Catch LLM Regressions Before Your Users Do: Unit Testing AI Outputs with deepeval

deepeval brings pytest-style unit testing to LLM outputs — write a threshold assertion, run it in CI, catch regressions before they ship.

What you'll be able to do after this:

• Write LLM unit tests that assert on answer quality, faithfulness, and hallucination — not just string equality
• Apply 50+ research-backed metrics (GEval, answer relevancy, contextual faithfulness) to any Claude or OpenAI output
• Wire deepeval into CI so quality regressions fail the build before reaching production

Why testing LLM outputs matters

Once you've built a RAG system or agent, the hardest part isn't getting it to work — it's knowing when it stopped. A model upgrade, a prompt tweak, or a retrieval change can silently degrade answers. deepeval catches those regressions with the same discipline you apply to unit tests.

Install

pip install -U deepeval

Write your first test (test_agent.py)

from deepeval import assert_test
from deepeval.test_case import LLMTestCase, SingleTurnParams
from deepeval.metrics import GEval, AnswerRelevancyMetric

def test_answer_quality():
    # Call your LLM/agent here and capture the output
    output = my_agent.answer("What is the capital of France?")

    test_case = LLMTestCase(
        input="What is the capital of France?",
        actual_output=output,
        expected_output="Paris"
    )

    correctness = GEval(
        name="Correctness",
        criteria="Is the actual output factually correct based on the expected output?",
        evaluation_params=[SingleTurnParams.ACTUAL_OUTPUT, SingleTurnParams.EXPECTED_OUTPUT],
        threshold=0.8
    )
    relevancy = AnswerRelevancyMetric(threshold=0.7)

    assert_test(test_case, [correctness, relevancy])

Run it

deepeval test run test_agent.py

The output shows per-metric scores, reasons for failures, and a pass/fail verdict. If actual_output scores below the threshold, the test fails — exactly like a pytest assertion.

Add RAG-specific metrics

from deepeval.metrics import FaithfulnessMetric, ContextualRelevancyMetric

test_case = LLMTestCase(
    input="What's our return policy?",
    actual_output=rag_output,
    retrieval_context=retrieved_chunks  # the passages from your vector store
)

assert_test(test_case, [
    FaithfulnessMetric(threshold=0.8),        # output doesn't hallucinate beyond context
    ContextualRelevancyMetric(threshold=0.7)  # retrieved context was relevant to the question
])

Wire into CI

# .github/workflows/eval.yml
- name: Run LLM eval
  run: deepeval test run tests/

Any test falling below its threshold breaks the build. Add deepeval login to push results to Confident AI for trend tracking across runs — useful for catching slow quality drift that no single test-run exposes.

Sources: deepeval Getting Started · deepeval — The LLM Evaluation Framework · Using DeepEval for LLM Evaluation in Python — Codecademy