Definition
NDCG@K (Normalized Discounted Cumulative Gain at Rank K) is an evaluation metric for RAG retrieval stages that measures the quality of document ranking by penalizing relevant documents found at lower positions using a logarithmic decay. Optimizing for high NDCG@K often requires a trade-off between retrieval latency and accuracy, typically involving the introduction of a compute-heavy Cross-Encoder re-ranker.
Focuses on the quality of result ordering, whereas Recall@K only tracks the presence of relevant items.
"A tiered sports podium where the score of the athlete at the top is exponentially more valuable to the team's total than those in lower spots."
Conceptual Overview
NDCG@K (Normalized Discounted Cumulative Gain at Rank K) is an evaluation metric for RAG retrieval stages that measures the quality of document ranking by penalizing relevant documents found at lower positions using a logarithmic decay. Optimizing for high NDCG@K often requires a trade-off between retrieval latency and accuracy, typically involving the introduction of a compute-heavy Cross-Encoder re-ranker.
Disambiguation
Focuses on the quality of result ordering, whereas Recall@K only tracks the presence of relevant items.
Visual Analog
A tiered sports podium where the score of the athlete at the top is exponentially more valuable to the team's total than those in lower spots.