Skip to content
FDE 101

Maximal marginal relevance

Maximal marginal relevance, or MMR, selects results by balancing relevance to the query with novelty compared with items already selected.1

Tiny example

Section titled “Tiny example”

Ordinary top four results may contain three copies of the same refund paragraph. MMR can keep one copy and promote the enterprise-exception passage, giving the model more varied evidence.

What it cannot do

Section titled “What it cannot do”

MMR cannot recover a passage missing from the candidate set. Too much diversity can also promote a less relevant item. It reduces redundancy; it does not verify factual accuracy.

Footnotes

Section titled “Footnotes”

  1. Goldstein and Carbonell, “Using MMR for Diversity-Based Reranking”, 1998. ↩