MMR Diversification
Diversify search results by balancing relevance with novelty to reduce redundancy
Why do anything?
Vector search returns similar results that cluster together. Users want diverse, non-redundant results.
Why now?
Redundant results waste user attention. Diversity improves user experience.
Why this feature?
MMR algorithm balances relevance (similarity to query) with diversity (dissimilarity to already-selected results).
How It Works
MMR provides result diversification by penalizing similarity to already-selected documents.
1
Candidate Scoring
Score all candidates by query relevance
2
Iterative Selection
Select documents maximizing marginal relevance
3
Diversity Balance
Lambda controls relevance-diversity tradeoff
Why This Approach
MMR is a proven algorithm for result diversification. Lambda parameter enables tuning for different use cases.
Where This Is Used
Integration
retriever = client.retrievers.create(stages=[{"type": "mmr", "lambda": 0.5}])