Unit-Test Your LLM App in 5 Minutes: DeepEval's Answer Relevancy, Faithfulness, and G-Eval

TL;DR: DeepEval brings pytest-style unit testing to LLM apps — score answer relevancy, faithfulness, and custom criteria with an LLM judge in under 15 lines of Python.

What you'll be able to do after this:

• Write a test case that automatically scores your LLM's response on answer relevancy and faithfulness
• Define a custom G-Eval metric for any criterion your standard benchmarks don't cover (tone, format, safety)
• Run your eval suite in CI to catch output regressions before they reach users

---

Your LLM "worked in testing" often means "you read some outputs and they looked fine." That doesn't scale. DeepEval is an open-source evaluation framework that works like pytest for LLM outputs — define test cases, pick metrics, run deepeval test run. The metrics use an LLM judge internally, so they handle nuanced quality criteria that string matching can't.

Install:

pip install -U deepeval
export OPENAI_API_KEY=sk-...   # used as the default judge model (GPT-4o); swap via config

Write your first test case:

# test_my_llm.py
from deepeval import evaluate
from deepeval.test_case import LLMTestCase, LLMTestCaseParams
from deepeval.metrics import AnswerRelevancyMetric, FaithfulnessMetric, GEval

test_case = LLMTestCase(
    input="What are the benefits of prompt caching?",
    actual_output=your_llm_response,         # the string you're testing
    retrieval_context=["Prompt caching stores the KV cache..."]  # for RAG; omit if not RAG
)

Run built-in metrics:

relevancy   = AnswerRelevancyMetric(threshold=0.7)
faithfulness = FaithfulnessMetric(threshold=0.5)

evaluate([test_case], [relevancy, faithfulness])

AnswerRelevancyMetric checks whether the response actually answers the question. FaithfulnessMetric checks whether claims are grounded in the provided context — it flags hallucinated facts. Both score 0–1; the threshold sets pass/fail.

Add a custom G-Eval metric for anything else:

When no built-in metric fits, G-Eval lets you write the criterion in plain English. It uses chain-of-thought prompting to reason before scoring:

tone_check = GEval(
    name="Professional Tone",
    criteria="Determine if the response is professional and avoids colloquial language.",
    evaluation_params=[LLMTestCaseParams.ACTUAL_OUTPUT],
    threshold=0.6
)
evaluate([test_case], [tone_check])

Run in pytest (CI-ready):

deepeval test run test_my_llm.py

Each failed metric prints the judge's reasoning, not just a score — so you see why it failed. Gate your CI on eval pass rate the same way you gate it on unit tests.

Sources: DeepEval 5-min Quickstart · G-Eval metric · AnswerRelevancy metric · Faithfulness metric