Maximum Marginal Relevance: Balancing Relevance and Diversity in Grocery AI Retrieval (2026 Guide)
Imagine a shopper asks:
What do I need for taco night?
A semantic retrieval system searches a grocery knowledge base and returns:
Ground beef taco recipe
Ground beef tacos for families
Easy beef taco dinner
Weeknight ground beef tacos
Classic beef taco recipe
Every result is:
relevant.
But together, they are also:
highly redundant.
The shopper might have been better served by a set containing:
Taco protein options
Tortillas
Salsa
Cheese
Avocado or guacamole
Vegetarian filling
Grocery list
Store-specific availability
The retrieval problem is therefore not merely:
Which documents are most relevant to the query?
It can also be:
Which relevant documents add useful information that the other selected documents do not already provide?
That is the problem addressed by Maximal Marginal Relevance, commonly abbreviated MMR and also frequently called Maximum Marginal Relevance in modern retrieval tooling.
The original work by Jaime Carbonell and Jade Goldstein introduced MMR as a method for combining query relevance with information novelty, with the goal of reducing redundancy while maintaining relevance in retrieved documents and summaries.
Modern retrieval libraries continue to use the same underlying idea. LangChain, for example, describes maximal marginal relevance search as optimizing simultaneously for similarity to the query and diversity among selected documents.
For grocery AI retrieval, that tradeoff can be extremely useful.
A shopper asking:
healthy breakfast ideas from my grocery store
does not necessarily need:
five near-duplicate oatmeal passages.
A RAG system may perform better with context spanning:
Oatmeal
Greek yogurt
Fruit
Cottage cheese
Whole-grain toast
provided those results remain:
relevant enough to the request.
This leads to the central principle of this guide:
Similarity finds relevant candidates. MMR asks whether the next candidate contributes something new.
And the defining retrieval principle is:
The best context set is not always the collection of individually most similar passages.
Table of Contents
What Is Maximum Marginal Relevance?
Why Retrieval Diversity Matters
Relevance vs Redundancy
What Does “Marginal” Mean?
How MMR Works
The MMR Objective
Query Similarity
Candidate Redundancy
The Lambda Tradeoff
How MMR Selects Results Iteratively
Grocery Retrieval Example
Taco Night Example
Product Discovery Example
Recipe Retrieval Example
Department Search Example
Location Search Example
Multi-Location Retrieval
Dietary Search Example
Weekly Ad Retrieval
MMR and RAG
Why Redundant RAG Context Can Be Wasteful
MMR and Context Windows
MMR and Context Selection
MMR vs Top-k Similarity Search
MMR vs Reranking
MMR vs Cross-Encoder Reranking
MMR vs Hybrid Retrieval
MMR vs Dense Retrieval
MMR vs Sparse Retrieval
MMR vs Late Interaction and ColBERT
MMR and HNSW
MMR and Vector Databases
MMR and Multi-Stage Retrieval
MMR and Query Fan-Out
MMR and Entity Resolution
MMR and Content Chunking
MMR and Passage-Level Optimization
MMR and Information Gain
MMR and Source Freshness
Relevance-Diversity Failure Modes
Too Much Relevance
Too Much Diversity
Diversity Without Constraint Satisfaction
Diversity Without Freshness
Diversity Across the Wrong Entity
The Grocery MMR Selection Model
The Grocery Context Diversity Matrix
The DIVERSE Framework
What Grocery Marketers Can Control
What Grocery Marketers Cannot Control
Does Google Use MMR?
Does ChatGPT Search Use MMR?
Common MMR Myths
Grocery MMR Retrieval Checklist
Frequently Asked Questions
Final Thoughts
What Is Maximum Marginal Relevance?
Maximum Marginal Relevance is a selection strategy that tries to balance:
relevance to the query
against:
redundancy with items already selected.
Suppose retrieval produces:
20 candidate passages.
A normal top-k similarity approach might simply choose:
the five passages with the highest query similarity.
MMR instead chooses results:
iteratively.
At each step, it asks something conceptually like:
Which remaining candidate is highly relevant to the query while also being sufficiently different from what I have already selected?
That difference is essential.
MMR is not simply:
find diverse documents.
It is:
find relevant documents while penalizing excessive redundancy.
Why Retrieval Diversity Matters
Suppose an AI assistant retrieves five passages for:
best high-protein breakfast foods at a grocery store
Pure Similarity Results
Greek yogurt is high in protein.
