Prevent extra similarity computation for single-level graphs (#12866)

### Description

[`#findBestEntryPoint`](4bc7850465/lucene/core/src/java/org/apache/lucene/util/hnsw/HnswGraphSearcher.java (L151)) is used to determine the entry point for the last level of HNSW search

It finds the single best-scoring node from [all upper levels](4bc7850465/lucene/core/src/java/org/apache/lucene/util/hnsw/HnswGraphSearcher.java (L159)) - but performs an [unnecessary computation](4bc7850465/lucene/core/src/java/org/apache/lucene/util/hnsw/HnswGraphSearcher.java (L157)) (along with [recording one visited node](4bc7850465/lucene/core/src/java/org/apache/lucene/util/hnsw/HnswGraphSearcher.java (L154))) when the graph just has 1 level (so the entry node is just the overall graph's entry node)

Also added a test to demonstrate this (fails without the changes in PR) -- where we visit `graph.size() + 1` nodes when the `topK` is high (should be a maximum of `graph.size()`)

---------

Co-authored-by: Kaival Parikh <kaivalp2000@gmail.com>

lucene/CHANGES.txt

@@ -140,6 +140,33 @@ Other
 
 * GITHUB#9049: Fixing bug in UnescapedCharSequence#toStringEscaped() (Jakub Slowinski)
 
+======================== Lucene 9.10.0 =======================
+
+API Changes
+---------------------
+(No changes)
+
+New Features
+---------------------
+(No changes)
+
+Improvements
+---------------------
+(No changes)
+
+Optimizations
+---------------------
+(No changes)
+
+Bug Fixes
+---------------------
+
+* GITHUB#12866: Prevent extra similarity computation for single-level HNSW graphs. (Kaival Parikh)
+
+Other
+---------------------
+(No changes)
+
 ======================== Lucene 9.9.0 =======================
 
 API Changes

lucene/core/src/java/org/apache/lucene/util/Version.java

@@ -112,10 +112,17 @@ public final class Version {
   /**
    * Match settings and bugs in Lucene's 9.9.0 release.
    *
-   * @deprecated Use latest
+   * @deprecated (9.10.0) Use latest
    */
   @Deprecated public static final Version LUCENE_9_9_0 = new Version(9, 9, 0);
 
+  /**
+   * Match settings and bugs in Lucene's 9.10.0 release.
+   *
+   * @deprecated Use latest
+   */
+  @Deprecated public static final Version LUCENE_9_10_0 = new Version(9, 10, 0);
+
   /**
    * Match settings and bugs in Lucene's 10.0.0 release.
    *

lucene/core/src/java/org/apache/lucene/util/hnsw/HnswGraphSearcher.java

@@ -100,19 +100,10 @@ public class HnswGraphSearcher {
       HnswGraphSearcher graphSearcher,
       Bits acceptOrds)
       throws IOException {
-    int initialEp = graph.entryNode();
+    int ep = graphSearcher.findBestEntryPoint(scorer, graph, knnCollector);
-    if (initialEp == -1) {
+    if (ep != -1) {
-      return;
+      graphSearcher.searchLevel(knnCollector, scorer, 0, new int[] {ep}, graph, acceptOrds);
     }
-    int[] epAndVisited = graphSearcher.findBestEntryPoint(scorer, graph, knnCollector.visitLimit());
-    int numVisited = epAndVisited[1];
-    int ep = epAndVisited[0];
-    if (ep == -1) {
-      knnCollector.incVisitedCount(numVisited);
-      return;
-    }
-    knnCollector.incVisitedCount(numVisited);
-    graphSearcher.searchLevel(knnCollector, scorer, 0, new int[] {ep}, graph, acceptOrds);
   }
 
