Get rid of VectorByte graphs. We only use VectorBytes as sequences of…

… bytes now.

jvector-base/src/main/java/io/github/jbellis/jvector/disk/ADCGraphCache.java

@@ -46,7 +46,7 @@ public CachedNode(VectorFloat<?> vector, int[] neighbors, VectorByte<?> packedNe
     /** return the cached node if present, or null if not */
     public abstract CachedNode getNode(int ordinal);
 
-    public static ADCGraphCache load(OnDiskADCGraphIndex<VectorFloat<?>> graph, int distance) throws IOException
+    public static ADCGraphCache load(OnDiskADCGraphIndex graph, int distance) throws IOException
     {
         if (distance < 0)
             return new EmptyGraphCache();
@@ -74,7 +74,7 @@ private static final class HMGraphCache extends ADCGraphCache
         private final Int2ObjectHashMap<CachedNode> cache;
         private long ramBytesUsed = 0;
 
-        public HMGraphCache(OnDiskADCGraphIndex<VectorFloat<?>> graph, int distance) {
+        public HMGraphCache(OnDiskADCGraphIndex graph, int distance) {
             try (var view = graph.getView()) {
                 Int2ObjectHashMap<CachedNode> tmpCache = new Int2ObjectHashMap<>();
                 cacheNeighborsOf(tmpCache, view, view.entryNode(), distance);
@@ -85,7 +85,7 @@ public HMGraphCache(OnDiskADCGraphIndex<VectorFloat<?>> graph, int distance) {
             }
         }
 
-        private void cacheNeighborsOf(Int2ObjectHashMap<CachedNode> tmpCache, ADCView<VectorFloat<?>> view, int ordinal, int distance) {
+        private void cacheNeighborsOf(Int2ObjectHashMap<CachedNode> tmpCache, ADCView view, int ordinal, int distance) {
             // cache this node
             var it = view.getNeighborsIterator(ordinal);
             int[] neighbors = new int[it.size()];

jvector-base/src/main/java/io/github/jbellis/jvector/disk/CachingADCGraphIndex.java

@@ -39,14 +39,14 @@
  * TODO: Refactor so that caching is pluggable for different GraphIndex implementations.
  */
 @Experimental
-public class CachingADCGraphIndex implements GraphIndex<VectorFloat<?>>, AutoCloseable, Accountable
+public class CachingADCGraphIndex implements GraphIndex, AutoCloseable, Accountable
 {
     private static final int CACHE_DISTANCE = 3;
 
     private final ADCGraphCache cache;
-    private final OnDiskADCGraphIndex<VectorFloat<?>> graph;
+    private final OnDiskADCGraphIndex graph;
 
-    public CachingADCGraphIndex(OnDiskADCGraphIndex<VectorFloat<?>> graph)
+    public CachingADCGraphIndex(OnDiskADCGraphIndex graph)
     {
         this.graph = graph;
         try {
@@ -86,10 +86,10 @@ public void close() throws IOException {
         graph.close();
     }
 
-    public class CachedView implements ADCView<VectorFloat<?>>, ApproximateScoreProvider {
-        private final ADCView<VectorFloat<?>> view;
+    public class CachedView implements ADCView, ApproximateScoreProvider {
+        private final ADCView view;
 
-        public CachedView(ADCView<VectorFloat<?>> view) {
+        public CachedView(ADCView view) {
             this.view = view;
         }
 

jvector-base/src/main/java/io/github/jbellis/jvector/disk/CachingGraphIndex.java

@@ -25,19 +25,19 @@
 import java.io.IOException;
 import java.io.UncheckedIOException;
 
-public class CachingGraphIndex implements GraphIndex<VectorFloat<?>>, AutoCloseable, Accountable
+public class CachingGraphIndex implements GraphIndex, AutoCloseable, Accountable
 {
     private static final int CACHE_DISTANCE = 3;
 
     private final GraphCache cache;
-    private final OnDiskGraphIndex<VectorFloat<?>> graph;
+    private final OnDiskGraphIndex graph;
 
-    public CachingGraphIndex(OnDiskGraphIndex<VectorFloat<?>> graph)
+    public CachingGraphIndex(OnDiskGraphIndex graph)
     {
         this(graph, CACHE_DISTANCE);
     }
 
-    public CachingGraphIndex(OnDiskGraphIndex<VectorFloat<?>> graph, int cacheDistance)
+    public CachingGraphIndex(OnDiskGraphIndex graph, int cacheDistance)
     {
         this.graph = graph;
         try {
@@ -58,7 +58,7 @@ public NodesIterator getNodes() {
     }
 
