mccullocht commented on code in PR #16030: URL: https://github.com/apache/lucene/pull/16030#discussion_r3910424627
########## lucene/core/src/java/org/apache/lucene/codecs/lucene106/Lucene106ScalarQuantizedVectorsFormat.java: ########## @@ -0,0 +1,170 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.lucene.codecs.lucene106; + +import java.io.IOException; +import org.apache.lucene.codecs.hnsw.FlatVectorScorerUtil; +import org.apache.lucene.codecs.hnsw.FlatVectorsFormat; +import org.apache.lucene.codecs.hnsw.FlatVectorsReader; +import org.apache.lucene.codecs.hnsw.FlatVectorsWriter; +import org.apache.lucene.codecs.lucene99.Lucene99FlatVectorsFormat; +import org.apache.lucene.index.SegmentReadState; +import org.apache.lucene.index.SegmentWriteState; +import org.apache.lucene.util.quantization.QuantizedByteVectorValues.ScalarEncoding; + +/** + * The quantization format used here is a per-vector optimized scalar quantization. These ideas are + * evolutions of LVQ proposed in <a href="https://arxiv.org/abs/2304.04759";>Similarity search in the + * blink of an eye with compressed indices</a> by Cecilia Aguerrebere et al., the previous work on + * globally optimized scalar quantization in Apache Lucene, and <a + * href="https://arxiv.org/abs/1908.10396";>Accelerating Large-Scale Inference with Anisotropic + * Vector Quantization </a> by Ruiqi Guo et. al. Also see {@link + * org.apache.lucene.util.quantization.OptimizedScalarQuantizer}. Some of key features are: + * + * <ul> + * <li>Estimating the distance between two vectors using their centroid centered distance. This + * requires some additional corrective factors, but allows for centroid centering to occur. + * <li>Optimized scalar quantization to single bit level of centroid centered vectors. + * <li>Asymmetric quantization of vectors, where query vectors are quantized to half-byte (4 bits) + * precision (normalized to the centroid) and then compared directly against the single bit + * quantized vectors in the index. + * <li>Transforming the half-byte quantized query vectors in such a way that the comparison with + * single bit vectors can be done with bit arithmetic. + * </ul> + * + * A previous work related to improvements over regular LVQ is <a + * href="https://arxiv.org/abs/2409.09913";>Practical and Asymptotically Optimal Quantization of + * High-Dimensional Vectors in Euclidean Space for Approximate Nearest Neighbor Search </a> by + * Jianyang Gao, et. al. + * + * <p>The format is stored within two files: + * + * <h2>.veq (vector data) file</h2> + * + * <p>Stores the quantized vectors in a flat format. Additionally, it stores each vector's + * corrective factors. At the end of the file, additional information is stored for vector ordinal + * to centroid ordinal mapping and sparse vector information. + * + * <ul> + * <li>For each vector: + * <ul> + * <li><b>[byte]</b> the quantized values. Each dimension may be up to 8 bits, and multiple + * dimensions may be packed into a single byte. + * <li><b>[float]</b> the optimized quantiles and an additional similarity dependent + * corrective factor. + * <li><b>[int]</b> the sum of the quantized components + * </ul> + * <li>After the vectors, sparse vector information keeping track of monotonic blocks. + * </ul> + * + * <h2>.vemq (vector metadata) file</h2> + * + * <p>Stores the metadata for the vectors. This includes the number of vectors, the number of + * dimensions, and file offset information. + * + * <ul> + * <li><b>int</b> the field number + * <li><b>int</b> the vector encoding ordinal + * <li><b>int</b> the vector similarity ordinal + * <li><b>vint</b> the vector dimensions + * <li><b>vlong</b> the offset to the vector data in the .veq file + * <li><b>vlong</b> the length of the vector data in the .veq file + * <li><b>vint</b> the number of vectors + * <li><b>vint</b> the wire number for ScalarEncoding + * <li><b>[float]</b> the centroid + * <li><b>float</b> the centroid square magnitude + * <li>The sparse vector information, if required, mapping vector ordinal to doc ID Review Comment: This is largely the same. There's no guidance on this and I probably could merge it back into the Lucene104 codec as a version bump. I have mixed feelings about new version vs codec internal version bump because it's less obvious with the codec internal version bump that I am breaking