Greek yogurt provides a protein-rich breakfast.
High-protein Greek yogurt is useful at breakfast.
Greek yogurt can increase breakfast protein.
Greek yogurt varieties offer protein.
Very relevant.
But informationally:
repetitive.
Now compare:
Diverse Relevant Results
Greek yogurt
Eggs
Cottage cheese
Protein oatmeal
Smoked salmon
The second set may provide:
more useful coverage of the information need.
The goal is not diversity for its own sake.
It is:
useful non-redundancy.
Relevance vs Redundancy
This article should make this distinction visually obvious.
Relevance
How well does this candidate address the query?
Redundancy
How much does this candidate repeat information already selected?
A document can be:
extremely relevant
and still add:
almost nothing new.
That means:
high relevance does not automatically imply high marginal value.
What Does “Marginal” Mean?
In MMR, “marginal” refers conceptually to:
the additional value of selecting one more item.
Suppose the selected set already contains:
a detailed passage about organic strawberries.
Now another passage is:
another detailed passage about organic strawberries.
Its incremental value may be:
low.
But a relevant passage about:
organic blueberries
could add:
a new dimension of useful information.
Therefore:
marginal relevance asks what a candidate adds beyond what is already selected.
How MMR Works
MMR operates:
iteratively.
Imagine a candidate pool:
ABCDEF
The system first selects:
a strong query-relevant result.
Then it evaluates remaining candidates based on:
Query relevance
Similarity to already selected results
Conceptually:
Relevant to Query − Redundant With Selected Set
The process repeats until:
the desired number of results is selected.
The MMR Objective
A common conceptual formulation is:
MMR = λ × Query Relevance − (1 − λ) × Redundancy
More formally, at each selection step the system chooses a remaining candidate that balances:
similarity to the query
against:
maximum similarity to an already selected result.
A commonly used form is:
arg max [ λ Sim(candidate, query) − (1 − λ) max Sim(candidate, selected) ]
The original MMR research framed the method around combining relevance and novelty, while modern implementations expose a tunable parameter controlling that tradeoff.
The exact scoring implementation can vary.
The important concept is:
reward query relevance, penalize redundancy.
Query Similarity
The first component asks:
How relevant is this candidate to the shopper's question?
Example query:
vegetarian taco dinner
Candidate:
Black bean tacos with avocado and salsa.
Strong relevance.
Candidate:
Apple pie recipe.
Very low relevance.
MMR still requires:
relevance.
Diversity does not rescue:
irrelevant content.
Candidate Redundancy
The second component asks:
How similar is this candidate to what has already been selected?
Suppose:
Selected
Black bean tacos with avocado.
Candidate A
Easy black bean taco recipe.
Candidate B
Roasted cauliflower tacos.
Candidate A might be:
more similar to both the query and selected passage.
Candidate B may provide:
more new information.
MMR can choose Candidate B if:
its relevance remains strong enough.
The Lambda Tradeoff
Many MMR implementations use:
λ
to control the balance between:
relevance and diversity.
LangChain's current implementation exposes lambda_mult from 0 to 1, where values closer to 1 emphasize query similarity and values closer to 0 emphasize diversity.
Conceptually:
λ → 1
Prioritize:
relevance.
Results may become:
more redundant.
λ → 0
Prioritize:
diversity.
Results may become:
less directly relevant.
Somewhere Between
Attempt to balance:
both.
But there is no universal:
best grocery MMR lambda.
That should be evaluated for:
Query type
Corpus
Retrieval stage
Context size
Use case
How MMR Selects Results Iteratively
Suppose the query is:
healthy lunch ideas
Candidate scores begin with:
Chicken salad — 0.92
Grilled chicken salad — 0.90
Caesar chicken salad — 0.89
Turkey wrap — 0.86
Lentil bowl — 0.84
Tuna sandwich — 0.82
Pure top-3 similarity could return:
three chicken salads.
MMR might instead select:
Chicken salad
Turkey wrap
Lentil bowl
because after the first chicken salad:
additional chicken-salad candidates become increasingly redundant.
This is the MMR intuition.
Grocery Retrieval Example
Query:
foods for a high-protein breakfast
Candidate pool:
Greek yogurt
Greek yogurt guide
Best Greek yogurt brands
Eggs
Cottage cheese
Protein oatmeal
Peanut butter
Smoked salmon
Pure semantic similarity could over-select:
Greek yogurt content.
MMR may produce:
a broader high-protein breakfast set.
