Langfuse CI Integration

Overview

Integrate Langfuse into CI/CD pipelines: trace validation tests, prompt regression testing, experiment-driven quality gates, automated prompt deployment from version control, and score monitoring.

Prerequisites

• Langfuse API keys stored as GitHub secrets (LANGFUSEPUBLICKEY, LANGFUSESECRETKEY)
• Test framework (Vitest or Jest)
• OpenAI API key for LLM tests

Instructions

Step 1: GitHub Actions Workflow for AI Quality Tests

# .github/workflows/langfuse-tests.yml
name: AI Quality Tests

on:
  pull_request:
    paths: ["src/ai/**", "src/prompts/**", "tests/ai/**"]

jobs:
  ai-quality:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with: { node-version: "20", cache: "npm" }
      - run: npm ci

      - name: Run AI quality tests with tracing
        env:
          LANGFUSE_PUBLIC_KEY: ${{ secrets.LANGFUSE_PUBLIC_KEY }}
          LANGFUSE_SECRET_KEY: ${{ secrets.LANGFUSE_SECRET_KEY }}
          LANGFUSE_BASE_URL: ${{ vars.LANGFUSE_BASE_URL || 'https://cloud.langfuse.com' }}
          OPENAI_API_KEY: ${{ secrets.OPENAI_API_KEY }}
        run: npx vitest run tests/ai/ --reporter=verbose

      - name: Langfuse connectivity check
        env:
          LANGFUSE_PUBLIC_KEY: ${{ secrets.LANGFUSE_PUBLIC_KEY }}
          LANGFUSE_SECRET_KEY: ${{ secrets.LANGFUSE_SECRET_KEY }}
        run: |
          node -e "
            const { LangfuseClient } = require('@langfuse/client');
            const lf = new LangfuseClient();
            lf.prompt.get('__ci-health__').catch(() => {});
            console.log('Langfuse SDK initialized OK');
          "

Step 2: Prompt Regression Tests

// tests/ai/prompt-quality.test.ts
import { describe, it, expect, afterAll } from "vitest";
import { LangfuseClient } from "@langfuse/client";
import { startActiveObservation, updateActiveObservation } from "@langfuse/tracing";
import OpenAI from "openai";

const langfuse = new LangfuseClient();
const openai = new OpenAI();

describe("Prompt Quality Regression", () => {
  it("summarization prompt produces valid output", async () => {
    const prompt = await langfuse.prompt.get("summarize-article", { type: "text" });
    const compiled = prompt.compile({ maxLength: "100 words" });

    const result = await startActiveObservation(
      { name: "ci-test-summarize", asType: "generation" },
      async () => {
        updateActiveObservation({ model: "gpt-4o-mini", input: compiled });

        const response = await openai.chat.completions.create({
          model: "gpt-4o-mini",
          messages: [{ role: "user", content: compiled }],
          temperature: 0,
        });

        const output = response.choices[0].message.content || "";
        updateActiveObservation({
          output,
          usage: {
            promptTokens: response.usage?.prompt_tokens,
            completionTokens: response.usage?.completion_tokens,
          },
        });
        return output;
      }
    );

    expect(result.length).toBeGreaterThan(20);
    expect(result.length).toBeLessThan(600);
  });

  it("classification prompt returns valid intent", async () => {
    const prompt = await langfuse.prompt.get("classify-intent", { type: "text" });
    const compiled = prompt.compile({ userMessage: "I want to cancel my subscription" });

    const response = await openai.chat.completions.create({
      model: "gpt-4o-mini",
      messages: [{ role: "user", content: compiled }],
      temperature: 0,
    });

    const intent = response.choices[0].message.content?.trim().toLowerCase() || "";
    const validIntents = ["billing", "cancellation", "support", "feedback"];
    expect(validIntents).toContain(intent);
  });
});

