Skip to content

Apache Lucene open-source search software

Corrects the VectorScorer.Bulk interface to be iterator friendly (#15496)

Benjamin Trentf7c5140dc4d3
diff --git a/lucene/CHANGES.txt b/lucene/CHANGES.txt
index 6ce3d131a26..6d8d157f009 100644
--- a/lucene/CHANGES.txt
+++ b/lucene/CHANGES.txt
@@ -176,7 +176,7 @@ New Features
* GITHUB#14565: Add ParentsChildrenBlockJoinQuery that supports parent and child filter in the same query
along with limiting number of child documents to retrieve per parent. (Jinny Wang)
-* GITHUB#15171: Add `VectorScorer.Bulk` for bulk iteration and scoring of vectors. This new interface is now
+* GITHUB#15171, GITHUB#15496: Add `VectorScorer.Bulk` for bulk iteration and scoring of vectors. This new interface is now
used in AbstractKnnVectorQuery to make exact matches faster. (Ben Trent)
* GITHUB#15176: Add `[Float|Byte]VectorValues#rescorer(element[])` interface to allow optimized rescoring of vectors.
diff --git a/lucene/core/src/java/org/apache/lucene/search/AbstractKnnVectorQuery.java b/lucene/core/src/java/org/apache/lucene/search/AbstractKnnVectorQuery.java
index c2a097da55d..8f618c92ddf 100644
--- a/lucene/core/src/java/org/apache/lucene/search/AbstractKnnVectorQuery.java
+++ b/lucene/core/src/java/org/apache/lucene/search/AbstractKnnVectorQuery.java
@@ -316,17 +316,16 @@ abstract class AbstractKnnVectorQuery extends Query {
HitQueue queue = new HitQueue(queueSize, true);
TotalHits.Relation relation = TotalHits.Relation.EQUAL_TO;
ScoreDoc topDoc = queue.top();
- DocIdSetIterator vectorIterator = vectorScorer.iterator();
DocAndFloatFeatureBuffer buffer = new DocAndFloatFeatureBuffer();
VectorScorer.Bulk bulkScorer = vectorScorer.bulk(acceptIterator);
- while (vectorIterator.docID() != DocIdSetIterator.NO_MORE_DOCS) {
+ for (float maxScore = bulkScorer.nextDocsAndScores(DocIdSetIterator.NO_MORE_DOCS, null, buffer);
+ buffer.size > 0;
+ maxScore = bulkScorer.nextDocsAndScores(DocIdSetIterator.NO_MORE_DOCS, null, buffer)) {
// Mark results as partial if timeout is met
if (queryTimeout != null && queryTimeout.shouldExit()) {
relation = TotalHits.Relation.GREATER_THAN_OR_EQUAL_TO;
break;
}
- // iterator already takes live docs into account
- float maxScore = bulkScorer.nextDocsAndScores(64, null, buffer);
if (maxScore < topDoc.score) {
// all the scores in this batch are too low, skip
continue;
diff --git a/lucene/core/src/java/org/apache/lucene/search/VectorScorer.java b/lucene/core/src/java/org/apache/lucene/search/VectorScorer.java
index b86e20af618..f6a2c1f6d6b 100644
--- a/lucene/core/src/java/org/apache/lucene/search/VectorScorer.java
+++ b/lucene/core/src/java/org/apache/lucene/search/VectorScorer.java
@@ -30,6 +30,7 @@ import org.apache.lucene.util.hnsw.RandomVectorScorer;
* @lucene.experimental
*/
public interface VectorScorer {
+ int DEFAULT_BULK_BATCH_SIZE = 64;
/**
* Compute the score for the current document ID.
@@ -52,7 +53,7 @@ public interface VectorScorer {
* Additionally, if the iterators are unpositioned (docID() == -1), this method should position
* them to the first document.
*
- * @param matchingDocs the documents to score
+ * @param matchingDocs Optional filter to iterate over the documents to score