This can be especially useful when the downstream AI system needs:
multiple complementary pieces of evidence.
Taco Night Example
Query:
What should I buy for taco night?
Candidate pool:
Ground beef taco recipe
Classic taco recipe
Taco seasoning guide
Tortilla guide
Salsa page
Guacamole page
Mexican cheese guide
Vegetarian taco filling
Store weekly ad
A pure relevance ranking may cluster around:
taco recipes.
MMR can encourage coverage across:
Protein
Tortillas
Seasoning
Toppings
Salsa
Cheese
The result becomes:
more useful as a shopping answer.
Product Discovery Example
Query:
snacks for a school lunch
A similarity-only retriever could return:
five granola-bar products.
MMR could return:
Granola bar
Applesauce
Pretzels
Cheese snack
Fruit cup
But diversity must not overpower:
dietary constraints.
If the query is:
nut-free snacks for school
every selected candidate must still satisfy:
the nut-free requirement.
This gives us a critical rule:
Diversity should operate inside the eligible candidate set—not replace eligibility.
Recipe Retrieval Example
Query:
easy dinners using chicken
Pure similarity:
Chicken pasta
Creamy chicken pasta
Garlic chicken pasta
One-pot chicken pasta
MMR:
Chicken pasta
Sheet-pan chicken
Chicken stir-fry
Chicken tacos
The latter provides:
broader meal possibilities.
Department Search Example
Query:
What services does your meat department offer?
Relevant candidate passages might cover:
Custom cutting
Grinding
Special orders
Marinated meats
Butcher consultation
If the top five results all discuss:
custom cutting,
the generated response may miss:
the broader service set.
MMR can help surface:
complementary evidence.
Location Search Example
Query:
What makes the Downtown grocery location useful?
Potential information:
Pharmacy
Seafood counter
Bakery
Pickup
Late hours
Parking
Prepared foods
A similarity retriever might over-select:
several location-description passages.
A diversity-aware selector might surface:
different useful attributes.
But all selected content must still belong to:
Downtown.
Multi-Location Retrieval
This is where diversity can become dangerous.
Suppose the query is:
What services are available at FreshTown Downtown?
Candidate set:
Downtown
Bakery
Curbside pickup
Northside
Pharmacy
Delivery
A diversity algorithm might prefer:
Pharmacy + Delivery
because they add novel information.
But they are:
the wrong location.
Therefore:
entity filtering must precede or constrain diversity selection.
A perfect diversity objective applied to:
the wrong candidate universe
still produces:
the wrong answer.
Dietary Search Example
Query:
gluten-free breakfast options
MMR might diversify results across:
Gluten-free cereal
Yogurt
Eggs
Gluten-free bread
Fruit
That can improve:
informational breadth.
But gluten-free status should come from:
explicit, reliable product information.
MMR does not verify:
dietary safety.
Weekly Ad Retrieval
Query:
best dinner deals this week
Suppose candidates include:
Chicken breast promotion
Another chicken breast promotion snippet
Chicken weekly-ad landing page
Salmon sale
Pasta sale
Vegetable sale
MMR could reduce:
duplicate chicken information
and increase:
promotional variety.
This can produce richer context for:
meal planning.
But current promotion dates must still be:
enforced.
MMR and RAG
MMR became especially relevant again with:
retrieval-augmented generation.
A RAG system may retrieve several passages and insert them into:
the LLM context.
If those passages are highly redundant:
valuable context capacity is wasted.
MMR can help select:
a diverse set of relevant passages
from a larger candidate pool.
The simplified pipeline:
Query → Retrieve Candidates → MMR Selection → Context → Generate
This is one reason modern vector-store frameworks expose MMR retrieval directly. LangChain's vector-store interface, for example, fetches a larger candidate pool and then selects a smaller set using maximal marginal relevance.
Why Redundant RAG Context Can Be Wasteful
Suppose an LLM has room for:
five grocery passages.
If all five say:
Downtown has a bakery,
then the context may fail to include:
Bakery hours
Custom cakes
Pickup
Deli
Pharmacy
Redundant passages consume:
context budget
without contributing equivalent new information.
This gives us one of the strongest lines in the article:
Retrieval quality is not just about finding strong passages. It is about assembling a strong set of passages.
MMR and Context Windows
A context window has:
finite capacity.
Even very large context windows should not automatically be filled with:
duplicate evidence.
The objective is not:
maximum content.
It is:
maximum useful coverage.