   /**
@@ -143,18 +134,21 @@ public class HnswGraphSearcher {
    *
    * @param scorer the scorer to compare the query with the nodes
    * @param graph the HNSWGraph
-   * @param visitLimit How many vectors are allowed to be visited
+   * @param collector the knn result collector
-   * @return An integer array whose first element is the best entry point, and second is the number
+   * @return the best entry point, `-1` indicates graph entry node not set, or visitation limit
-   *     of candidates visited. Entry point of `-1` indicates visitation limit exceed
+   *     exceeded
    * @throws IOException When accessing the vector fails
    */
-  private int[] findBestEntryPoint(RandomVectorScorer scorer, HnswGraph graph, long visitLimit)
+  private int findBestEntryPoint(RandomVectorScorer scorer, HnswGraph graph, KnnCollector collector)
       throws IOException {
-    int size = getGraphSize(graph);
-    int visitedCount = 1;
-    prepareScratchState(size);
     int currentEp = graph.entryNode();
+    if (currentEp == -1 || graph.numLevels() == 1) {
+      return currentEp;
+    }
+    int size = getGraphSize(graph);
+    prepareScratchState(size);
     float currentScore = scorer.score(currentEp);
+    collector.incVisitedCount(1);
     boolean foundBetter;
     for (int level = graph.numLevels() - 1; level >= 1; level--) {
       foundBetter = true;
@@ -169,11 +163,11 @@ public class HnswGraphSearcher {
           if (visited.getAndSet(friendOrd)) {
             continue;
           }
-          if (visitedCount >= visitLimit) {
+          if (collector.earlyTerminated()) {
-            return new int[] {-1, visitedCount};
+            return -1;
           }
           float friendSimilarity = scorer.score(friendOrd);
-          visitedCount++;
+          collector.incVisitedCount(1);
           if (friendSimilarity > currentScore) {
             currentScore = friendSimilarity;
             currentEp = friendOrd;
@@ -182,7 +176,7 @@ public class HnswGraphSearcher {
         }
       }
     }
-    return new int[] {currentEp, visitedCount};
+    return collector.earlyTerminated() ? -1 : currentEp;
   }
 
   /**

lucene/core/src/test/org/apache/lucene/util/hnsw/HnswGraphTestCase.java

@@ -70,6 +70,7 @@ import org.apache.lucene.search.Sort;
 import org.apache.lucene.search.SortField;
 import org.apache.lucene.search.TaskExecutor;
 import org.apache.lucene.search.TopDocs;
+import org.apache.lucene.search.TopKnnCollector;
 import org.apache.lucene.store.Directory;
 import org.apache.lucene.tests.util.LuceneTestCase;
 import org.apache.lucene.tests.util.TestUtil;
@@ -1026,6 +1027,37 @@ abstract class HnswGraphTestCase<T> extends LuceneTestCase {
     }
   }
 
+  public void testAllNodesVisitedInSingleLevel() throws IOException {
+    int size = atLeast(100);
+    int dim = atLeast(50);
+
+    // Search for a large number of results
+    int topK = size - 1;
+
+    AbstractMockVectorValues<T> docVectors = vectorValues(size, dim);
+    HnswGraph graph =
+        HnswGraphBuilder.create(buildScorerSupplier(docVectors), 10, 30, random().nextLong())
+            .build(size);
+
+    HnswGraph singleLevelGraph =
+        new DelegateHnswGraph(graph) {
+          @Override
+          public int numLevels() {
+            // Only retain the last level
+            return 1;
+          }
+        };
+
+    AbstractMockVectorValues<T> queryVectors = vectorValues(1, dim);
+    RandomVectorScorer queryScorer = buildScorer(docVectors, queryVectors.vectorValue(0));
+
+    KnnCollector collector = new TopKnnCollector(topK, Integer.MAX_VALUE);
+    HnswGraphSearcher.search(queryScorer, collector, singleLevelGraph, null);
+
+    // Check that we visit all nodes
+    assertEquals(graph.size(), collector.visitedCount());
+  }
+
   private int computeOverlap(int[] a, int[] b) {
     Arrays.sort(a);
     Arrays.sort(b);
@@ -1297,4 +1329,42 @@ abstract class HnswGraphTestCase<T> extends LuceneTestCase {
 
     return sb.toString();
   }
+
+  private static class DelegateHnswGraph extends HnswGraph {
+    final HnswGraph delegate;
+
+    DelegateHnswGraph(HnswGraph delegate) {
+      this.delegate = delegate;
+    }
+
+    @Override
+    public void seek(int level, int target) throws IOException {
+      delegate.seek(level, target);
+    }
+
+    @Override
+    public int size() {
+      return delegate.size();
+    }
+
+    @Override
+    public int nextNeighbor() throws IOException {
+      return delegate.nextNeighbor();
+    }
+
+    @Override
+    public int numLevels() throws IOException {
+      return delegate.numLevels();
+    }
+
+    @Override
+    public int entryNode() throws IOException {
+      return delegate.entryNode();
+    }
+
+    @Override
+    public NodesIterator getNodesOnLevel(int level) throws IOException {
+      return delegate.getNodesOnLevel(level);
+    }
+  }
 }