Main rule
Treat AI output like a product behavior that must be tested, measured, and regressed safely.
LLM Evaluation Cheat Sheet
A testing-first reference for evaluating LLM behavior, including correctness, safety, groundedness, and regression tracking.
Use it for
Prompt validation, RAG quality, hallucination checks, and AI workflow regression in CI.
Default habit
Keep a small evaluation dataset before trusting improvements.
Pair it with
Use this with AI basics, RAG, and AI test generation.
What to Evaluate
Evaluation starts with explicit quality dimensions.
- correctness - relevance - completeness - groundedness - safety - consistency
Hallucination Checks
Ask whether the model invented facts or exceeded the available evidence.
Good check: Can this answer be supported by the source context? If not, mark it as hallucinated or unsupported.
Structured Output Validation
Evaluate whether the response matches the expected schema as well as the meaning.
Check: - required fields exist - field types are correct - enums are valid - business rules still hold
RAG Evaluation
RAG systems need retrieval and answer quality checks, not just final-output review.
Key checks: - context relevance - groundedness - answer relevance
DeepEval
Useful for test-style LLM evaluations embedded into code-based workflows.
Good for: - metric-based evaluation - regression suites - code-driven checks - CI integration
PromptFoo
Useful for prompt comparison, model comparison, and output quality benchmarking.
Good for: - comparing prompts - comparing models - scoring answer quality - config-driven runs
Trulens
Strong when you need observability and deeper RAG-specific introspection.
Good for: - RAG tracing - retrieval debugging - groundedness analysis - pipeline observability
Evaluation Dataset
Even a small curated set of examples is better than trusting vibes.
Include: - representative happy paths - edge cases - known failure cases - tricky ambiguous inputs - unsafe prompts
CI Integration
Run evals like any regression suite and fail builds only on meaningful thresholds.
- nightly evaluation run - prompt change checks - score thresholds - trend tracking - artifact output for review