MMR can therefore serve as:
a context-efficiency mechanism.
But it is not itself:
context-window management.
MMR and Context Selection
The previous multi-stage retrieval article distinguished:
retrieval
from:
context selection.
MMR can operate:
between them.
Conceptually:
Candidate Retrieval → MMR Diversification → Evidence Evaluation → Context Selection
Or MMR itself may serve as part of:
context selection.
The architectural location depends on:
implementation.
Its conceptual job remains:
reduce redundant selections while preserving relevance.
MMR vs Top-k Similarity Search
This is the simplest comparison.
Top-k Similarity
Choose:
the k candidates most similar to the query.
MMR
Choose:
candidates that balance similarity to the query with difference from already selected candidates.
So:
Top-k asks: Which are closest?
MMR asks: Which relevant candidate adds the most new value next?
MMR vs Reranking
MMR can be considered:
a form of reranking or selection.
But not every reranker is:
MMR.
A conventional relevance reranker might reorder candidates based on:
query-document relevance.
MMR additionally considers:
relationships among the selected documents.
This is key.
Traditional reranking evaluates candidates against the query. MMR also evaluates candidates against one another.
MMR vs Cross-Encoder Reranking
Cross-encoder reranking asks:
How relevant is this candidate to the query?
MMR asks:
How relevant is this candidate to the query, and how much does it duplicate what I already selected?
They can therefore work together:
Retrieve → Cross-Encoder Rerank → MMR Diversify
or potentially:
Retrieve → MMR → Rerank
depending on architecture.
No universal ordering applies.
The critical distinction:
Cross-encoders improve relevance precision. MMR improves set-level relevance diversity.
MMR vs Hybrid Retrieval
Hybrid retrieval combines:
multiple candidate-generation methods.
For example:
sparse + dense.
MMR selects:
a diverse relevant subset.
So:
Hybrid Retrieval = Where candidates come from.
MMR = How a useful non-redundant set can be selected from candidates.
MMR vs Dense Retrieval
Dense retrieval can return:
semantically similar candidates.
Because semantically similar candidates can themselves be:
highly similar to one another,
dense retrieval is a natural place where:
redundancy can emerge.
MMR can operate after:
dense candidate retrieval.
MMR vs Sparse Retrieval
Sparse search can also return:
repetitive results.
For example, exact phrase matching may retrieve:
multiple pages containing nearly identical wording.
MMR is not limited to:
dense retrieval.
The diversity logic can conceptually operate over:
any candidate pool where relevance and candidate-candidate similarity can be evaluated.
MMR vs Late Interaction and ColBERT
The previous article addressed:
fine-grained relevance matching.
ColBERT asks:
how well do detailed query representations interact with detailed document representations?
MMR asks:
after relevant candidates exist, which ones form the strongest non-redundant set?
So:
ColBERT improves candidate relevance resolution.
MMR improves candidate-set composition.
These are fundamentally different layers.
MMR and HNSW
HNSW can generate:
approximate nearest-neighbor candidates.
Those candidates may then be:
passed to MMR.
Conceptually:
Query → Embedding → HNSW → 50 Candidates → MMR → 8 Diverse Candidates
The specific numbers are illustrative.
LangChain's MMR API explicitly separates:
fetch_k
from:
k,
illustrating the common pattern of fetching a larger candidate pool before selecting a smaller diversified result set.
MMR and Vector Databases
Some vector-store interfaces support MMR as:
a retrieval option.
The system can:
Run similarity search
Fetch a candidate pool
Apply MMR
Return a smaller set
This demonstrates that MMR often sits:
downstream of base vector similarity search.
But:
MMR is not itself a vector database.
MMR and Multi-Stage Retrieval
Now place MMR in the broader architecture.
Question↓Candidate Generation↓Hybrid Fusion↓Filtering↓Entity Resolution↓Relevance Reranking↓MMR Diversity Selection↓Evidence Evaluation↓Context Selection↓Grounding↓Generation
This is one possible pipeline.
MMR's unique responsibility is:
set-level redundancy control.
MMR and Query Fan-Out
Query fan-out can create:
multiple related retrieval branches.
Example:
best grocery store for taco night
Potential subqueries:
Taco meat
Tortillas
Salsa
Produce
Cheese
Store hours
When results from those branches are combined:
redundancy can appear.
MMR can potentially help select:
complementary evidence across the combined pool.
But query fan-out itself already introduces:
information diversity.
So these mechanisms should not be confused.
