Evals-Driven Development: Building Automated CI/CD Quality Gates for LLM Applications

Software Engineering Without Unit Tests is Chaos

In traditional software development, no competent team would deploy code to production without continuous integration: unit tests, integration tests, and linting checks that must pass before merging a pull request.

Yet in generative AI, engineering teams routinely modify system prompts, upgrade model checkpoints (e.g. from gpt-4o-2024-05-13 to gpt-4o-2024-08-06), or tweak temperature values based on nothing more than 2 manual checks in a playground UI.

Two days later, production error rates spike because the new prompt broke structured JSON outputs for 12% of mobile users.

Evals-Driven Development (EDD) brings software engineering discipline to AI: you define objective assertion suites and run automated regression benchmarks in GitHub Actions before any prompt change merges into production.

---

1. The Anatomy of an LLM Eval

An evaluation suite tests non-deterministic systems using three distinct layers of assertions:

graph TD
    Output[Model Generation] --> L1[Layer 1: Deterministic Schema & Type Assertions]
    Output --> L2[Layer 2: Rule-Based Heuristic Tests (Regex, Length, Blacklist)]
    Output --> L3[Layer 3: Model-as-a-Judge Semantic Scoring (Faithfulness, Coherence)]
    
    L1 --> Pass[Pass / Fail Gate in CI]
    L2 --> Pass
    L3 --> Pass

1. Layer 1 (Deterministic): Did the output conform strictly to the Zod schema? Is it valid JSON? Does it contain required keys?
2. Layer 2 (Heuristic): Did the output avoid blacklisted toxic terms? Is the token length within budget bounds? Did it include mandatory disclaimer strings?
3. Layer 3 (Semantic LLM-as-a-Judge): Given the reference documentation, is the answer strictly faithful (free of hallucinations)?

---

2. Implementing Automated Evals with GitHub Actions

Here is a practical eval runner using TypeScript and Vitest that runs in CI on every pull request:

// evals/prompt-regression.test.ts
import { describe, it, expect } from 'vitest';
import { generateCustomerSummary } from '../src/ai/summarizer';
import goldenDataset from './golden-dataset.json';

describe('Customer Support Summarizer Regression Suite', () => {
  goldenDataset.forEach((fixture) => {
    it(`evaluates test case: ${fixture.id} - ${fixture.scenario}`, async () => {
      const result = await generateCustomerSummary(fixture.transcript);

      // 1. Deterministic checks
      expect(result).toHaveProperty('sentiment');
      expect(['positive', 'neutral', 'negative']).toContain(result.sentiment);
      expect(result.summary.length).toBeLessThan(500);

      // 2. Expected Key Information Extraction
      fixture.mustContainEntities.forEach((entity: string) => {
        expect(result.summary.toLowerCase()).toContain(entity.toLowerCase());
      });

      // 3. Negative Constraint Check
      fixture.forbiddenPhrases.forEach((phrase: string) => {
        expect(result.summary.toLowerCase()).not.toContain(phrase.toLowerCase());
      });
    }, 15000); // 15s timeout for API call
  });
});

---

3. Ragas Metrics for RAG Pipelines

When evaluating RAG systems, relying purely on lexical string matching (like BLEU or ROUGE) is ineffective. Modern pipelines employ Ragas metrics:

• Faithfulness: Quantifies whether every statement in the generated answer can be directly inferred from the retrieved context chunks (catches hallucinations).
• Answer Relevance: Measures whether the response directly addresses the user prompt rather than deflecting.
• Context Precision: Evaluates whether the vector retriever ranked relevant chunks higher than irrelevant noise.

# run_ragas_eval.py
from ragas import evaluate
from ragas.metrics import faithfulness, answer_relevancy, context_precision
from datasets import Dataset

def evaluate_rag_pipeline(eval_data: dict):
    dataset = Dataset.from_dict(eval_data)
    results = evaluate(
        dataset=dataset,
        metrics=[faithfulness, answer_relevancy, context_precision]
    )
    print("Faithfulness Score:", results["faithfulness"])
    assert results["faithfulness"] > 0.90, "Regression: Faithfulness fell below 90% threshold!"

---

4. Key Takeaways

• Never merge prompt changes without automated evals: Treat prompts with the same rigor as database migration scripts.
• Curate 100 Golden Fixtures: Store ground-truth test cases in JSON fixtures within version control.
• Fail Fast in CI: Run low-cost deterministic checks first, followed by tiered semantic evaluations on sample subsets.