KNN benchy: include numDocs in indexPath so benchy reindexes when requested numDocs changes; remove obsolete workaround for bug that returns NO_MORE_DOCS as a Lucene docid; fix typo in generated nightly benchy HTML; add TODOs

Author: mikemccand

src/main/knn/KnnGraphTester.java

@@ -466,7 +466,7 @@ private void run(String... args) throws Exception {
       throw new IllegalArgumentException("-prefilter requires filterSelectivity between 0 and 1");
     }
     if (indexPath == null) {
-      indexPath = Paths.get(formatIndexPath(docVectorsPath)); // derive index path
+      indexPath = Paths.get(formatIndexPath(docVectorsPath, numDocs)); // derive index path
       log("Index Path = %s\n", indexPath);
     }
     if (parentJoin && reindex == false && isParentJoinIndex(indexPath) == false) {
@@ -665,7 +665,8 @@ private static Query generateRandomQuery(Random random, Path indexPath, int size
     }
   }
 
-  private String formatIndexPath(Path docsPath) {
+  private String formatIndexPath(Path docsPath, int numDocs) {
+    // TODO: shouldn't this use the same hashing that we use when saving exact results to cache file?
     List<String> suffix = new ArrayList<>();
     if (indexType == IndexType.FLAT) {
       suffix.add("flat");
@@ -685,6 +686,8 @@ private String formatIndexPath(Path docsPath) {
     if (parentJoin) {
       suffix.add("parentJoin");
     }
+    // make sure we reindex if numDocs has changed:
+    suffix.add(Integer.toString(numDocs));
     return INDEX_DIR + "/" + docsPath.getFileName() + "-" + String.join("-", suffix) + ".index";
   }
 

src/main/knn/KnnTesterUtils.java

@@ -32,15 +32,14 @@ public class KnnTesterUtils {
   public static int[] getResultIds(TopDocs topDocs, StoredFields storedFields) throws IOException {
     int[] resultIds = new int[topDocs.scoreDocs.length];
     int i = 0;
+    // TODO: switch to doc values for this id field?  more efficent than stored fields
+    // TODO: or, at least load the stored documents in index (Lucene docid) order to
+    //       amortize cost of decompressing each stored doc block (hmm, though, this cost/time
+    //       is not included in the reported benchy results... this is called after all KNN
+    //       queries have run)
     for (ScoreDoc doc : topDocs.scoreDocs) {
-      if (doc.doc != NO_MORE_DOCS) {
-        // there is a bug somewhere that can result in doc=NO_MORE_DOCS!  I think it happens
-        // in some degenerate case (like input query has NaN in it?) that causes no results to
-        // be returned from HNSW search?
-        resultIds[i++] = Integer.parseInt(storedFields.document(doc.doc).get(KnnGraphTester.ID_FIELD));
-      } else {
-        System.out.println("NO_MORE_DOCS!");
-      }
+      assert doc.doc != NO_MORE_DOCS: "illegal docid " + doc.doc + " returned from KNN search?";
+      resultIds[i++] = Integer.parseInt(storedFields.document(doc.doc).get(KnnGraphTester.ID_FIELD));
     }
     return resultIds;
   }

src/python/runNightlyKnn.py

@@ -292,7 +292,7 @@ def write_graph():
 </style>
 
 <div id="summary" style="height:17%%; width:95%%">
-This benchmark indexes and searches Cohere 768 dimensin vectors from https://huggingface.co/datasets/Cohere/wikipedia-22-12-en-embeddings.
+This benchmark indexes 8.0M and searches Cohere 768 dimension vectors from https://huggingface.co/datasets/Cohere/wikipedia-22-12-en-embeddings.
 </div>
 """)