AGENT SKILL

langfuse-core-workflow-b

Execute Langfuse secondary workflow: Evaluation and scoring. Use when implementing LLM evaluation, adding user feedback, or setting up automated quality scoring for AI outputs. Trigger with phrases like "langfuse evaluation", "langfuse scoring", "rate llm outputs", "langfuse feedback", "evaluate ai responses".

License 许可
MIT
最近更新
2026-11-15
Stars
★ 2.0k
Forks
233

01 前言 Front Matter

name
langfuse-core-workflow-b
author
Jeremy Longshore <jeremy@intentsolutions.io>
license
MIT
version
1.0.0
description
Execute Langfuse secondary workflow: Evaluation and scoring. Use when implementing LLM evaluation, adding user feedback, or setting up automated quality scoring for AI outputs. Trigger with phrases like "langfuse evaluation", "langfuse scoring", "rate llm outputs", "langfuse feedback", "evaluate ai responses".
allowed-tools
Read, Write, Edit, Bash(npm:*), Grep

03 SKILL.md 正文

Langfuse Core Workflow B: Evaluation & Scoring

Overview

Secondary workflow for Langfuse: evaluating and scoring LLM outputs for quality monitoring.

Prerequisites

  • Completed langfuse-core-workflow-a (tracing setup)
  • Understanding of evaluation metrics
  • (Optional) Evaluation framework like RAGAS or custom evaluators

Instructions

Step 1: Add Manual Scores to Traces

import { Langfuse } from "langfuse";

const langfuse = new Langfuse();

async function scoreTrace(
  traceId: string,
  scores: { name: string; value: number; comment?: string }[]
) {
  for (const score of scores) {
    await langfuse.score({
      traceId,
      name: score.name,
      value: score.value,
      comment: score.comment,
    });
  }
}

// Usage after generating a response
const trace = langfuse.trace({ name: "chat" });
// ... do LLM work ...

// Add scores
await scoreTrace(trace.id, [
  { name: "relevance", value: 0.9, comment: "Directly answers question" },
  { name: "accuracy", value: 0.85, comment: "Minor factual imprecision" },
  { name: "helpfulness", value: 1.0, comment: "Very helpful response" },
]);

Step 2: Capture User Feedback

// API endpoint for user feedback
async function handleFeedback(req: {
  traceId: string;
  rating: "good" | "bad";
  comment?: string;
}) {
  await langfuse.score({
    traceId: req.traceId,
    name: "user-feedback",
    value: req.rating === "good" ? 1 : 0,
    comment: req.comment,
    dataType: "BOOLEAN", // Renders as thumbs up/down in UI
  });

  return { success: true };
}

// Frontend integration
function FeedbackButtons({ traceId }: { traceId: string }) {
  const submitFeedback = async (rating: "good" | "bad") => {
    await fetch("/api/feedback", {
      method: "POST",
      body: JSON.stringify({ traceId, rating }),
    });
  };

  return (
    <div>
      <button onClick={() => submitFeedback("good")}>?</button>
      <button onClick={() => submitFeedback("bad")}>?</button>
    </div>
  );
}

Step 3: Automated Evaluation with LLM-as-Judge

async function evaluateWithLLM(
  traceId: string,
  question: string,
  answer: string,
  context: string[]
) {
  const evaluationPrompt = `
You are evaluating an AI assistant's response.

Question: ${question}
Context provided: ${context.join("\n")}
AI Response: ${answer}

Rate the response on these criteria (0.0 to 1.0):
1. Relevance: Does it answer the question?
2. Accuracy: Is it factually correct based on context?
3. Completeness: Does it fully address the question?

Respond in JSON: {"relevance": 0.X, "accuracy": 0.X, "completeness": 0.X}
`;

  // Create evaluation trace linked to original
  const evalTrace = langfuse.trace({
    name: "llm-evaluation",
    metadata: { evaluatedTraceId: traceId },
    tags: ["evaluation", "automated"],
  });

  const generation = evalTrace.generation({
    name: "judge-response",
    model: "gpt-4",
    input: evaluationPrompt,
  });

  const response = await openai.chat.completions.create({
    model: "gpt-4",
    messages: [{ role: "user", content: evaluationPrompt }],
    response_format: { type: "json_object" },
  });

  const scores = JSON.parse(response.choices[0].message.content!);

  generation.end({
    output: scores,
    usage: {
      promptTokens: response.usage?.prompt_tokens || 0,
      completionTokens: response.usage?.completion_tokens || 0,
    },
  });

  // Apply scores to original trace
  for (const [name, value] of Object.entries(scores)) {
    await langfuse.score({
      traceId,
      name: `llm-judge/${name}`,
      value: value as number,
      comment: "Automated LLM evaluation",
    });
  }

  return scores;
}

Step 4: Score Specific Generations

async function scoreGeneration(
  traceId: string,
  observationId: string, // Generation ID
  scores: Record<string, number>
) {
  for (const [name, value] of Object.entries(scores)) {
    await langfuse.score({
      traceId,
      observationId, // Links score to specific generation
      name,
      value,
    });
  }
}

// Usage
const trace = langfuse.trace({ name: "multi-step" });

const gen1 = trace.generation({ name: "step-1", model: "gpt-4" });
// ... LLM call ...
gen1.end({ output: result1 });

const gen2 = trace.generation({ name: "step-2", model: "gpt-4" });
// ... LLM call ...
gen2.end({ output: result2 });

