Maximal Marginal Relevance (MMR)

MMR operates by greedily selecting the next document D that maximizes the following score:

MMR = Argmax [ λ * Sim(D, Q) - (1 - λ) * max Sim(D, Di) ]

• λ (Lambda): A hyperparameter between 0 and 1 controlling the trade-off.
• Sim(D, Q): The similarity between a candidate document and the query.
• max Sim(D, Di): The maximum similarity between the candidate and any already-selected document.

When λ=1, the algorithm is a standard relevance-only ranking. When λ=0, it maximizes diversity by selecting documents most dissimilar to everything already chosen.

The key innovation of MMR is the subtractive penalty term that actively suppresses redundant information.

• A candidate document highly relevant to the query but nearly identical to an already-selected document receives a high penalty, reducing its MMR score.
• A slightly less relevant but topically distinct document can outrank it because its penalty term is low.
• This prevents the common failure mode of standard similarity search where the top-k results are near-duplicates of each other, ensuring the final list covers diverse aspects of the query.

MMR builds the result set one document at a time using a greedy selection process, not a global optimization.

• Step 1: The first document selected is always the single most relevant item to the query.
• Step 2: The algorithm iterates, recalculating the MMR score for all remaining candidates against the growing set of chosen documents.
• Trade-off: This greedy approach is computationally efficient (polynomial time) but does not guarantee a globally optimal diverse set. It is a practical heuristic that works well in real-time retrieval systems.

The λ parameter is the single control knob for balancing relevance and diversity, and tuning it is highly application-specific.

• High λ (e.g., 0.7-0.9): Favors relevance. Useful for question-answering where the top document must be correct, and diversity is a secondary concern.
• Low λ (e.g., 0.3-0.5): Favors diversity. Ideal for exploratory search, summarization of large document collections, or product recommendation where showing varied options is critical.
• Practical Tuning: The optimal λ is typically found through offline evaluation on a labeled dataset measuring both precision and subtopic recall.

MMR is a critical post-retrieval step in modern Retrieval-Augmented Generation (RAG) pipelines to manage the context window.

• Context Diversity: A vector store may retrieve 20 semantically similar chunks. Applying MMR selects the top-5 most diverse chunks to fit within the LLM's token limit, providing broader grounding.
• Hallucination Reduction: By forcing the model to synthesize an answer from distinct, non-redundant sources, MMR reduces the risk of the model overfitting to a single repetitive piece of evidence.
• Implementation: LangChain and LlamaIndex provide native MMR vector store retrievers that apply the algorithm directly on embedding similarity scores.

Despite its effectiveness, MMR has distinct limitations compared to modern learned alternatives.

• Linear Similarity Assumption: MMR uses a simple linear combination of cosine similarities. It cannot learn complex, non-linear patterns of redundancy that a cross-encoder reranker might capture.
• Quadratic Complexity: The greedy algorithm requires computing pairwise similarity between the candidate set and the growing selected set, leading to O(n*k) complexity where n is the candidate pool size and k is the number of selections.
• Static Lambda: A single global λ cannot adapt to different query types. A factual query ("capital of France") needs no diversity, while an ambiguous query ("jaguar") needs high diversity. Adaptive MMR variants attempt to predict λ per-query.