     @Override
-    public View<VectorFloat<?>> getView() {
+    public View getView() {
         return new CachedView(graph.getView());
     }
 
@@ -77,10 +77,10 @@ public void close() throws IOException {
         graph.close();
     }
 
-    private class CachedView implements View<VectorFloat<?>> {
-        private final View<VectorFloat<?>> view;
+    private class CachedView implements View {
+        private final View view;
 
-        public CachedView(View<VectorFloat<?>> view) {
+        public CachedView(View view) {
             this.view = view;
         }
 

jvector-base/src/main/java/io/github/jbellis/jvector/disk/GraphCache.java

@@ -39,7 +39,7 @@ public CachedNode(VectorFloat<?> vector, int[] neighbors) {
     /** return the cached node if present, or null if not */
     public abstract CachedNode getNode(int ordinal);
 
-    public static GraphCache load(GraphIndex<VectorFloat<?>> graph, int distance) throws IOException
+    public static GraphCache load(GraphIndex graph, int distance) throws IOException
     {
         if (distance < 0)
             return new EmptyGraphCache();
@@ -68,7 +68,7 @@ private static final class HMGraphCache extends GraphCache
         private final Int2ObjectHashMap<CachedNode> cache;
         private long ramBytesUsed = 0;
 
-        public HMGraphCache(GraphIndex<VectorFloat<?>> graph, int distance) {
+        public HMGraphCache(GraphIndex graph, int distance) {
             try (var view = graph.getView()) {
                 Int2ObjectHashMap<CachedNode> tmpCache = new Int2ObjectHashMap<>();
                 cacheNeighborsOf(tmpCache, view, view.entryNode(), distance);
@@ -79,7 +79,7 @@ public HMGraphCache(GraphIndex<VectorFloat<?>> graph, int distance) {
             }
         }
 
-        private void cacheNeighborsOf(Int2ObjectHashMap<CachedNode> tmpCache, GraphIndex.View<VectorFloat<?>> view, int ordinal, int distance) {
+        private void cacheNeighborsOf(Int2ObjectHashMap<CachedNode> tmpCache, GraphIndex.View view, int ordinal, int distance) {
             // cache this node
             var it = view.getNeighborsIterator(ordinal);
             int[] neighbors = new int[it.size()];

jvector-base/src/main/java/io/github/jbellis/jvector/disk/OnDiskADCGraphIndex.java

@@ -52,10 +52,9 @@
  * entry points and it's fine to go to disk.
  * TODO: Permit maxDegree != 32.
  * TODO: Permit 256 PQ clusters by quantizing floats to one byte.
- * @param <T> vector type
  */
 @Experimental
-public class OnDiskADCGraphIndex<T> implements GraphIndex<T>, AutoCloseable, Accountable
+public class OnDiskADCGraphIndex implements GraphIndex, AutoCloseable, Accountable
 {
     private static final VectorTypeSupport vectorTypeSupport = VectorizationProvider.getInstance().getVectorTypeSupport();
     private final ReaderSupplier readerSupplier;
@@ -93,7 +92,7 @@ public OnDiskADCGraphIndex(ReaderSupplier readerSupplier, long offset)
      * while preserving the original relative ordering in `graph`.  That is, for all node ids i and j,
      * if i &lt; j in `graph` then map[i] &lt; map[j] in the returned map.
      */
-    public static <T> Map<Integer, Integer> getSequentialRenumbering(GraphIndex<T> graph) {
+    public static <T> Map<Integer, Integer> getSequentialRenumbering(GraphIndex graph) {
         try (var view = graph.getView()) {
             Int2IntHashMap oldToNewMap = new Int2IntHashMap(-1);
             int nextOrdinal = 0;
@@ -124,8 +123,7 @@ public OnDiskView getView()
         return new OnDiskView(readerSupplier.get());
     }
 
-    // TODO: This is fake generic until the reading functionality uses T
-    public class OnDiskView implements ADCView<T>, ApproximateScoreProvider, AutoCloseable
+    public class OnDiskView implements ADCView, ApproximateScoreProvider, AutoCloseable
     {
         private final RandomAccessReader reader;
         private final int[] neighbors;
@@ -139,13 +137,13 @@ public OnDiskView(RandomAccessReader reader)
             this.packedNeighbors = vectorTypeSupport.createByteVector(maxDegree * pqv.getCompressedSize());
         }
 
