Skip to content

Metrics Reference

RAGhelm evaluates RAG and agent systems as release candidates, not just as isolated model calls. Metrics feed scorecards, manifests, gates, and dashboards.

Retrieval metrics

Recall@k

Fraction of relevant contexts retrieved in the top-k results.

Recall@k = |relevant in top-k| / |total relevant|

MRR

Mean reciprocal rank of the first relevant result.

MRR = 1/|Q| * sum(1/rank_i)

NDCG@k

Ranking quality metric that rewards relevant results near the top and penalizes relevant results lower in the ranking.

Generation metrics

Generation scores use a 1-5 scale.

Metric Description
faithfulness Whether the answer is grounded in retrieved/source material.
relevance Whether the answer addresses the question.
completeness Whether the answer covers required aspects.
overall Weighted quality summary.

Gate usage

Metrics become release-control evidence only when linked to:

  • the target under test
  • eval suite and mode
  • retrieval/generation/judge configuration
  • artifact hashes
  • privacy classification
  • RAGRunManifest and ReadinessScorecard IDs where available

Badge colors

Badge colors are display hints, not release decisions.

Threshold Badge color
>= 0.90 Green #9ece6a
>= 0.80 Yellow #e0af68
< 0.80 Red #f7768e

Per ADR-010, published quality badges must be backed by production scorecard and manifest evidence or fail closed.