* @return a {@link Bulk} scorer
* @throws IOException if an exception occurs during bulk scorer creation
* @lucene.experimental
@@ -65,12 +66,13 @@ public interface VectorScorer {
if (iterator.docID() == -1) {
iterator.nextDoc();
}
- return (nextCount, liveDocs, buffer) -> {
- buffer.growNoCopy(nextCount);
+ return (upTo, liveDocs, buffer) -> {
+ assert upTo > 0;
+ buffer.growNoCopy(DEFAULT_BULK_BATCH_SIZE);
int size = 0;
float maxScore = Float.NEGATIVE_INFINITY;
for (int doc = iterator.docID();
- doc != DocIdSetIterator.NO_MORE_DOCS && size < nextCount;
+ doc < upTo && size < DEFAULT_BULK_BATCH_SIZE;
doc = iterator.nextDoc()) {
if (liveDocs == null || liveDocs.get(doc)) {
buffer.docs[size] = doc;
@@ -91,16 +93,16 @@ public interface VectorScorer {
*/
interface Bulk {
/**
- * Score up to nextCount documents, store the results in the provided buffer. Behaves similarly
- * to {@link Scorer#nextDocsAndScores(int, Bits, DocAndFloatFeatureBuffer)}
+ * Score docs ids iterating to upTo documents, store the results in the provided buffer. Behaves
+ * similarly to {@link Scorer#nextDocsAndScores(int, Bits, DocAndFloatFeatureBuffer)}
*
- * @param nextCount the maximum number of documents to score
+ * @param upTo the maximum doc ID to score
* @param liveDocs the live docs, or null if all docs are live
* @param buffer the buffer to store the results
* @return the max score of the scored documents
* @throws IOException if an exception occurs during scoring
*/
- float nextDocsAndScores(int nextCount, Bits liveDocs, DocAndFloatFeatureBuffer buffer)
+ float nextDocsAndScores(int upTo, Bits liveDocs, DocAndFloatFeatureBuffer buffer)
throws IOException;
static Bulk fromRandomScorerDense(
@@ -111,14 +113,15 @@ public interface VectorScorer {
matchingDocs == null
? iterator
: ConjunctionUtils.createConjunction(List.of(matchingDocs, iterator), List.of());
- return (nextCount, liveDocs, buffer) -> {
+ return (upTo, liveDocs, buffer) -> {
+ assert upTo > 0;
if (matches.docID() == -1) {
matches.nextDoc();
}
- buffer.growNoCopy(nextCount);
+ buffer.growNoCopy(DEFAULT_BULK_BATCH_SIZE);
int size = 0;
for (int doc = matches.docID();
- doc != DocIdSetIterator.NO_MORE_DOCS && size < nextCount;
+ doc < upTo && size < DEFAULT_BULK_BATCH_SIZE;
doc = matches.nextDoc()) {
if (liveDocs == null || liveDocs.get(doc)) {
buffer.docs[size++] = doc;
@@ -141,16 +144,17 @@ public interface VectorScorer {
int[] docIds = new int[0];
@Override
- public float nextDocsAndScores(
- int nextCount, Bits liveDocs, DocAndFloatFeatureBuffer buffer) throws IOException {
+ public float nextDocsAndScores(int upTo, Bits liveDocs, DocAndFloatFeatureBuffer buffer)
+ throws IOException {
+ assert upTo > 0;
if (matches.docID() == -1) {
matches.nextDoc();
}
- buffer.growNoCopy(nextCount);
- docIds = ArrayUtil.growNoCopy(docIds, nextCount);
+ buffer.growNoCopy(DEFAULT_BULK_BATCH_SIZE);
+ docIds = ArrayUtil.growNoCopy(docIds, DEFAULT_BULK_BATCH_SIZE);
int size = 0;
for (int doc = matches.docID();
- doc != DocIdSetIterator.NO_MORE_DOCS && size < nextCount;
+ doc < upTo && size < DEFAULT_BULK_BATCH_SIZE;
doc = matches.nextDoc()) {
if (liveDocs == null || liveDocs.get(doc)) {
buffer.docs[size] = iterator.index();