SHEN YIFull-Stack & AI

Paris • Singapore • Shanghai

RAG Formation · Reference

Trustworthy RAG evaluation sets

A standalone cheat sheet for learning and testing RAG quality on this site.

$ reference --status
▸ course RAG Formation
▸ mode standalone reference
▸ status Complete
● learn → test → verify

Reference sheet

Trustworthy RAG evaluation sets

The guided RAG Formation course content is complete; this reference is available for self-directed practice.

The evaluation tuple

Question + category + expected sources + required facts + expected behavior

For an answerable question, the system should give a correct answer grounded in the evidence. For an unanswerable question, it should refuse.

What a good set tests

DimensionWhy it mattersMeridian demo
Factual lookupFind a precise number, name, date, or requirement.35 questions
Reference lookupFind the exact SOP or document identifier.7 questions
Multi-hopPull together evidence from more than one document.9 questions
Cross-lingualMatch a question with a source written in another language.5 questions
AcronymResolve domain terms and abbreviations.3 questions
UnanswerableMeasure whether the system refuses instead of rewarding a hallucination.6 questions

Ground truth rules

  1. Every expected source must be present in the corpus.
  2. Every required fact must appear literally in an expected source.
  3. An unanswerable question must have neither an expected source nor a required fact.
  4. Match retrieved sources by metadata, not by arbitrary text in a chunk.
  5. Keep the evaluation set unchanged while comparing system changes.

Bad measurement

Count a vector result as a hit just because its chunk text mentions the expected document ID.

Better measurement

Count it as a hit only when the source metadata identifies the expected document.

Trust boundary

An evaluation harness can be wrong and still print a confident-looking percentage. Test the scorer itself with accents, multi-source questions, body-text cross-references, missing citations, and refusals.

Continue learning: Open the matching standalone lesson · Course roadmap