Skip to content

ES|QL: Optimize MMR by reducing cache size and lookup #145014

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
ioanatia merged 5 commits into from
Mar 31, 2026
Merged

ES|QL: Optimize MMR by reducing cache size and lookup #145014

Changes from 4 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
58a1be0
Optimize MMR by reducing cache size and lookup
ioanatia Mar 26, 2026
a04be0b
Use min instead of max
ioanatia Mar 26, 2026
e75f379
Merge branch 'main' into mmr_optimization
ioanatia Mar 27, 2026
7e34522
Refactor
ioanatia Mar 30, 2026
3bed78a
Merge branch 'main' into mmr_optimization
ioanatia Mar 31, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
84 changes: 21 additions & 63 deletions .../src/main/java/org/elasticsearch/search/diversification/mmr/MMRResultDiversification.java
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,11 +16,11 @@

import java.io.IOException;
import java.util.ArrayList;
import java.util.Comparator;
import java.util.HashMap;
import java.util.Arrays;
import java.util.LinkedHashMap;
import java.util.List;
import java.util.Map;
import java.util.stream.IntStream;

public class MMRResultDiversification extends ResultDiversification<MMRResultDiversificationContext> {

Expand All @@ -46,14 +46,17 @@ public RankDoc[] diversify(RankDoc[] docs) throws IOException {
List<Integer> selectedDocRanks = new ArrayList<>();

// cache the similarity scores for the query vector vs. searchHits
Map<Integer, Float> querySimilarity = getQuerySimilarityForDocs(docs);

float[] querySimilarities = getQuerySimilarityForDocs(docs);
// always add the highest relevant doc to the list
selectedDocRanks.add(getHighestRelevantDocRank(docs, querySimilarity));
int prevSelectedDocRank = 1 + IntStream.range(0, querySimilarities.length)
.reduce(0, (a, b) -> querySimilarities[a] >= querySimilarities[b] ? a : b);

Map<Integer, Map<Integer, Float>> cachedSimilarities = new HashMap<>();
selectedDocRanks.add(prevSelectedDocRank);
int topDocsSize = context.getSize();

float[] maxSimilarityToSelected = new float[docs.length];
Arrays.fill(maxSimilarityToSelected, Float.NEGATIVE_INFINITY);

for (int x = 0; x < topDocsSize && selectedDocRanks.size() < topDocsSize && selectedDocRanks.size() < docs.length; x++) {
int thisMaxMMRDocRank = -1;
float thisMaxMMRScore = Float.NEGATIVE_INFINITY;
Expand All @@ -69,29 +72,26 @@ public RankDoc[] diversify(RankDoc[] docs) throws IOException {
continue;
}

var cachedScoresForDoc = cachedSimilarities.getOrDefault(docRank, new LinkedHashMap<>());

// compute MMR scores for remaining searchHits
float highestSimilarityScoreToSelected = getHighestSimilarityScoreToSelectedVectors(
selectedDocRanks,
float similarityToLastSelected = getVectorComparisonScore(

@markjhoy markjhoy Mar 30, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think I get at what the optimization is here -- you're only comparing the current document to the last selected document, correct? (the original implementation, and the implementation in the paper, computes MMR in respect to all the previously selected documents)...

I think this will work, but it's still unsure in my head if it will produce the most correct results...

@ioanatia ioanatia Mar 31, 2026

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We still compute MMR wrt all previously selected documents.

We keep an array of the computed max similarity between each doc and the selected set.
Then as we select new diversified docs and we iterate through the remaining docs to find a new candidate:

  • We compute the similarity between the current doc for which we calculate MMR and the doc that was added to the selected docs in the prev iteration.
  • We update the maxSimilarityToSelected[docRank - 1] for the current doc.
  • We compute the MMR score using the maxSimilarityToSelected value for the current doc.

similarityFunction,
thisDocVector,
cachedScoresForDoc
context.getFieldVector(prevSelectedDocRank)
);
maxSimilarityToSelected[docRank - 1] = Float.max(similarityToLastSelected, maxSimilarityToSelected[docRank - 1]);

