Category: AI Search & Retrieval
Definition
Maximal Marginal Relevance (MMR) is a retrieval and ranking technique that selects results by balancing two factors:
- Relevance to the user’s query
- Difference from results that have already been selected
The goal is to return results that are highly relevant without filling the result set with multiple documents that say essentially the same thing.
Why It Matters
Traditional similarity search can sometimes return several highly similar results.
For example, a search for:
“best project management software”
might retrieve ten articles that all discuss the same popular software products.
While each article may individually be relevant, the overall result set may lack diversity.
MMR attempts to solve this by rewarding relevance while penalizing excessive similarity to already-selected results.
This can produce a more useful and varied set of retrieved information.
How It Works
A simplified MMR objective can be represented as:
MMR = relevance − redundancy
More formally, the system considers the relevance of a candidate to the query while also considering how similar that candidate is to documents that have already been selected.
A higher relevance score helps a document get selected.
A high similarity to an already-selected document can reduce its priority.
The process continues until the desired number of results has been chosen.
Example
Suppose a retrieval system has these candidates:
- Article A — highly relevant to the query
- Article B — highly relevant but almost identical to Article A
- Article C — highly relevant and covers a different aspect
- Article D — moderately relevant but provides unique information
A basic similarity ranking might select:
A → B → C
MMR may instead select:
A → C → D
because C and D add information that is less redundant with A.
MMR in AI Retrieval
MMR can be particularly useful when an AI system needs to assemble a collection of passages or documents to use as context.
If the retrieved context contains five passages that repeat the same information, the system may waste valuable context space.
A more diverse selection could provide:
- A definition
- A practical example
- Supporting evidence
- A contrasting viewpoint
- A relevant technical explanation
This can give a downstream language model a broader evidence base.
MMR and AI Visibility
MMR does not mean that content should be optimized specifically to “beat” other sources.
Instead, it highlights an important principle of AI retrieval: being relevant is not always enough.
When multiple sources are similarly relevant, retrieval systems may benefit from selecting information that adds something new.
For publishers and brands, this reinforces the value of creating genuinely useful, differentiated content rather than producing multiple pages that repeat the same information with minor wording changes.
Distinct expertise, original research, examples, data, and unique perspectives can make content more valuable within a broader information set.
MMR vs. Relevance Ranking
These approaches have different objectives.
| Approach | Primary Goal |
|---|---|
| Relevance Ranking | Find the most relevant results |
| MMR | Find relevant results while reducing redundancy |
| Re-Ranking | Improve the ordering of retrieved candidates |
MMR can itself be used as part of a broader re-ranking or retrieval process.
Related Terms
- Re-Ranking
- Relevance Scoring
- Retrieval Relevance
- Similarity Score
- Vector Similarity
- Dense Retrieval
- Hybrid Retrieval
- Passage Retrieval
- Cross-Encoder
- Bi-Encoder
In Simple Terms
Maximal Marginal Relevance (MMR) helps a retrieval system avoid returning the same idea repeatedly.
It tries to select results that are both relevant and meaningfully different, giving the system a more diverse and useful collection of information.
