MMR Retrieval Lab

The Problem

Rank by relevance alone — your top results repeat themselves.

Query "London" across 21 documents. Drag λ and watch MMR trade relevance for diversity in real time.

Lambda (λ) — Relevance vs Diversity

0.70 production sweet spot

← most diverse most relevant →

Results (k):

Result Metrics

Max Pair Sim

0.064

base 0.654

Redundancy

0.048

base 0.128

Mean Relevance

0.220

base 0.227

Diversity

0.952

base 0.872

−90%

worst-case redundancy removed vs pure-relevance at this λ

MMR = λ · relevance(doc, query)
− (1−λ) · max_sim(doc, selected)

At each step, pick the document that's most relevant to the query and least similar to what's already been selected. Computed live in your browser — no backend.

Carbonell & Goldstein 1998 · TF-IDF embeddings
Full Python implementation + benchmarks in the repo