// Score each generation separately
await scoreGeneration(trace.id, gen1.id, { quality: 0.9 });
await scoreGeneration(trace.id, gen2.id, { quality: 0.7 });

Step 5: Dataset-Based Evaluation

// Create evaluation dataset in Langfuse
async function runDatasetEvaluation(datasetName: string) {
  // Fetch dataset from Langfuse
  const dataset = await langfuse.getDataset(datasetName);

  const results = [];

  for (const item of dataset.items) {
    // Run your pipeline
    const trace = langfuse.trace({
      name: "dataset-eval",
      metadata: { datasetItemId: item.id },
    });

    const output = await runPipeline(item.input);

    trace.update({ output });

    // Compare to expected output
    const score = compareOutputs(output, item.expectedOutput);

    await langfuse.score({
      traceId: trace.id,
      name: "dataset-match",
      value: score,
    });

    // Link trace to dataset item
    await langfuse.datasetItem.link({
      datasetItemId: item.id,
      traceId: trace.id,
    });

    results.push({ input: item.input, output, score });
  }

  return results;
}

Output

  • Manual scoring for human evaluation
  • User feedback capture (thumbs up/down)
  • Automated LLM-as-judge evaluation
  • Generation-level scoring
  • Dataset-based evaluation runs

Error Handling

Issue Cause Solution
Score not appearing Wrong traceId Verify trace ID exists
Invalid score value Out of range Use 0.0-1.0 or define range
Feedback delayed Not flushed Call flushAsync() if needed
Eval trace missing Not linked Use metadata to link traces

Examples

RAGAS Integration (Python)

from langfuse import Langfuse
from ragas import evaluate
from ragas.metrics import faithfulness, answer_relevancy

langfuse = Langfuse()

async def ragas_evaluation(trace_id: str, question: str, answer: str, contexts: list):
    # Run RAGAS evaluation
    result = evaluate(
        questions=[question],
        answers=[answer],
        contexts=[contexts],
        metrics=[faithfulness, answer_relevancy],
    )

    # Send scores to Langfuse
    for metric_name, score in result.items():
        langfuse.score(
            trace_id=trace_id,
            name=f"ragas/{metric_name}",
            value=score,
            comment="RAGAS automated evaluation",
        )

    return result

Evaluation Dashboard Query

// Fetch traces with scores for analysis
const traces = await langfuse.fetchTraces({
  name: "chat",
  fromTimestamp: new Date(Date.now() - 7 * 24 * 60 * 60 * 1000),
  toTimestamp: new Date(),
});

// Calculate average scores
const scoresByName: Record<string, number[]> = {};

for (const trace of traces.data) {
  for (const score of trace.scores || []) {
    if (!scoresByName[score.name]) scoresByName[score.name] = [];
    scoresByName[score.name].push(score.value);
  }
}

const averages = Object.fromEntries(
  Object.entries(scoresByName).map(([name, values]) => [
    name,
    values.reduce((a, b) => a + b, 0) / values.length,
  ])
);

console.log("Average scores:", averages);

Resources

Next Steps

For debugging issues, see langfuse-common-errors.

04 同仓库更多 Skills

Dicklesworthstone
Dicklesworthstone/pi_agent_rust

theme-factory

Toolkit for styling artifacts with a theme. These artifacts can be slides, docs, reportings, HTML landing pages, etc. There are 10 pre-set themes with colors/fonts that you can apply to any artifact that has been creating, or can generate a new theme on-the-fly.

视觉设计
★ 2.0k
Dicklesworthstone
Dicklesworthstone/pi_agent_rust

web-artifacts-builder

Suite of tools for creating elaborate, multi-component claude.ai HTML artifacts using modern frontend web technologies (React, Tailwind CSS, shadcn/ui). Use for complex artifacts requiring state management, routing, or shadcn/ui components - not for simple single-file HTML/JSX artifacts.

前端开发
★ 2.0k
Dicklesworthstone
Dicklesworthstone/pi_agent_rust

webapp-testing

Toolkit for interacting with and testing local web applications using Playwright. Supports verifying frontend functionality, debugging UI behavior, capturing browser screenshots, and viewing browser logs.

测试与质量前端开发
★ 2.0k
Dicklesworthstone
Dicklesworthstone/pi_agent_rust

neuropixels-analysis

Neuropixels neural recording analysis. Load SpikeGLX/OpenEphys data, preprocess, motion correction, Kilosort4 spike sorting, quality metrics, Allen/IBL curation, AI-assisted visual analysis, for Neuropixels 1.0/2.0 extracellular electrophysiology. Use when working with neural recordings, spike sorting, extracellular electrophysiology, or when the user mentions Neuropixels, SpikeGLX, Open Ephys, Kilosort, quality metrics, or unit curation.

★ 2.0k
Dicklesworthstone
Dicklesworthstone/pi_agent_rust

skill-creator

Guide for creating effective skills. This skill should be used when users want to create a new skill (or update an existing skill) that extends Claude's capabilities with specialized knowledge, workflows, or tool integrations.

Skill与智能体开发
★ 2.0k
Dicklesworthstone
Dicklesworthstone/pi_agent_rust

notebooklm

Use this skill to query your Google NotebookLM notebooks directly from Claude Code for source-grounded, citation-backed answers from Gemini. Browser automation, library management, persistent auth. Drastically reduced hallucinations through document-only responses.

Skill与智能体开发
★ 2.0k