āĻĒā§āϰāϝ⧁āĻ•ā§āϤāĻŋāĻ—āϤ āĻ—āĻžāχāĻĄ

āĻĒ⧁āύāϰ⧁āĻĻā§āϧāĻžāϰ āĻŽā§‡āĻŸā§āϰāĻŋāĻ•ā§āϏ: Recall@K, MRR āĻāĻŦāĻ‚ nDCG

Recall@K measures how many known relevant items appear in the top K results, MRR emphasizes the rank of the first relevant item, and nDCG evaluates ranking quality using graded relevance with position discounts.

āĻāχ āύāĻŋāĻ°ā§āĻĻ⧇āĻļāĻŋāĻ•āĻž āĻļ⧁āύ⧁āύāĻāφāχ-āĻŦāĻ°ā§āĻŖāĻŋāϤ ¡ ~3 min

āĻāχ āĻĒ⧃āĻˇā§āĻ āĻžāϝāĻŧ3 āĻŽāĻŋāύāĻŋāϟ āĻĒāĻĄāĻŧāĻž āĻšāϝāĻŧ⧇āϛ⧇
  1. āĻ“āĻ­āĻžāϰāĻ­āĻŋāω
  2. āĻ—āĻ­ā§€āϰ āĻĄā§āĻŦ
  3. āĻŦāĻžāĻ¸ā§āϤāĻŦ-āĻŦāĻŋāĻļā§āĻŦ āĻŦāĻžāĻ¸ā§āϤāĻŦāĻžāϝāĻŧāύ
  4. The Future of Retrieval Metrics: Recall@K, MRR and nDCG
  5. āĻāϟāĻžāϕ⧇ āĻŦāĻžāĻ¸ā§āϤāĻŦ⧇ āĻĒā§āϰāϝāĻŧā§‹āĻ— āĻ•āϰāĻž āĻšāĻšā§āϛ⧇
  6. āĻ…āĻ¨ā§āĻŦ⧇āώāĻŖ āϚāĻžāϞāĻŋāϝāĻŧ⧇ āϝāĻžāύ
  7. āĻĒā§āϰāĻžāϝāĻŧāĻļāχ āϜāĻŋāĻœā§āĻžāĻžāϏāĻŋāϤ āĻĒā§āϰāĻļā§āύāĻžāĻŦāϞ⧀

āĻ“āĻ­āĻžāϰāĻ­āĻŋāω

Each metric reflects different judgments and use cases, and none directly measures whether a generated answer is correct.

āĻ—āĻ­ā§€āϰ āĻĄā§āĻŦ

Retrieval evaluation starts with queries, a document collection, and relevance judgments. A metric summarizes how a ranked result list compares with those judgments. Recall@K asks what fraction of relevant items appear within the first K results. It is useful when finding all relevant evidence matters, but it ignores the ordering among those retrieved items. Mean reciprocal rank (MRR) scores the rank of the first relevant result: a relevant item at rank one contributes 1, at rank two contributes one-half, and so on; query scores are averaged. This fits tasks where users mainly need one good result near the top. It does not reward additional relevant results after the first. Normalized discounted cumulative gain (nDCG) uses relevance grades and discounts items lower in the ranking. It can distinguish highly relevant from partially relevant documents and rewards placing stronger matches earlier. The exact gain and discount conventions should be documented when comparing implementations. These formulas summarize judged rankings and do not measure answer generation or user satisfaction. These measures depend on a test collection and relevance labels. Incomplete judgments can make a relevant but unjudged result look wrong, and query selection affects conclusions. For retrieval-augmented generation, report retrieval metrics separately from answer-level grounding, correctness, and usefulness. Choose metrics that match the product goal, inspect examples, and evaluate on representative queries rather than selecting whichever score looks largest.

āĻŦāĻžāĻ¸ā§āϤāĻŦ-āĻŦāĻŋāĻļā§āĻŦ āĻŦāĻžāĻ¸ā§āϤāĻŦāĻžāϝāĻŧāύ

A legal search team uses Recall@20 when it needs to surface most known relevant passages for review.