Step 3: Experiment-Driven Quality Gates

// tests/ai/experiment-gate.test.ts
import { describe, it, expect } from "vitest";
import { LangfuseClient } from "@langfuse/client";
import OpenAI from "openai";

const langfuse = new LangfuseClient();
const openai = new OpenAI();

describe("Quality Gate: Intent Classification", () => {
  it("scores above 80% accuracy on test dataset", async () => {
    async function classifyIntent(input: { query: string }) {
      const response = await openai.chat.completions.create({
        model: "gpt-4o-mini",
        messages: [
          { role: "system", content: "Classify intent. Return one word." },
          { role: "user", content: input.query },
        ],
        temperature: 0,
      });
      return response.choices[0].message.content?.trim() || "";
    }

    const result = await langfuse.runExperiment({
      datasetName: "intent-classification-test",
      runName: `ci-${process.env.GITHUB_SHA?.slice(0, 7) || "local"}`,
      task: classifyIntent,
      evaluators: [
        ({ output, expectedOutput }) => ({
          name: "exact-match",
          value: output.toLowerCase() === expectedOutput.intent.toLowerCase() ? 1 : 0,
          dataType: "BOOLEAN" as const,
        }),
      ],
    });

    // Calculate accuracy
    const scores = result.runs.flatMap((r) => r.scores || []);
    const accuracy = scores.filter((s) => s.value === 1).length / scores.length;

    console.log(`Accuracy: ${(accuracy * 100).toFixed(1)}%`);
    expect(accuracy).toBeGreaterThanOrEqual(0.8);
  });
});

Step 4: Automated Prompt Deployment

# .github/workflows/deploy-prompts.yml
name: Deploy Prompts to Langfuse

on:
  push:
    branches: [main]
    paths: ["src/prompts/**"]

jobs:
  deploy:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-node@v4
        with: { node-version: "20", cache: "npm" }
      - run: npm ci

      - name: Deploy prompts
        env:
          LANGFUSE_PUBLIC_KEY: ${{ secrets.LANGFUSE_PUBLIC_KEY }}
          LANGFUSE_SECRET_KEY: ${{ secrets.LANGFUSE_SECRET_KEY }}
        run: node scripts/deploy-prompts.mjs

// scripts/deploy-prompts.mjs
import { LangfuseClient } from "@langfuse/client";
import { readdirSync, readFileSync } from "fs";
import { join } from "path";

const langfuse = new LangfuseClient();
const promptDir = join(process.cwd(), "src/prompts");

for (const file of readdirSync(promptDir).filter((f) => f.endsWith(".json"))) {
  const config = JSON.parse(readFileSync(join(promptDir, file), "utf-8"));

  await langfuse.api.prompts.create({
    name: config.name,
    prompt: config.template,
    type: config.type || "text",
    config: config.config || {},
    labels: ["production", `deploy-${new Date().toISOString().split("T")[0]}`],
  });

  console.log(`Deployed: ${config.name}`);
}

Step 5: Score Regression Monitoring

// scripts/check-quality-regression.ts
import { LangfuseClient } from "@langfuse/client";

const langfuse = new LangfuseClient();

async function checkRegression() {
  const scores = await langfuse.api.scores.list({
    name: "quality",
    limit: 100,
  });

  const values = scores.data.map((s) => s.value).filter((v): v is number => v !== null);
  const avg = values.reduce((a, b) => a + b, 0) / values.length;

  console.log(`Average quality score: ${avg.toFixed(3)} (n=${values.length})`);

  if (avg < 0.7) {
    console.error("QUALITY REGRESSION: Score below 0.7 threshold");
    process.exit(1);
  }
}

checkRegression();

CI Best Practices

Error Handling

Output

The pipeline reports the test suite result, experiment/run identifier, and aggregate quality score. A regression exits non-zero so the pull request is blocked; a pass leaves a traceable Langfuse run linked to the commit SHA.

Examples

Run the focused quality suite locally before opening a pull request:

LANGFUSE_PUBLIC_KEY=pk-lf-... LANGFUSE_SECRET_KEY=sk-lf-... \
  npx vitest run tests/ai/ --reporter=verbose

For an intentionally changed prompt, first update the expected dataset result, then run the experiment gate and inspect the resulting run in Langfuse before merging.

Resources

• Experiments via SDK
• Prompt Management
• Scores via SDK