MMR and Entity Resolution
Diversity is useful only after:
entity relationships are correct.
For query:
What can I buy at Downtown?
Do not diversify by pulling:
Northside-only services.
MMR must operate within:
the correct entity boundary.
This produces another signature line:
Diversify the answer—not the entity.
MMR and Content Chunking
Poor chunking can produce:
many nearly identical chunks.
For example, five overlapping chunks may all repeat:
the same department description.
Retrieval then returns:
redundant passages.
MMR can reduce:
the downstream impact.
But it does not fix:
bad source chunking.
The better approach is:
good semantic chunking + diversity-aware selection.
MMR and Passage-Level Optimization
Strong passages should be:
Focused
Specific
Independently understandable
Entity-aligned
If every passage says:
the same generic thing,
MMR has little useful diversity to work with.
Therefore:
retrieval diversity begins with information diversity.
MMR and Information Gain
MMR has a natural relationship to:
information gain.
Information Gain asks:
does this content contribute something genuinely useful?
MMR asks:
does this candidate add useful novelty relative to what is already selected?
They are not the same.
But both resist:
repetitive information.
A useful distinction is:
Information Gain = Does the source add something new to the information environment?
MMR = Does this candidate add something new to this selected result set?
MMR and Source Freshness
Suppose:
Candidate A
Current chicken promotion.
Candidate B
Current salmon promotion.
Candidate C
Expired vegetable promotion.
Candidate C may offer:
excellent diversity.
But it is:
stale.
Therefore:
freshness should constrain or precede diversity.
Do not allow novelty to reward:
incorrect information.
Relevance-Diversity Failure Modes
MMR introduces:
a tradeoff.
Tradeoffs create:
failure modes.
The two broad extremes are:
relevance collapse
and:
diversity collapse.
Too Much Relevance
If diversity is given:
almost no weight,
MMR becomes increasingly similar to:
top-k relevance retrieval.
The set may contain:
Duplicate passages
Near-identical products
Repeated facts
Overlapping chunks
Result:
high relevance, low information coverage.
Too Much Diversity
If diversity receives:
too much weight,
the system can select candidates that are:
novel but weakly related.
Query:
best vegetables for stir-fry
A poorly balanced system might select:
Broccoli
Bell peppers
Soy sauce
Cast-iron skillet
Chicken
The set becomes:
varied,
but increasingly:
off-task.
This leads to:
Diversity without relevance is noise.
Diversity Without Constraint Satisfaction
Query:
vegan dinner ideas
Do not diversify into:
Chicken curry
Salmon bowl
Beef tacos
because:
they are different.
Eligibility must remain:
satisfied first.
Diversity Without Freshness
Query:
deals this week
Do not select an expired promotion simply because:
it adds variety.
Diversity Across the Wrong Entity
Query:
Downtown store services
Do not diversify across:
Northside and Westside services.
The rule:
Diversity must remain inside the truth boundary of the query.
The Grocery MMR Selection Model
This should be the article's primary signature visual.
Start with:
CANDIDATE POOL
Result A
Result B
Result C
Result D
Result E
Result F
↓
QUERY RELEVANCE
Which candidates answer:
the shopper's request?
↓
REDUNDANCY CHECK
Which candidates repeat:
already selected information?
↓
MARGINAL VALUE
Which candidate adds:
the strongest useful new information?
↓
SELECT
↓
REPEAT
This becomes:
Relevant → Compare → Diversify → Select → Repeat
The core principle:
MMR evaluates the value of the next result in the context of the results already chosen.
The Grocery Context Diversity Matrix
This should be the most intuitive visual.
Query:
What should I buy for taco night?
Candidate | Query Relevance | Redundancy | Marginal Value |
Beef taco recipe | High | Low initially | High |
Easy beef taco recipe | High | High | Low |
Tortilla guide | High | Low | High |
Salsa options | High | Low | High |
Guacamole guide | High | Low | High |
Chocolate cake guide | Low | Low | Low |
This demonstrates:
novel does not automatically mean useful.
The ideal result has:
high relevance + low unnecessary redundancy.
The DIVERSE Framework
This should be the Grocery Growth marketer-facing framework.
It is not the mathematical MMR algorithm.
D — Define the Information Need
Understand whether the shopper needs:
One exact answer
Several options
Multiple categories
Supporting evidence
Comparative information
Not every query needs:
maximum diversity.