A help center tracks MRR when users need one useful answer near the top.

A recommender uses graded relevance labels and nDCG to favor stronger matches earlier.

A RAG evaluation reports retrieval recall separately from whether the generated answer is supported.

The Future of Retrieval Metrics: Recall@K, MRR and nDCG

Retrieval systems will increasingly combine lexical, vector, hybrid, and reranking stages, making stage-specific evaluation important. Better relevance judgments and task-aligned graded labels can make nDCG more useful, while user studies can test whether rank gains improve outcomes. RAG evaluation will continue to pair retrieval metrics with answer-level checks. Future benchmarks should publish query sets, judgments, metric settings, and limitations to support meaningful comparisons. Teams should revisit judgments as collections and user needs change over time for all relevant users consistently.

āĻāϟāĻžāϕ⧇ āĻŦāĻžāĻ¸ā§āϤāĻŦ⧇ āĻĒā§āϰāϝāĻŧā§‹āĻ— āĻ•āϰāĻž āĻšāĻšā§āϛ⧇

āϕ⧇āύ āĻāϟāĻž āϗ⧁āϰ⧁āĻ¤ā§āĻŦāĻĒā§‚āĻ°ā§āĻŖ

  • āĻ–āϰāϚ āĻ“ āĻŦāĻžāĻœā§‡āϟ. āφāĻ°ā§āĻ•āĻŋāĻŸā§‡āĻ•āϚāĻžāϰ⧇āϰ āϏāĻŋāĻĻā§āϧāĻžāĻ¨ā§āϤāϗ⧁āϞāĻŋ āĻŦāĻ›āϰ⧇āϰ āĻĒāϰ āĻŦāĻ›āϰ āϧāϰ⧇ āĻ•āĻ°ā§āĻŽāĻ•ā§āώāĻŽāϤāĻž āĻāĻŦāĻ‚ āĻ…āĻĒāĻžāϰ⧇āϟāĻŋāĻ‚ āĻ–āϰāϚ āϚāĻžāϞāĻžāϝāĻŧāĨ¤
  • āϏ⧁āĻ¸ā§āĻĒāĻˇā§āϟ āϏāĻŋāĻĻā§āϧāĻžāĻ¨ā§āϤ. āĻ•āĻžāϰāĻŋāĻ—āϰāĻŋ āĻļāĻŋāĻ•ā§āώāĻž āĻĻāϞāϗ⧁āϞ⧋āϕ⧇ āϏāĻ āĻŋāĻ• āĻ¸ā§āĻŸā§āϝāĻžāĻ• āĻŦ⧇āϛ⧇ āύāĻŋāϤ⧇ āϏāĻžāĻšāĻžāĻ¯ā§āϝ āĻ•āϰ⧇, āĻļ⧁āϧ⧁ āύāϤ⧁āύāϟāĻŋ āύāϝāĻŧāĨ¤
  • āĻŽāĻžāύ āύāĻŋāϝāĻŧāĻ¨ā§āĻ¤ā§āϰāĻŖ. āĻ­āĻžāϞ⧋ āχāĻžā§āϜāĻŋāύāĻŋāϝāĻŧāĻžāϰāĻŋāĻ‚ āĻĒāĻ›āĻ¨ā§āĻĻ āĻ‰ā§ŽāĻĒāĻžāĻĻāύ⧇ āύāĻŋāĻ°ā§āĻ­āϰāϝ⧋āĻ—ā§āϝāϤāĻžāϰ āϘāϟāύāĻž āĻ•āĻŽāĻŋāϝāĻŧ⧇ āĻĻ⧇āϝāĻŧāĨ¤

