Search

Search IconIcon to open search

IR Metrics

Last updatedUpdated: by Jakub Žovák · 1 min read

Properties
created 01.11.2025, 17:59
modified 21.03.2026, 13:30
published Empty
topics Evaluation
authors Empty
ai-assisted No
  • This note aggregates information about different approaches to use for evaluation of the information retrieval performed in order to find relevant context
  • For more info about ranking metrics see Ranking Metrics

# Classical Metrics

  • Mean Average Precision ( MAP)
    • Averages precision across all recall levels and queries—captures both precision and ranking quality
  • Mean Reciprocal Rank ( MRR)
    • Focuses on how high the first relevant result appears in the ranking
  • Normalized Discounted Cumulative Gain ( nDCG)
    • Weighs the usefulness of each result by its position—higher-ranked relevant items count more

# RAG-tailored Metric

# Metrics Comparison

  • Note the below table considers question answering task
  • Table 3: Spearman correlation between IR metrics and end-to-end RAG accuracy. Bold indicates best per model dataset combination. UDCG improvements are statistically significant (bootstrap, p < 0.05) ( Source).