I — Identify Hard Constraints
Before diversification, enforce:
Location
Dietary requirements
Price
Time
Inventory
Store eligibility
V — Verify Entity Ownership
Make sure selected facts belong to:
the correct entity.
E — Establish Relevance
Candidates must still:
answer the query.
R — Reduce Redundancy
Avoid selecting:
multiple passages that contribute essentially the same information.
S — Select Complementary Evidence
Favor relevant candidates that:
expand useful coverage.
E — Evaluate the Final Set
Ask:
Does the collection collectively answer the question better than the individually highest-similarity results would?
So:
DIVERSE = Define → Identify → Verify → Establish → Reduce → Select → Evaluate
This is a Grocery Growth retrieval-content framework.
It is not Google's, OpenAI's, or MMR's internal algorithm.
What Grocery Marketers Can Control
Most grocery marketers will not:
tune an MMR lambda.
But they can improve the diversity and usefulness of the information available to retrieval systems.
Create distinct information about:
Products
Departments
Services
Locations
Recipes
Dietary options
Promotions
Shopping occasions
Customer questions
Local specialties
Avoid having:
20 pages saying essentially the same thing.
The source-level principle is:
A retrieval system cannot assemble diverse useful evidence if the website only publishes redundant information.
What Grocery Marketers Cannot Control
For external search and AI platforms, grocery marketers generally cannot control:
Whether MMR is used
Candidate pool size
Lambda
Diversity penalty
Similarity metric
Selection order
MMR implementation
Context-selection architecture
Diversity threshold
Result-set optimization
Do not try to optimize:
an imaginary external MMR score.
Does Google Use MMR?
Do not claim:
Google Search or AI Overviews universally use Maximum Marginal Relevance.
Google's public documentation establishes that modern Search and generative features can retrieve and synthesize information from multiple sources.
It does not publicly establish:
a universal MMR implementation, lambda, redundancy formula, or document-diversity score.
MMR should therefore be explained as:
a well-established retrieval and selection technique, not a secret Google ranking mechanism.
Does ChatGPT Search Use MMR?
OpenAI does not publicly specify:
Maximum Marginal Relevance
as a universal ChatGPT Search algorithm.
OpenAI documents web searching, query rewriting, relevance, reliability, and source citations, but not a publisher-facing:
MMR lambda or diversity-selection formula.
So do not claim:
ChatGPT universally uses MMR
ChatGPT has an MMR score
OAI-SearchBot calculates diversity penalties
MMR determines ChatGPT citations
Those would be unsupported.
Common MMR Myths
“MMR means Maximum Marginal Relevance.”
You will see both Maximum Marginal Relevance and Maximal Marginal Relevance used. The original Carbonell and Goldstein work uses Maximal Marginal Relevance terminology. Modern tools also commonly use that phrasing.
For this article's title, “Maximum Marginal Relevance” is understandable search language, but the body should mention:
Maximal Marginal Relevance (MMR)
as the original/formal terminology.
“MMR just finds diverse content.”
No.
It balances:
relevance + diversity.
“MMR intentionally selects irrelevant results.”
It should not.
“MMR is a similarity metric.”
No.
It is:
a selection/reranking strategy.
“MMR replaces vector search.”
No.
It often operates:
on candidates already retrieved.
“MMR replaces HNSW.”
No.
“MMR replaces hybrid retrieval.”
No.
“MMR replaces cross-encoder reranking.”
No.
“MMR is a cross-encoder.”
No.
“MMR is ColBERT.”
No.
“MMR is RAG.”
No.
“MMR only works with embeddings.”
The concept predates modern embedding retrieval and was originally developed for information retrieval and summarization.
“MMR always improves results.”
No.
It can hurt relevance if:
diversity is overemphasized.
“More diversity is always better.”
No.
“Lambda 0.5 is the best setting.”
No.
Some software uses 0.5 as a default, but that is not a universal optimum. LangChain currently defaults lambda_mult to 0.5 in its vector-store interface.
“Lambda 1 means maximum diversity.”
Not in LangChain's current lambda_mult convention. There, 1 emphasizes query similarity and 0 maximizes diversity.
Different APIs can expose diversity controls differently, so:
always check implementation semantics.
“MMR verifies whether a product is in stock.”
No.
“MMR checks allergens.”
No.
“MMR determines the correct location.”
No.
“MMR solves stale content.”
No.
“MMR guarantees better grounding.”
No.
It can improve:
context composition,
but grounding remains downstream.
“MMR guarantees citation diversity.”
No.