āϜāĻ¨ā§āϝ āϏāϤāĻ°ā§āĻ•

  • āĻāĻ•āϟāĻŋ āĻŦ⧇āĻžā§āϚāĻŽāĻžāĻ°ā§āĻ• āĻ…āĻĒā§āϟāĻŋāĻŽāĻžāχāϜ āĻ•āϰāĻž āĻŦ⧃āĻšāĻ¤ā§āϤāϰ āϏāĻŋāĻ¸ā§āĻŸā§‡āĻŽ āĻĻ⧁āĻ°ā§āĻŦāϞāϤāĻž āφāĻĄāĻŧāĻžāϞ āĻ•āϰāϤ⧇ āĻĒāĻžāϰ⧇āĨ¤
  • āĻ…āĻŦāĻ•āĻžāĻ āĻžāĻŽā§‹ āĻāĻŦāĻ‚ āϰāĻ•ā§āώāĻŖāĻžāĻŦ⧇āĻ•ā§āώāϪ⧇āϰ āĻ–āϰāϚ āĻĒā§āϰāĻžāϝāĻŧāχ āĻ…āĻŦāĻŽā§‚āĻ˛ā§āϝāĻžāϝāĻŧāύ āĻ•āϰāĻž āĻšāϝāĻŧāĨ¤
  • āϏāĻŋāĻ¸ā§āĻŸā§‡āĻŽāϗ⧁āϞāĻŋ āφāϰāĻ“ āϜāϟāĻŋāϞ āĻšāĻ“āϝāĻŧāĻžāϰ āϏāĻžāĻĨ⧇ āϏāĻžāĻĨ⧇ āϏ⧁āϰāĻ•ā§āώāĻž āĻāĻŦāĻ‚ āĻĒāĻ°ā§āϝāĻŦ⧇āĻ•ā§āώāĻŖāϝ⧋āĻ—ā§āϝāϤāĻžāϰ āĻĢāĻžāρāĻ• āĻŦāĻžāĻĄāĻŧāϤ⧇ āĻĒāĻžāϰ⧇āĨ¤

āĻĒāϰāĻŦāĻ°ā§āϤ⧀ āĻĒāĻĻāĻ•ā§āώ⧇āĻĒ

  1. āĻŦāĻžāĻ¸ā§āϤāĻŦāĻžāϝāĻŧāύ⧇āϰ āφāϗ⧇ āĻŦāĻŋāϞāĻŽā§āĻŦ, āϗ⧁āĻŖāĻŽāĻžāύ āĻāĻŦāĻ‚ āĻ–āϰāĻšā§‡āϰ āϞāĻ•ā§āĻˇā§āϝ āύāĻŋāĻ°ā§āϧāĻžāϰāĻŖ āĻ•āϰ⧁āύāĨ¤
  2. āĻŦāĻžāĻ¸ā§āϤāĻŦāϏāĻŽā§āĻŽāϤ āϞ⧋āĻĄ āĻāĻŦāĻ‚ āĻĄā§‡āϟāĻž āĻ…āĻŦāĻ¸ā§āĻĨāĻžāϰ āĻ…āϧ⧀āύ⧇ āĻŦ⧇āĻžā§āϚāĻŽāĻžāĻ°ā§āĻ•āĨ¤
  3. āĻ¤ā§āϰ⧁āϟāĻŋ, āĻĒā§āϰāĻŦāĻžāĻš, āĻāĻŦāĻ‚ āĻŦā§āϝāĻŦāĻšāĻžāϰāĻ•āĻžāϰ⧀āϰ āĻĒā§āϰāĻ­āĻžāĻŦ⧇āϰ āϜāĻ¨ā§āϝ āϝāĻ¨ā§āĻ¤ā§āϰ āĻĒāĻ°ā§āϝāĻŦ⧇āĻ•ā§āώāĻŖāĨ¤
  4. āĻ¸ā§āϕ⧇āϞ āĻ•āϰāĻžāϰ āφāϗ⧇ āϰ⧋āϞāĻŦā§āϝāĻžāĻ• āĻāĻŦāĻ‚ āϘāϟāύāĻžāϰ āĻĒā§āϰāϤāĻŋāĻ•ā§āϰāĻŋāϝāĻŧāĻž āĻĒāĻžāĻĨ āĻĒā§āϰāĻ¸ā§āϤ⧁āϤ āĻ•āϰ⧁āύāĨ¤

