Maximal Marginal Relevance (MMR)

MMR formalizes the fundamental tension in result ranking. It selects the next document d_i from a candidate set R that maximizes a weighted combination of two scores:

• Relevance Score: How well d_i matches the original query Q, typically computed via cosine similarity with a query embedding or a score like BM25.
• Novelty Score: How dissimilar d_i is from all documents S already selected in the result set, measured by maximal similarity to any document in S.

The algorithm explicitly prevents the top results from being near-duplicates of each other, ensuring each selected item adds new information.

The lambda (λ) parameter is a tunable hyperparameter between 0 and 1 that controls the balance between relevance and diversity in the MMR equation: MMR(d_i) = λ * Sim1(d_i, Q) - (1 - λ) * max_{d_j in S} Sim2(d_i, d_j)

• λ = 1.0: Pure relevance ranking. Diversity is ignored, reverting to a standard similarity search.
• λ = 0.5: Equal weight to relevance and novelty. A common default for general-purpose diversification.
• λ = 0.0: Pure diversity ranking. The query is ignored; results are selected solely to be different from each other.

Engineering teams tune λ based on the use case: lower for exploratory search, higher for fact-finding missions.

MMR uses a greedy, iterative selection process because the optimal set selection is NP-hard. The algorithm proceeds as follows:

1. Start with an empty result set S and a full candidate set R.
2. Select the document from R with the highest relevance score to Q and move it to S.
3. For each remaining position k:
   • For every document d_i in R, compute its MMR score.
   • Select the document d_i with the highest MMR score.
   • Remove d_i from R and add it to S.
4. Repeat step 3 until S reaches the desired size.

This greedy approach is computationally efficient, requiring O(kn) similarity computations for k results and n candidates.

MMR is agnostic to the specific similarity functions used for Sim1 (query-document) and Sim2 (document-document). This allows integration with various retrieval backends:

• Sparse Retrieval: Use BM25 or TF-IDF for Sim1 and Sim2.
• Dense Retrieval: Use cosine similarity between vector embeddings (e.g., from Sentence-BERT).
• Hybrid Retrieval: Use a weighted combination of sparse and dense scores for Sim1.
• Cross-Encoder Reranking: Use a more accurate but slower neural model for final Sim1 scoring after an initial fast retrieval.

The choice of similarity function directly impacts the semantic definition of both "relevance" and "novelty."

In Retrieval-Augmented Generation (RAG) and autonomous agent workflows, MMR is critical for context window management.

• Preventing Redundancy: When constructing a context window from multiple retrieved documents, MMR ensures each chunk provides unique information, maximizing the utility of limited model tokens.
• Mitigating Bias: Without MMR, a retriever might return several highly similar documents that reinforce a single perspective or fact, potentially biasing the LLM's synthesis.
• Supporting Multi-Step Reasoning: For agents, diverse context chunks can contain different pieces of evidence required for complex, multi-faceted reasoning, as opposed to repetitive evidence for a single sub-task.

The naive MMR implementation has a complexity of O(kn) for k results and n candidates, with each step requiring a similarity computation against all items in S. For large-scale systems, this can be optimized:

• Pre-computed Document-Document Similarity: For static corpora, a pre-computed similarity matrix can cache Sim2 values, trading memory for speed.
• Approximate Nearest Neighbor (ANN) Search: Use vector indexes like HNSW for fast retrieval of top-relevance candidates (Sim1), reducing n in the initial candidate set R.
• Cluster-Based Pruning: First cluster candidates, then apply MRR to select diverse clusters before selecting within clusters.
• Threshold-Based Novelty: Ignore the max operation if similarity to any doc in S is below a threshold, speeding up the calculation.

The foundational preprocessing step for MMR. Semantic chunking segments documents into coherent units based on contextual meaning and topic boundaries, rather than arbitrary character counts. This creates the candidate pool of passages that MMR will later rank and diversify.

• Purpose: To produce chunks where internal content is semantically cohesive, improving the quality of retrieved text for language models.
• Contrast with MMR: Chunking creates the items; MMR selects and orders a subset of those items for final output.

A retrieval strategy that MMR can be applied to post-fusion. Hybrid search combines results from:

• Sparse retrieval (e.g., BM25) for exact keyword matching.
• Dense retrieval (e.g., vector similarity) for semantic understanding.

The combined, ranked list is then an ideal candidate for MMR processing to reduce redundancy among the top results from both retrieval methods, ensuring the final set is both relevant and diverse.

A technique often used before retrieval to improve recall, which changes the input to MMR. Query expansion augments a user's original search query with additional related terms or phrases.

• Effect on MMR: By retrieving a broader set of initially relevant documents, the pool for MMR's diversity selection becomes richer and potentially more varied.
• Synergy: Query expansion casts a wider net; MMR refines the catch to avoid repetitive results.

The core infrastructure that enables the similarity calculations MMR relies on. A dense vector index (e.g., in a vector database) stores high-dimensional embeddings of text chunks.

• MMR's Dependency: To compute the "marginal relevance" of a new chunk, MMR must efficiently find its similarity to already-selected chunks. This requires fast Approximate Nearest Neighbor (ANN) search provided by indices like HNSW or IVF.
• Performance: The efficiency of the underlying index directly impacts the latency of an MMR reranking step.

An alternative ranking method to MMR for combining multiple result lists. RRF is a simple, score-free method used in hybrid search:

• It takes ranked lists from different retrievers (e.g., keyword and vector).
• It assigns a score based on the reciprocal of the sum of each item's rank across lists.
• It reranks all items based on this fused score.

Key Difference from MMR: RRF focuses on aggregating rankings from multiple sources but does not explicitly optimize for diversity within the final list, which is MMR's primary objective.

The framework for measuring algorithms like MMR. Key metrics assess the trade-off MMR manages:

• Precision@K: Measures the fraction of relevant items in the top K results. MMR aims to maintain high precision while introducing diversity.
• Recall: Measures the fraction of all relevant items retrieved. MMR typically operates on a pre-retrieved set, so it doesn't directly affect recall.
• Novelty & Diversity: Metrics like Alpha-nDCG or Intent-Aware Metrics explicitly measure the diversity of subtopics covered in a result set, which is what MMR is designed to maximize.