“Google has an MMR ranking factor.”
No public ranking factor of this kind is documented.
“ChatGPT has a publisher-visible MMR setting.”
No.
“There is MMR schema markup.”
No.
Grocery MMR Retrieval Checklist
Candidate Quality
Candidates genuinely relevant
Wrong-topic candidates removed
Hard constraints enforced
Entity ownership verified
Stale content excluded
Redundancy
Duplicate pages minimized
Overlapping chunks reviewed
Near-identical passages identified
Syndicated/duplicate content controlled
Repeated facts consolidated
Diversity
Different relevant subtopics represented
Different useful product options surfaced
Different supporting evidence considered
Diversity remains query-relevant
Diversity remains entity-correct
If Building Your Own Retrieval System
Similarity-search baseline measured
MMR tested against top-k retrieval
Candidate pool size tested
Final k tested
Lambda tested
Human relevance evaluated
Redundancy measured
Context coverage measured
Latency measured
Downstream RAG answers evaluated
High-Consequence Constraints
Dietary claims explicit
Allergy claims verified
Price current
Inventory current where required
Location correct
Promotion dates current
Claims Discipline
No claim MMR = pure diversity
No claim MMR replaces retrieval
No universal lambda recommendation
No claim MMR verifies facts
No claim Google universally uses MMR
No claim ChatGPT universally uses MMR
No invented MMR ranking factor
Frequently Asked Questions
What is Maximum Marginal Relevance?
Maximum Marginal Relevance, more formally introduced as Maximal Marginal Relevance, is a method for balancing:
query relevance and information novelty.
The original Carbonell and Goldstein work used MMR to reduce redundancy while maintaining relevance in document reranking and summarization.
What does MMR stand for?
Maximal Marginal Relevance.
“Maximum Marginal Relevance” is also widely used colloquially and in search discussions.
What problem does MMR solve?
It helps prevent a selected result set from containing:
too many highly similar or redundant items.
How does MMR work?
Conceptually, it rewards:
similarity to the query
and penalizes:
similarity to already selected candidates.
What is the MMR formula?
A common conceptual form is:
λ × Query Relevance − (1 − λ) × Redundancy
with candidates selected:
iteratively.
What does lambda do?
Lambda controls:
the relevance-diversity tradeoff.
In LangChain's current implementation:
1 emphasizes similarity
0 emphasizes maximum diversity
with values between them balancing the two.
Is 0.5 the best lambda?
No.
It is a common default in some implementations, not:
a universal optimum.
What is fetch_k?
In LangChain's MMR implementation, fetch_k is the number of candidate documents initially fetched before MMR selects the final k documents.
Conceptually:
fetch broadly → diversify/select narrowly.
Why retrieve more candidates than you return?
Because MMR needs:
alternatives to choose among.
If you retrieve only:
exactly the number you plan to return,
the opportunity for diversification is:
limited.
Does MMR use cosine similarity?
It can use similarity values based on:
vector representations,
but the exact similarity function depends on:
implementation.
MMR itself is not:
cosine similarity.
Is MMR a reranker?
It can be used as:
a reranking/selection technique.
Is MMR a cross-encoder?
No.
Is MMR vector search?
No.
It can operate:
downstream of vector search.
Is MMR an embedding model?
No.
Is MMR the same as hybrid retrieval?
No.
Is MMR the same as ColBERT?
No.
Is MMR useful for RAG?
Yes.
It can help produce:
less redundant retrieved context.
Why does diversity matter in RAG?
Because a context window filled with:
duplicate information
can waste space that could contain:
complementary evidence.
Does MMR increase factual accuracy?
Not inherently.
It improves:
result-set composition.
Truth and accuracy require:
reliable source information and downstream validation.
Can MMR hurt retrieval?
Yes.
Excessive diversity weighting can:
reduce query relevance.
Should filters happen before MMR?
Often hard constraints such as:
Location
Dietary status
Time
Eligibility
should define the valid candidate universe before or during diversification.
The exact architecture varies.
Can MMR select multiple product categories?
Yes, if:
they remain relevant to the query.
Can MMR help conversational grocery queries?
Potentially.
Broad queries often benefit from:
multiple complementary pieces of information.
Does MMR improve exact product lookup?
Possibly less useful.
If a shopper asks:
UPC 123456
the goal is likely:
one exact match,
not result diversity.
What kinds of grocery queries benefit most from MMR?