āĻ…āĻ¨ā§āĻŦ⧇āώāĻŖ āϚāĻžāϞāĻŋāϝāĻŧ⧇ āϝāĻžāύ

āĻĢā§āϰāĻŋ āύāĻŋāωāϜāϞ⧇āϟāĻžāϰ

āĻĒā§āϰāϤāĻŋāĻĻāĻŋāύ⧇āϰ āĻāφāχ āĻŦā§āϰāĻŋāĻĢāĻŋāĻ‚ āĻĒāĻžāύ

āĻĒā§āϰāϤāĻŋ āϏāĻĒā§āϤāĻžāĻšā§‡āϰ āĻ•āĻžāĻœā§‡āϰ āĻĻāĻŋāύ āϏāĻ•āĻžāϞ⧇ āϤāĻŋāύāϟāĻŋ āϝāĻžāϚāĻžāχāĻ•ā§ƒāϤ āĻāφāχ āĻ¸ā§āĻŸā§‹āϰāĻŋ, āϏāĻšāϜ āχāĻ‚āϰ⧇āϜāĻŋāϤ⧇ āϞ⧇āĻ–āĻžāĨ¤ āϚāĻŋāϰāĻ•āĻžāϞ⧇āϰ āϜāĻ¨ā§āϝ āĻĢā§āϰāĻŋ, āϕ⧋āύ⧋ āĻŦāĻŋāĻœā§āĻžāĻžāĻĒāύ āύ⧇āχāĨ¤

āĻĒā§āϰāϤāĻŋ āϏāĻĒā§āϤāĻžāĻšā§‡āϰ āĻ•āĻžāĻœā§‡āϰ āĻĻāĻŋāύ⧇ āĻāĻ•āϟāĻŋ āχāĻŽā§‡āϞāĨ¤ āĻāĻ• āĻ•ā§āϞāĻŋāϕ⧇āχ āφāύāϏāĻžāĻŦāĻ¸ā§āĻ•ā§āϰāĻžāχāĻŦ āĻ•āϰ⧁āύāĨ¤ āφāĻŽāϰāĻž āĻ•āĻ–āύ⧋āχ āφāĻĒāύāĻžāϰ āĻ āĻŋāĻ•āĻžāύāĻž āĻŦāĻŋāĻ•ā§āϰāĻŋ āĻŦāĻž āĻļā§‡ā§ŸāĻžāϰ āĻ•āϰāĻŋ āύāĻžāĨ¤

āύāĻŋāĻœā§‡āϕ⧇ āĻĒāϰ⧀āĻ•ā§āώāĻž āĻ•āϰ⧁āύ

Take the Retrieval Metrics: Recall@K, MRR and nDCG quiz

āĻĒā§āϰāϤāĻŋāϟāĻŋ āωāĻ¤ā§āϤāϰ⧇āϰ āϤāĻžā§ŽāĻ•ā§āώāĻŖāĻŋāĻ• āĻĢāĻŋāĻĄāĻŦā§āϝāĻžāĻ•, āĻāĻŦāĻ‚ āϕ⧋āĻ°ā§āϏ āĻĒāĻžāϏ āĻ•āϰāĻžāϰ āĻĒāϰ āϝāĻžāϚāĻžāχāϝ⧋āĻ—ā§āϝ āφāχāĻĄāĻŋ āϏāĻš āĻāĻ•āϟāĻŋ āĻļā§‡ā§ŸāĻžāϰāϝ⧋āĻ—ā§āϝ āϏāĻžāĻ°ā§āϟāĻŋāĻĢāĻŋāϕ⧇āϟāĨ¤