-        public T getVector(int node) {
+        public VectorFloat<?> getVector(int node) {
             try {
                 long offset = neighborsOffset +
                         node * (Integer.BYTES + (long) dimension * Float.BYTES + pqv.getCompressedSize() * maxDegree + (long) Integer.BYTES * (maxDegree + 1)) // earlier entries
                         + Integer.BYTES; // skip the ID
                 reader.seek(offset);
-                return (T) vectorTypeSupport.readFloatVector(reader, dimension);
+                return vectorTypeSupport.readFloatVector(reader, dimension);
             }
             catch (IOException e) {
                 throw new UncheckedIOException(e);
@@ -182,7 +180,7 @@ public VectorByte<?> getPackedNeighbors(int node) {
 
         @Override
         public NodeSimilarity.ApproximateScoreFunction approximateScoreFunctionFor(VectorFloat<?> query, VectorSimilarityFunction similarityFunction) {
-            return QuickADCPQDecoder.newDecoder((ADCView<VectorFloat<?>>) this, query, similarityFunction);
+            return QuickADCPQDecoder.newDecoder(this, query, similarityFunction);
         }
 
         public VectorFloat<?> reusableResults() {
@@ -236,7 +234,7 @@ public void close() throws IOException {
      *
      * If any nodes have been deleted, you must use the overload specifying `oldToNewOrdinals` instead.
      */
-    public static <T> void write(GraphIndex<T> graph, RandomAccessVectorValues<T> vectors, PQVectors pqVectors, DataOutput out)
+    public static <T> void write(GraphIndex graph, RandomAccessVectorValues vectors, PQVectors pqVectors, DataOutput out)
             throws IOException
     {
         try (var view = graph.getView()) {
@@ -259,8 +257,8 @@ public static <T> void write(GraphIndex<T> graph, RandomAccessVectorValues<T> ve
      *                  compressed representations are embedded in the serialized graph to support accelerated ADC.
      * @param out the output to write to
      */
-    public static <T> void write(GraphIndex<T> graph,
-                                 RandomAccessVectorValues<T> vectors,
+    public static <T> void write(GraphIndex graph,
+                                 RandomAccessVectorValues vectors,
                                  Map<Integer, Integer> oldToNewOrdinals,
                                  PQVectors pqVectors,
                                  DataOutput out)
@@ -275,7 +273,7 @@ public static <T> void write(GraphIndex<T> graph,
         }
 
         if (graph instanceof OnHeapGraphIndex) {
-            var ohgi = (OnHeapGraphIndex<T>) graph;
+            var ohgi = (OnHeapGraphIndex) graph;
             if (ohgi.getDeletedNodes().cardinality() > 0) {
                 throw new IllegalArgumentException("Run builder.cleanup() before writing the graph");
             }

jvector-base/src/main/java/io/github/jbellis/jvector/disk/OnDiskGraphIndex.java

@@ -35,7 +35,7 @@
 import java.util.Map;
 import java.util.stream.IntStream;
 
-public class OnDiskGraphIndex<T> implements GraphIndex<T>, AutoCloseable, Accountable
+public class OnDiskGraphIndex implements GraphIndex, AutoCloseable, Accountable
 {
     private static final VectorTypeSupport vectorTypeSupport = VectorizationProvider.getInstance().getVectorTypeSupport();
     private final ReaderSupplier readerSupplier;
@@ -65,7 +65,7 @@ public OnDiskGraphIndex(ReaderSupplier readerSupplier, long offset)
      * while preserving the original relative ordering in `graph`.  That is, for all node ids i and j,
      * if i &lt; j in `graph` then map[i] &lt; map[j] in the returned map.
      */
-    public static <T> Map<Integer, Integer> getSequentialRenumbering(GraphIndex<T> graph) {
+    public static <T> Map<Integer, Integer> getSequentialRenumbering(GraphIndex graph) {
         try (var view = graph.getView()) {
             Int2IntHashMap oldToNewMap = new Int2IntHashMap(-1);
             int nextOrdinal = 0;
@@ -91,13 +91,12 @@ public int maxDegree() {
     }
 
     /** return a Graph that can be safely queried concurrently */
-    public OnDiskGraphIndex<T>.OnDiskView getView()
+    public OnDiskGraphIndex.OnDiskView getView()
     {
         return new OnDiskView(readerSupplier.get());
     }
 
-    // TODO: This is fake generic until the reading functionality uses T
-    public class OnDiskView implements GraphIndex.View<T>, AutoCloseable
+    public class OnDiskView implements GraphIndex.View, AutoCloseable
     {
         private final RandomAccessReader reader;
         private final int[] neighbors;
@@ -109,13 +108,13 @@ public OnDiskView(RandomAccessReader reader)
             this.neighbors = new int[maxDegree];
         }
 
