LLM Eval Frameworks: Promptfoo and DeepEval

Promptfoo’s documentation describes a configuration-driven workflow: list prompts, providers, variables, test cases, and optional assertions, then run evaluations to compare outputs. Assertions can be deterministic, such as schema or exact-match checks, or model-graded; a test suite can also be reviewed manually. DeepEval’s current documentation centers on Python test cases and metrics and supports end-to-end, component, and trajectory evaluations; its CLI can run tests in CI. Many of its built-in metrics use LLM judges, which can add cost, latency, and judge error. Promptfoo also offers model-graded assertions and comparisons, so neither framework is limited to one metric style.

These frameworks organize test execution and result review; they do not decide what “good” means for a product. Define test inputs that reflect actual use, expected outcomes or rubrics, thresholds that match the risk, and a process for reviewing failures. A model-graded metric is an estimate under its rubric, not a fact. Calibrate important metrics against human judgments and inspect disagreement. Use deterministic checks when the expected result is exact, such as required JSON fields, and human review where context or safety requires interpretation.

CI integration can catch regressions before release, but nondeterministic outputs and model-service changes can make a check noisy. Pin relevant model and framework versions, record run settings, repeat or adjudicate unstable cases, and avoid failing a build on a metric that has not been validated for the task. Preserve representative failures as new regression cases after confirming their expected behavior.