āϕ⧁āχāϜ āĻļ⧁āϰ⧁ āĻ•āϰ⧁āύ

āĻĢā§āϰāĻŋ āĻāφāχ āĻļāĻŋāĻ•ā§āώāĻžāϕ⧇ āϏāĻŽāĻ°ā§āĻĨāύ āĻ•āϰ⧁āύāĨ¤ AI Understanding āĻāĻ•āϟāĻŋ 501(c)(3) āĻ…āϞāĻžāĻ­āϜāύāĻ• āĻĒā§āϰāϤāĻŋāĻˇā§āĻ āĻžāύ — āϕ⧋āύ⧋ āĻŦāĻŋāĻœā§āĻžāĻžāĻĒāύ āύ⧇āχ, āϕ⧋āύ⧋ āĻĒ⧇āĻ“ā§ŸāĻžāϞ āύ⧇āχ, āĻ•āĻ–āύ⧋āχ āύāĻžāĨ¤ āĻāĻ•āϟāĻŋ āĻĻāĻžāύ āĻ•āϰ⧁āύ

āĻĒā§āϰāĻžāϝāĻŧāĻļāχ āϜāĻŋāĻœā§āĻžāĻžāϏāĻŋāϤ āĻĒā§āϰāĻļā§āύāĻžāĻŦāϞ⧀

What is Retrieval Metrics: Recall@K, MRR and nDCG?

Recall@K measures how many known relevant items appear in the top K results, MRR emphasizes the rank of the first relevant item, and nDCG evaluates ranking quality using graded relevance with position discounts. Each metric reflects different judgments and use cases, and none directly measures whether a generated answer is correct.

What does Recall@K measure?

Recall at a cutoff measures relevant-item coverage in the top K.

What does Recall@K fail to distinguish by itself?

Recall@K measures coverage at the cutoff, not ordering within those results.

Why can incomplete relevance judgments distort retrieval metrics?

Metrics depend on the judged collection and can penalize unjudged positives.

Does a high retrieval Recall@K prove that a RAG answer is correct?

Retrieval and answer generation are distinct evaluation stages.

āĻļāĻŋāĻ–āϤ⧇ āĻĨāĻžāϕ⧁āύ

āϏāĻŽā§āĻĒāĻ°ā§āĻ•āĻŋāϤ āĻ—āĻžāχāĻĄ

āĻāχ āĻŦāĻŋāώāϝāĻŧ⧇āϰ āϜāĻ¨ā§āϝ āĻŦāĻžāĻ›āĻžāχ āĻ•āϰāĻž āφāϰāĻ“ āĻ—āĻžāχāĻĄ

āĻĒāϰāĻŦāĻ°ā§āϤ⧀ āφāĻĒāĻĒāϰāĻŦāĻ°ā§āϤ⧀ āĻ—āĻžāχāĻĄ

āĻāύāĻĄāĻŋāϏāĻŋāϜāĻŋ āĻāĻŦāĻ‚ āĻ°â€Œā§āϝāĻžāĻ™ā§āĻ•āĻŋāĻ‚ āĻŽā§‡āĻŸā§āϰāĻŋāĻ•ā§āϏ

āĻŸā§‡āĻ•āύāĻŋāĻ•ā§āϝāĻžāϞ

āĻāĻ•āϟāĻŋ āĻŦāĻŋāύāĻžāĻŽā§‚āĻ˛ā§āϝ⧇ AI āĻļāĻ‚āϏāĻžāĻĒāĻ¤ā§āϰ āĻĒāĻžāύāĻāĻ•āϟāĻŋ āϏāĻ‚āĻ•ā§āώāĻŋāĻĒā§āϤ āĻŦāĻŋāύāĻžāĻŽā§‚āĻ˛ā§āϝ⧇ āϕ⧋āĻ°ā§āϏ āĻāĻŦāĻ‚ āĻĒāϰ⧀āĻ•ā§āώāĻž. āϕ⧋āύ⧋ āĻ–āϰāϚ āύ⧇āχ, āϕ⧋āύ⧋ āĻ•āĻžāĻ°ā§āĻĄ āύ⧇āχāĨ¤ āĻĒā§āϰāĻžāϝāĻŧ 2.5 āϘāĻ¨ā§āϟāĻžāĨ¤āĻŦāĻŋāύāĻžāĻŽā§‚āĻ˛ā§āϝ⧇ āĻļ⧁āϰ⧁ āĻ•āϰ⧁āύ →