-        public T getVector(int node) {
+        public VectorFloat<?> getVector(int node) {
             try {
                 long offset = neighborsOffset +
                         node * (Integer.BYTES + (long) dimension * Float.BYTES + (long) Integer.BYTES * (maxDegree + 1)) // earlier entries
                         + Integer.BYTES; // skip the ID
                 reader.seek(offset);
-                return (T) vectorTypeSupport.readFloatVector(reader, dimension);
+                return vectorTypeSupport.readFloatVector(reader, dimension);
             }
             catch (IOException e) {
                 throw new UncheckedIOException(e);
@@ -180,7 +179,7 @@ public void close() throws IOException {
      *
      * If any nodes have been deleted, you must use the overload specifying `oldToNewOrdinals` instead.
      */
-    public static <T> void write(GraphIndex<T> graph, RandomAccessVectorValues<T> vectors, DataOutput out)
+    public static <T> void write(GraphIndex graph, RandomAccessVectorValues vectors, DataOutput out)
             throws IOException
     {
         try (var view = graph.getView()) {
@@ -201,14 +200,14 @@ public static <T> void write(GraphIndex<T> graph, RandomAccessVectorValues<T> ve
      *                         any deleted nodes.
      * @param out the output to write to
      */
-    public static <T> void write(GraphIndex<T> graph,
-                                 RandomAccessVectorValues<T> vectors,
+    public static void write(GraphIndex graph,
+                                 RandomAccessVectorValues vectors,
                                  Map<Integer, Integer> oldToNewOrdinals,
                                  DataOutput out)
             throws IOException
     {
         if (graph instanceof OnHeapGraphIndex) {
-            var ohgi = (OnHeapGraphIndex<T>) graph;
+            var ohgi = (OnHeapGraphIndex) graph;
             if (ohgi.getDeletedNodes().cardinality() > 0) {
                 throw new IllegalArgumentException("Run builder.cleanup() before writing the graph");
             }

jvector-base/src/main/java/io/github/jbellis/jvector/graph/ADCView.java

@@ -20,7 +20,7 @@
 import io.github.jbellis.jvector.vector.types.VectorByte;
 import io.github.jbellis.jvector.vector.types.VectorFloat;
 
-public interface ADCView<T> extends GraphIndex.View<T> {
+public interface ADCView extends GraphIndex.View {
     VectorByte<?> getPackedNeighbors(int node);
     VectorFloat<?> reusableResults();
     PQVectors getPQVectors();

jvector-base/src/main/java/io/github/jbellis/jvector/graph/GraphIndex.java

@@ -25,9 +25,9 @@
 package io.github.jbellis.jvector.graph;
 
 import io.github.jbellis.jvector.util.Bits;
+import io.github.jbellis.jvector.vector.types.VectorFloat;
 
 import java.io.IOException;
-import java.lang.ref.Cleaner;
 
 /**
  * Represents a graph-based vector index.  Nodes are represented as ints, and edges are
@@ -39,11 +39,7 @@
  * All methods are threadsafe.  Operations that require persistent state are wrapped
  * in a View that should be created per accessing thread.
  */
-public interface GraphIndex<T> extends AutoCloseable {
-    // Cleaner used for closing GraphIndex Views in unstructured contexts.
-    // This includes Views used for score functions as well as searchers.
-    Cleaner cleaner = Cleaner.create();
-
+public interface GraphIndex extends AutoCloseable {
     /** Returns the number of nodes in the graph */
     int size();
 
@@ -58,7 +54,7 @@ public interface GraphIndex<T> extends AutoCloseable {
     /**
      * Return a View with which to navigate the graph.  Views are not threadsafe.
      */
-    View<T> getView();
+    View getView();
 
     /**
      * @return the maximum number of edges per node
@@ -83,7 +79,7 @@ default boolean containsNode(int nodeId) {
     @Override
     void close() throws IOException;
 
-    interface View<T> extends AutoCloseable {
+    interface View extends AutoCloseable {
         /**
          * Iterator over the neighbors of a given node.  Only the most recently instantiated iterator
          * is guaranteed to be valid.
@@ -107,7 +103,7 @@ interface View<T> extends AutoCloseable {
          * In that situation, we will want to reorder the results by the exact similarity
          * at the end of the search.
          */
-        T getVector(int node);
+        VectorFloat<?> getVector(int node);
 
         /**
          * Return a Bits instance indicating which nodes are live.  The result is undefined for
@@ -123,7 +119,7 @@ default int getIdUpperBound() {
         }
     }
 
-    static <T> String prettyPrint(GraphIndex<T> graph) {
+    static <T> String prettyPrint(GraphIndex graph) {
         StringBuilder sb = new StringBuilder();
         sb.append(graph);
         sb.append("\n");