// compute MMR
float querySimilarityScore = querySimilarity.getOrDefault(doc.rank, 0.0f);
float mmr = (context.getLambda() * querySimilarityScore) - ((1 - context.getLambda()) * highestSimilarityScoreToSelected);
float mmr = context.getLambda() * querySimilarities[docRank - 1] - (1 - context.getLambda())
* maxSimilarityToSelected[docRank - 1];

if (mmr > thisMaxMMRScore) {
thisMaxMMRScore = mmr;
thisMaxMMRDocRank = docRank;
}

// cache these scores
cachedSimilarities.put(docRank, cachedScoresForDoc);
}

if (thisMaxMMRDocRank >= 0) {
selectedDocRanks.add(thisMaxMMRDocRank);
prevSelectedDocRank = thisMaxMMRDocRank;
}
}

Expand All @@ -111,50 +111,9 @@ public RankDoc[] diversify(RankDoc[] docs) throws IOException {
return ret;
}

private Integer getHighestRelevantDocRank(RankDoc[] docs, Map<Integer, Float> querySimilarity) {
Map.Entry<Integer, Float> highestRelevantDoc = querySimilarity.entrySet()
.stream()
.max(Comparator.comparingDouble(Map.Entry::getValue))
.orElse(null);

if (highestRelevantDoc != null) {
return highestRelevantDoc.getKey();
}

// no query vectors? Just use the first document in the order
return docs[0].rank;
}

private float getHighestSimilarityScoreToSelectedVectors(
List<Integer> selectedDocRanks,
VectorData thisDocVector,
Map<Integer, Float> cachedScoresForDoc
) {
float highestScore = Float.NEGATIVE_INFINITY;
for (Integer compareToDocRank : selectedDocRanks) {
Float similarityScore = cachedScoresForDoc.getOrDefault(compareToDocRank, null);
if (similarityScore == null) {
VectorData comparisonVector = context.getFieldVector(compareToDocRank);
if (comparisonVector != null) {
if (comparisonVector.size() == 0) {
cachedScoresForDoc.put(compareToDocRank, Float.NEGATIVE_INFINITY);
continue;
}

similarityScore = getVectorComparisonScore(similarityFunction, thisDocVector, comparisonVector);
cachedScoresForDoc.put(compareToDocRank, similarityScore);
}
}
if (similarityScore != null && similarityScore > highestScore) {
highestScore = similarityScore;
}
}
return highestScore == Float.NEGATIVE_INFINITY ? 0.0f : highestScore;

@mromaios mromaios Mar 31, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💭 minor edge case, just from comparing implementations (not sure if it's a valid one). In the previous code if there was no valid similarity score, we would have 0.0f, but now we will get Float.NEGATIVE_INFINITY, that is if context.getFieldVector() returns null for every selected document.

@ioanatia ioanatia Mar 31, 2026

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We don't actually get to this path, because as we iterate through candidates, we skip those that don't have a vector value:

elasticsearch/server/src/main/java/org/elasticsearch/search/diversification/mmr/MMRResultDiversification.java

Lines 67 to 70 in dc43d5c

var thisDocVector = context.getFieldVector(docRank);
if (thisDocVector == null || thisDocVector.size() == 0) {
continue;
}

@mromaios mromaios Mar 31, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Oh right 🤦
Thanks!

}

protected Map<Integer, Float> getQuerySimilarityForDocs(RankDoc[] docs) {
Map<Integer, Float> querySimilarity = new HashMap<>();

protected float[] getQuerySimilarityForDocs(RankDoc[] docs) {
float[] querySimilarity = new float[docs.length];
Arrays.fill(querySimilarity, 0.0f);
VectorData queryVector = context.getQueryVector();
if (queryVector == null) {
return querySimilarity;
Expand All @@ -163,8 +122,7 @@ protected Map<Integer, Float> getQuerySimilarityForDocs(RankDoc[] docs) {
for (RankDoc doc : docs) {
VectorData vectorData = context.getFieldVector(doc.rank);
if (vectorData != null) {
float querySimilarityScore = getVectorComparisonScore(similarityFunction, vectorData, queryVector);
querySimilarity.put(doc.rank, querySimilarityScore);
querySimilarity[doc.rank - 1] = getVectorComparisonScore(similarityFunction, vectorData, queryVector);
}
}
return querySimilarity;
Expand Down
Loading