Part 7
Tooling turns evaluation from manual spreadsheets into repeatable practice. This part surveys the popular frameworks, what each is best at, and how to pick the ones that fit your stack.
OpenAI Evals is a framework for building benchmark-driven evaluation suites. You write a YAML or Python registry of eval cases, each with prompts and grading logic (from simple exact-match to LLM-judged checks), then run the suite against any model and compare the output. It is useful for reproducible, shareable model comparisons.
DeepEval provides a broad metric library — answer relevance, faithfulness, hallucination, bias, and more — plus a test harness that reports pass/fail per case. It integrates with pytest so evaluation runs become part of your normal test pipeline, which makes it a strong default for LLM application teams.
from deepeval import assert_test
from deepeval.test_case import LLMTestCase
from deepeval.metrics import AnswerRelevancyMetric
test_case = LLMTestCase(input="...", actual_output="...")
assert_test(test_case, [AnswerRelevancyMetric()])Promptfoo is built for comparing prompts and model outputs side by side. You define prompt variants and test cases in a config file, run them, and get a matrix of outputs you can diff visually. It is excellent for prompt iteration — "which wording scores best?" — before you spend effort on heavier infrastructure.
LangSmith traces LLM application runs, so you can see every prompt, call, and step inside a single request. On top of tracing, it offers dataset-based evaluation and feedback collection from live users. It shines when you need to debug why a metric changed, because the trace shows the exact path a bad answer took.
Phoenix is an open observability platform for AI. It ingests spans and traces from production and offline runs, and provides drift detection, embedding visualization, and evaluation dashboards. Use it when you want to watch production performance over time and dig into anomalies, not just run one-off tests.
MLflow's evaluation API computes standard metrics for models — classification and regression metrics out of the box, plus LLM metrics when a model is set up for it. Because MLflow already tracks experiments and artifacts, evaluation results live next to the runs that produced them, keeping the whole history in one place.
Hugging Face Evaluate is a lightweight library that computes standard metrics (accuracy, F1, BLEU, and hundreds more) with one call. It is the simplest way to get consistent, comparable numbers across experiments and is a natural fit in any Hugging Face workflow.
The EleutherAI lm-evaluation-harness is a benchmark runner for foundation models. It executes dozens of standard benchmarks with a fixed protocol so different models can be compared fairly. If you are choosing between pretrained models or tracking your own model's progress against public baselines, this is the reference tool.
| Framework | Best for | Watch out for |
|---|---|---|
| OpenAI Evals | Reusable benchmark suites | Larger setup effort |
| DeepEval | LLM app metrics in CI | Judge-model dependency |
| Promptfoo | Prompt A/B comparisons | Not a full platform |
| LangSmith | Tracing + evaluation | Vendor lock-in |
| Phoenix | Production observability | Requires instrumentation |
| MLflow | Experiment + eval tracking | LLM eval is newer |
| HF Evaluate | Quick standard metrics | Not a platform |
| lm-eval-harness | Foundation-model benchmarks | Focused on public tasks |