Ranking Metrics Playground

Enter the relevance grade of each retrieved item in the order your system ranked them, and see NDCG, Precision, Recall, MAP, MRR and Hit Rate computed step by step. A common bug is computing the ideal DCG from the retrieved list instead of from all relevant items, which hides what your system missed. Fill in the ground-truth judgments to see the difference.

Retrieved relevances (ranked order) Graded relevance of each returned item, top result first. 0 = not relevant.

Ground-truth judgments (optional) Grades of every relevant item for this query, including ones the system missed.

Gain

Edge cases

Metrics

Cumulative DCG vs ideal

your ranking ideal ranking

Step-by-step DCG

rank	relevance	gain	discount 1/log₂(i+1)	DCG contribution	cumulative DCG	ideal relevance	ideal cumulative

Definitions and conventions

Metric	Definition	Edge case handling
DCG@k	Σ gain(relᵢ) / log₂(i+1)	Positions past the list end count as 0
Gain	exponential: 2^rel − 1 · linear: rel	scikit-learn uses linear; TREC and LightGBM use exponential
IDCG@k	DCG@k of the judgments sorted descending	Uses all judged items, not only retrieved ones
NDCG@k	DCG@k / IDCG@k	0 when IDCG = 0 (some libraries return 1 or NaN)
Precision@k	relevant in top-k / k	Divides by k even if fewer than k items were returned
Recall@k	relevant in top-k / total relevant	0 when nothing is relevant
AP@k	Σ P@i·relᵢ / min(k, total relevant)	The normaliser differs across libraries
RR@k	1 / rank of first relevant item	0 if none in top-k

An item counts as relevant for the binary metrics when its grade is > 0. Everything runs in your browser. The reference Python implementation (metrics.py) is tested against sklearn.metrics.ndcg_score and kept in parity with this page's JavaScript.