Skip to content

Apache Lucene open-source search software

Allow implementers of AbstractKnnVectorQuery to access final topK results (#12590)

Kaival Parikh2474940bffe6
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 d81ab2806fe..f0de3ea7a39 100644
--- a/lucene/core/src/java/org/apache/lucene/search/AbstractKnnVectorQuery.java
+++ b/lucene/core/src/java/org/apache/lucene/search/AbstractKnnVectorQuery.java
@@ -88,7 +88,7 @@ abstract class AbstractKnnVectorQuery extends Query {
TopDocs[] perLeafResults = taskExecutor.invokeAll(tasks).toArray(TopDocs[]::new);
// Merge sort the results
- TopDocs topK = TopDocs.merge(k, perLeafResults);
+ TopDocs topK = mergeLeafResults(perLeafResults);
if (topK.scoreDocs.length == 0) {
return new MatchNoDocsQuery();
}
@@ -200,6 +200,22 @@ abstract class AbstractKnnVectorQuery extends Query {
return new TopDocs(totalHits, topScoreDocs);
}
+ /**
+ * Merges all segment-level kNN results to get the index-level kNN results.
+ *
+ * <p>The default implementation delegates to {@link TopDocs#merge(int, TopDocs[])} to find the
+ * overall top {@link #k}, which requires input results to be sorted.
+ *
+ * <p>This method is useful for reading and / or modifying the final results as needed.
+ *
+ * @param perLeafResults array of segment-level kNN results.
+ * @return index-level kNN results (no constraint on their ordering).
+ * @lucene.experimental
+ */
+ protected TopDocs mergeLeafResults(TopDocs[] perLeafResults) {
+ return TopDocs.merge(k, perLeafResults);
+ }
+
private Query createRewrittenQuery(IndexReader reader, TopDocs topK) {
int len = topK.scoreDocs.length;