Especially:
Recommendations
Meal planning
Recipe discovery
Product alternatives
Department overviews
Shopping-list generation
Broad research questions
RAG context assembly
What queries may need less diversity?
Especially:
Exact product lookup
Exact store hours
Exact price
Exact location
Specific inventory check
For these queries:
precision may dominate diversity.
Does Google use MMR?
Google does not publicly document MMR as a universal public Search or AI Overview retrieval mechanism.
Does ChatGPT Search use MMR?
OpenAI does not publicly document MMR as a universal ChatGPT Search retrieval mechanism.
Should grocery marketers optimize pages for MMR?
Not directly.
Instead:
publish distinct, useful, accurate information rather than many repetitive pages.
Final Thoughts
Retrieval systems often begin with a simple goal:
find the content most relevant to the query.
But when multiple pieces of information are selected, another problem appears:
What if all of the highly relevant results say essentially the same thing?
That is where Maximum—or more formally, Maximal—Marginal Relevance becomes valuable.
MMR introduces a second consideration:
novelty relative to what has already been selected.
The original MMR research explicitly framed the method around combining query relevance with information novelty to reduce redundancy.
Modern retrieval tooling continues to use the same core principle, including vector-store implementations that fetch a larger candidate set and return a smaller collection chosen for both query similarity and diversity.
For grocery AI retrieval, the distinction matters because many shopper questions are:
inherently multi-faceted.
A shopper asking:
What should I buy for taco night?
may benefit from:
tortillas + protein + salsa + cheese + produce
more than:
five nearly identical taco recipes.
A shopper asking:
healthy breakfast options
may benefit from:
yogurt + eggs + fruit + oatmeal + cottage cheese
more than:
five Greek-yogurt passages.
This makes MMR a:
set-optimization technique.
It does not simply ask:
“Which result is best?”
It asks:
“Which result should I add next, given what I have already selected?”
The Grocery MMR Selection Model becomes:
Candidate Pool → Relevance → Redundancy Check → Marginal Value → Select → Repeat
The DIVERSE Framework becomes:
Define → Identify → Verify → Establish → Reduce → Select → Evaluate
Within the wider Grocery Growth retrieval architecture:
Content → Embed → Candidate Retrieval → Filter → Entity Resolve → Relevance Rerank → MMR Diversify → Evidence Evaluate → Context Select → Ground → Generate → Cite
MMR owns:
the diversity and redundancy-control layer.
It does not replace:
Retrieval
Filtering
Entity resolution
Freshness
Reranking
Grounding
Citation
And it should never turn:
diversity
into:
randomness.
The three defining lines for this article should be:
Similarity finds relevant candidates. MMR asks whether the next candidate contributes something new.
Retrieval quality is not just about finding strong passages. It is about assembling a strong set of passages.
Diversity without relevance is noise. Relevance without diversity can be repetition.
Want Your Grocery Content to Provide Better Retrieval Coverage?
Grocery Growth helps grocery retailers build clearer, more useful information architectures for traditional search, semantic retrieval, AI-assisted discovery, and modern answer systems.
Optimization can include:
Product information
Department pages
Location architecture
Service information
Grocery content strategy
Semantic SEO
Entity optimization
Content chunking
Passage optimization
Information gain
Content freshness
Internal linking
Structured information
Crawlability
LLM retrieval optimization
GEO
AEO
SEO
AI-search optimization
The goal is not to manipulate:
an MMR lambda or an undisclosed diversity score.
It is to create enough:
distinct, accurate, relevant information
that retrieval systems have useful evidence to choose from.
Explore Grocery Growth's SEO and AI-search optimization services.
Written by Grocery Growth
Grocery Growth helps grocery retailers build clearer digital information architectures for traditional search, semantic retrieval, AI-assisted discovery, and modern grocery shopping experiences.
Related Articles
Representation & Matching: Vector Embeddings Explained for Grocery Store Marketers; Cosine Similarity and Semantic Retrieval; Late Interaction Retrieval: ColBERT and Token-Level Matching.
Candidate Generation: Dense vs Sparse Retrieval; Hybrid Retrieval; HNSW Explained.
Candidate Evaluation: Cross-Encoder Reranking; Bi-Encoders vs Cross-Encoders.
Pipeline Architecture: Multi-Stage Grocery Content Retrieval Pipelines.
Content Selection: Content Chunking; Passage-Level Optimization; Information Gain and AI Search.
Downstream AI: RAG and Grocery Store Websites; Grounding in AI Search; Citation Optimization.
Comments