: "Your product is amazing!" },
    expectedOutput: { intent: "feedback", sentiment: "positive" },
    metadata: { category: "feedback" },
  },
];

for (const testCase of testCases) {
  await langfuse.api.datasetItems.create({
    datasetName: "customer-support-v1",
    input: testCase.input,
    expectedOutput: testCase.expectedOutput,
    metadata: testCase.metadata,
  });
}

Step 5: Run Experiments with the Experiment Runner

import { LangfuseClient } from "@langfuse/client";

const langfuse = new LangfuseClient();

// Define the task function -- your LLM application logic
async function classifyIntent(input: { query: string }): Promise<string> {
  const response = await openai.chat.completions.create({
    model: "gpt-4o-mini",
    messages: [
      { role: "system", content: "Classify the user intent. Return one word." },
      { role: "user", content: input.query },
    ],
    temperature: 0,
  });
  return response.choices[0].message.content?.trim() || "";
}

// Define evaluator functions
function exactMatch({ output, expectedOutput }: {
  output: string;
  expectedOutput: { intent: string };
}) {
  return {
    name: "exact-match",
    value: output.toLowerCase() === expectedOutput.intent.toLowerCase() ? 1 : 0,
    dataType: "BOOLEAN" as const,
  };
}

// Run the experiment
const result = await langfuse.runExperiment({
  datasetName: "customer-support-v1",
  runName: "gpt-4o-mini-classifier-v1",
  runDescription: "Testing intent classification with gpt-4o-mini",
  task: classifyIntent,
  evaluators: [exactMatch],
});

console.log(`Experiment complete. ${result.runs.length} items evaluated.`);
// View results in Langfuse UI: Datasets > customer-support-v1 > Runs

Step 6: LLM-as-a-Judge Evaluation

async function llmJudge({ output, input, expectedOutput }: {
  output: string;
  input: { query: string };
  expectedOutput: { intent: string; sentiment: string };
}) {
  const judgment = await openai.chat.completions.create({
    model: "gpt-4o",
    temperature: 0,
    messages: [
      {
        role: "system",
        content: `You are an AI evaluator. Score the response 0-10 on accuracy and helpfulness.
Return JSON: {"score": <number>, "reasoning": "<explanation>"}`,
      },
      {
        role: "user",
        content: `Query: ${input.query}\nExpected: ${JSON.stringify(expectedOutput)}\nActual: ${output}`,
      },
    ],
    response_format: { type: "json_object" },
  });

  const result = JSON.parse(judgment.choices[0].message.content || "{}");

  return {
    name: "llm-judge-quality",
    value: result.score / 10, // Normalize to 0-1
    dataType: "NUMERIC" as const,
    comment: result.reasoning,
  };
}

// Use as an evaluator in experiments
await langfuse.runExperiment({
  datasetName: "customer-support-v1",
  runName: "judge-evaluation-v1",
  task: classifyIntent,
  evaluators: [exactMatch, llmJudge],
});

Error Handling

Output

Produce a versioned dataset or prompt reference, an experiment run identifier, and per-item plus aggregate scores. Summarize the threshold, sample size, and failed cases so a release decision is reproducible.

Examples

Create a small customer-support-v1 dataset, run the exact-match evaluator at temperature zero, and inspect failed items before changing the prompt. Add the LLM-as-a-judge only as a second score; retain deterministic exact-match or rubric evidence as the release gate.

Resources

• Scores via API/SDK
• Datasets & Experiments
• Experiment Runner SDK
• Prompt Management
• LLM-as-a-Judge

Next Steps

For common error debugging, see langfuse-common-errors. For CI/CD integration of evaluations, see langfuse-ci-integration.