Maximum Marginal Relevance (MMR)
MMR is a re-ranking strategy that balances relevance and diversity — ensuring “items returned are both relevant and sufficiently varied to address all facets of the query.”
Formula
MMR score = λ × Sim(document, query) − (1−λ) × max Sim(document, previously_selected)
- λ → 1: prioritizes relevance
- λ → 0: emphasizes diversity
How it works (iterative selection)
For each iteration, pick the candidate with the highest MMR score — balancing similarity to the query against similarity to already-selected documents.
Example: nodes [N2, N3, N1] for query Q1:
- Select N1 first (highest relevance: 0.9)
- Compute MMR scores for N2, N3 penalizing similarity to N1
- Select N3
- Final order: [N1, N3, N2]
Key benefit
Minimizes redundancy — the top N results provide meaningful variety, eliminating clustering of similar items. Enables comprehensive answers addressing multiple query dimensions.