Definition
A ranking metric used in RAG evaluation to measure the effectiveness of a retriever by rewarding relevant documents at the top of the results while applying a logarithmic discount to those appearing at lower ranks. It normalizes the score against an Ideal DCG (IDCG), ensuring the value ranges from 0 to 1 regardless of the query's total number of relevant documents.
Unlike Hit Rate or MRR, NDCG considers graded relevance levels rather than just binary success or the position of the first hit.
"A tiered awards podium where the height of the platform (relevance) matters, but the value of the prize drops exponentially the further the winner is seated from the front row."
- Discounted Cumulative Gain (DCG)(Prerequisite)
- Mean Reciprocal Rank (MRR)(Comparison)
- Graded Relevance(Component)
- Retrieval Evaluation(Context)
Conceptual Overview
A ranking metric used in RAG evaluation to measure the effectiveness of a retriever by rewarding relevant documents at the top of the results while applying a logarithmic discount to those appearing at lower ranks. It normalizes the score against an Ideal DCG (IDCG), ensuring the value ranges from 0 to 1 regardless of the query's total number of relevant documents.
Disambiguation
Unlike Hit Rate or MRR, NDCG considers graded relevance levels rather than just binary success or the position of the first hit.
Visual Analog
A tiered awards podium where the height of the platform (relevance) matters, but the value of the prize drops exponentially the further the winner is seated from the front row.