iter

Author: iverase

server/src/main/java/org/elasticsearch/index/codec/vectors/IVFVectorsReader.java

@@ -236,7 +236,7 @@ public final void search(String field, float[] target, KnnCollector knnCollector
             // not perfect, but a comparative heuristic.
             // TODO: we might want to consider the density of the centroids as experiments shows that for fewer vectors per centroid,
             // the least vectors we need to score to get a good recall.
-            maxVectorVisited = (long) (2.0 * Math.round(Math.log10(numVectors) * Math.log10(numVectors)) * (knnCollector.k()));
+            maxVectorVisited = Math.round(1.75f * Math.log10(knnCollector.k()) * Math.log10(knnCollector.k()) * (knnCollector.k()));
             // clip so we visit at least one vector
             maxVectorVisited = Math.max(maxVectorVisited, 1L);
         } else {

server/src/main/java/org/elasticsearch/search/vectors/AbstractIVFKnnVectorQuery.java

@@ -54,7 +54,7 @@ abstract class AbstractIVFKnnVectorQuery extends Query implements QueryProfilerP
     protected final int k;
     protected final int numCands;
     protected final Query filter;
-    protected final KnnSearchStrategy searchStrategy;
+    protected final IVFKnnSearchStrategy searchStrategy;
     protected int vectorOpsCount;
 
     protected AbstractIVFKnnVectorQuery(String field, float visitRatio, int k, int numCands, Query filter) {

server/src/main/java/org/elasticsearch/search/vectors/IVFKnnFloatVectorQuery.java

@@ -15,6 +15,7 @@
 import org.apache.lucene.search.Query;
 import org.apache.lucene.search.TopDocs;
 import org.apache.lucene.search.knn.KnnCollectorManager;
+import org.apache.lucene.search.knn.KnnSearchStrategy;
 import org.apache.lucene.util.Bits;
 
 import java.io.IOException;
@@ -79,17 +80,24 @@ protected TopDocs approximateSearch(
         int visitedLimit,
         KnnCollectorManager knnCollectorManager
     ) throws IOException {
-        KnnCollector knnCollector = knnCollectorManager.newCollector(visitedLimit, searchStrategy, context);
-        if (knnCollector == null) {
-            return NO_RESULTS;
-        }
         LeafReader reader = context.reader();
         FloatVectorValues floatVectorValues = reader.getFloatVectorValues(field);
         if (floatVectorValues == null) {
             FloatVectorValues.checkField(reader, field);
             return NO_RESULTS;
         }
-        if (Math.min(knnCollector.k(), floatVectorValues.size()) == 0) {
+        if (floatVectorValues.size() == 0) {
+            return NO_RESULTS;
+        }
+        KnnSearchStrategy strategy = searchStrategy;
+        if (searchStrategy.getVisitRatio() == 0.0f) {
+            // dynamically set the percentage
+            float expected = (float) Math.round(1.75f * Math.log10(numCands) * Math.log10(numCands) * (numCands));
+            float ratio = expected / floatVectorValues.size();
+            strategy = new IVFKnnSearchStrategy(ratio);
+        }
+        KnnCollector knnCollector = knnCollectorManager.newCollector(visitedLimit, strategy, context);
+        if (knnCollector == null) {
             return NO_RESULTS;
         }
         reader.searchNearestVectors(field, query, knnCollector, acceptDocs);