you on upgrade. ########## lucene/core/src/java/org/apache/lucene/codecs/lucene106/Lucene106ScalarQuantizedVectorsReader.java: ########## @@ -0,0 +1,738 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.lucene.codecs.lucene106; + +import static org.apache.lucene.codecs.lucene106.Lucene106ScalarQuantizedVectorsFormat.VECTOR_DATA_EXTENSION; +import static org.apache.lucene.codecs.lucene99.Lucene99HnswVectorsReader.readSimilarityFunction; +import static org.apache.lucene.codecs.lucene99.Lucene99HnswVectorsReader.readVectorEncoding; +import static org.apache.lucene.search.DocIdSetIterator.NO_MORE_DOCS; +import static org.apache.lucene.util.quantization.OptimizedScalarQuantizer.transposeHalfByte; + +import java.io.IOException; +import java.util.HashMap; +import java.util.Map; +import java.util.Objects; +import java.util.stream.Stream; +import org.apache.lucene.codecs.CodecUtil; +import org.apache.lucene.codecs.KnnVectorsReader; +import org.apache.lucene.codecs.hnsw.FlatVectorsReader; +import org.apache.lucene.codecs.hnsw.FlatVectorsScorer; +import org.apache.lucene.codecs.lucene95.OrdToDocDISIReaderConfiguration; +import org.apache.lucene.index.ByteVectorValues; +import org.apache.lucene.index.CorruptIndexException; +import org.apache.lucene.index.DocsWithFieldSet; +import org.apache.lucene.index.FieldInfo; +import org.apache.lucene.index.FieldInfos; +import org.apache.lucene.index.Float16VectorValues; +import org.apache.lucene.index.FloatVectorValues; +import org.apache.lucene.index.IndexFileNames; +import org.apache.lucene.index.KnnVectorValues; +import org.apache.lucene.index.MergePolicy; +import org.apache.lucene.index.SegmentReadState; +import org.apache.lucene.index.SegmentWriteState; +import org.apache.lucene.index.VectorEncoding; +import org.apache.lucene.index.VectorSimilarityFunction; +import org.apache.lucene.search.AcceptDocs; +import org.apache.lucene.search.KnnCollector; +import org.apache.lucene.search.VectorScorer; +import org.apache.lucene.store.ChecksumIndexInput; +import org.apache.lucene.store.DataAccessHint; +import org.apache.lucene.store.FileDataHint; +import org.apache.lucene.store.FileTypeHint; +import org.apache.lucene.store.IOContext; +import org.apache.lucene.store.IndexInput; +import org.apache.lucene.store.IndexOutput; +import org.apache.lucene.util.Bits; +import org.apache.lucene.util.IOUtils; +import org.apache.lucene.util.RamUsageEstimator; +import org.apache.lucene.util.hnsw.CloseableRandomVectorScorerSupplier; +import org.apache.lucene.util.hnsw.RandomVectorScorer; +import org.apache.lucene.util.hnsw.RandomVectorScorerSupplier; +import org.apache.lucene.util.quantization.OptimizedScalarQuantizer; +import org.apache.lucene.util.quantization.QuantizedByteVectorValues; +import org.apache.lucene.util.quantization.QuantizedByteVectorValues.ScalarEncoding; +import org.apache.lucene.util.quantization.QuantizedVectorsReader; +import org.apache.lucene.util.quantization.ScalarQuantizer; + +/** + * Reader for scalar quantized vectors in the Lucene 10.5 format. + * + * @lucene.experimental + */ +public class Lucene106ScalarQuantizedVectorsReader extends FlatVectorsReader + implements QuantizedVectorsReader { + + private static final long SHALLOW_SIZE = + RamUsageEstimator.shallowSizeOfInstance(Lucene106ScalarQuantizedVectorsReader.class); + + private final Map<String, FieldEntry> fields = new HashMap<>(); + private final IndexInput quantizedVectorData; + private final FlatVectorsReader rawVectorsReader; + private final Lucene106ScalarQuantizedVectorScorer vectorScorer; + public static final int EXHAUSTIVE_BULK_SCORE_ORDS = 64; + + public Lucene106ScalarQuantizedVectorsReader( + SegmentReadState state, + FlatVectorsReader rawVectorsReader, + Lucene106ScalarQuantizedVectorScorer vectorsScorer) + throws IOException { + // Quantized vectors are accessed randomly from their node ID stored in the HNSW + // graph. + this(state, rawVectorsReader, vectorsScorer, DataAccessHint.RANDOM); + } + + public Lucene106ScalarQuantizedVectorsReader( + SegmentReadState state, + FlatVectorsReader rawVectorsReader, + Lucene106ScalarQuantizedVectorScorer vectorsScorer, + DataAccessHint accessHint) + throws IOException { + this.vectorScorer = vectorsScorer; + this.rawVectorsReader = rawVectorsReader; + int versionMeta = -1; + String metaFileName = + IndexFileNames.segmentFileName( + state.segmentInfo.name, + state.segmentSuffix, + Lucene106ScalarQuantizedVectorsFormat.META_EXTENSION); + try (ChecksumIndexInput meta = state.directory.openChecksumInput(metaFileName)) { + Throwable priorE = null; + try { + versionMeta = + CodecUtil.checkIndexHeader( + meta, + Lucene106ScalarQuantizedVectorsFormat.META_CODEC_NAME, + Lucene106ScalarQuantizedVectorsFormat.VERSION_START, + Lucene106ScalarQuantizedVectorsFormat.VERSION_CURRENT, + state.segmentInfo.getId(), + state.segmentSuffix); + readFields(meta, state.fieldInfos); + } catch (Throwable exception) { + priorE = exception; + } finally { + CodecUtil.checkFooter(meta, priorE); + } + + final IOContext.FileOpenHint[] hints = + Stream.of(FileTypeHint.DATA, FileDataHint.KNN_VECTORS, accessHint) + .filter(Objects::nonNull) + .toArray(IOContext.FileOpenHint[]::new); + quantizedVectorData = + openDataInput( + state, + versionMeta, + VECTOR_DATA_EXTENSION, + Lucene106ScalarQuantizedVectorsFormat.VECTOR_DATA_CODEC_NAME, + state.context.withHints(hints)); + } catch (Throwable t) { + IOUtils.closeWhileSuppressingExceptions(t, this); + throw t; + } + } + + private void readFields(ChecksumIndexInput meta, FieldInfos infos) throws IOException { + for (int fieldNumber = meta.readInt(); fieldNumber != -1; fieldNumber = meta.readInt()) { + FieldInfo info = infos.fieldInfo(fieldNumber); + if (info == null) { + throw new CorruptIndexException("Invalid field number: " + fieldNumber, meta); + } + FieldEntry fieldEntry = readField(meta, info); + validateFieldEntry(info, fieldEntry); + fields.put(info.name, fieldEntry); + } + } + + static void validateFieldEntry(FieldInfo info, FieldEntry fieldEntry) { + int dimension = info.getVectorDimension(); + if (dimension != fieldEntry.dimension) { + throw new IllegalStateException( + "Inconsistent vector dimension for field=\"" + + info.name + + "\"; " + + dimension + + " != " + + fieldEntry.dimension); + } + + long numQuantizedVectorBytes = + Math.multiplyExact( + (fieldEntry.scalarEncoding.getDocPackedLength(dimension) + + (Float.BYTES * 3) + + Integer.BYTES), + (long) fieldEntry.size); + if (numQuantizedVectorBytes != fieldEntry.vectorDataLength) { + throw new IllegalStateException( + "vector data length " + + fieldEntry.vectorDataLength + + " not matching size = " + + fieldEntry.size + + " * (dims=" + + dimension + + " + 16" + + ") = " + + numQuantizedVectorBytes); + } + } + + @Override + public FlatVectorsScorer getFlatVectorScorer(String field) throws IOException { + return vectorScorer; + } + + @Override + public RandomVectorScorer getRandomVectorScorer(String field, float[] target) throws IOException { + FieldEntry fi = fields.get(field); + if (fi == null) { + return null; + } + return vectorScorer.getRandomVectorScorer( + fi.similarityFunction, + OffHeapScalarQuantizedVectorValues.load( + fi.ordToDocDISIReaderConfiguration, + fi.dimension, + fi.size, + new OptimizedScalarQuantizer(fi.similarityFunction), + fi.scalarEncoding, + fi.similarityFunction, + vectorScorer, + fi.centroid, + fi.centroidDP, + fi.vectorDataOffset, + fi.vectorDataLength, + quantizedVectorData), + target); + } + + @Override + public RandomVectorScorer getRandomVectorScorer(String field, short[] target) throws IOException { + return rawVectorsReader.getRandomVectorScorer(field, target); + } + + @Override + public RandomVectorScorer getRandomVectorScorer(String field, byte[] target) throws IOException { + return rawVectorsReader.getRandomVectorScorer(field, target); + } + + @Override + public void checkIntegrity(MergePolicy.OneMerge merge) throws IOException { + rawVectorsReader.checkIntegrity(merge); + CodecUtil.checksumEntireFile(quantizedVectorData); + } + + @Override + public FloatVectorValues getFloatVectorValues(String field) throws IOException { + FieldEntry fi = fields.get(field); + if (fi == null) { + return null; + } + if (fi.vectorEncoding != VectorEncoding.FLOAT32) { + throw new IllegalArgumentException( + "field=\"" + + field + + "\" is encoded as: " + + fi.vectorEncoding + + " expected: " + + VectorEncoding.FLOAT32); + } + + FloatVectorValues rawFloatVectorValues = + fi.isDataBlind() ? null : rawVectorsReader.getFloatVectorValues(field); + + if (rawFloatVectorValues == null) { + // Data-blind mode: full-precision float vectors were never stored. Reconstruct floats from Review Comment: There might be some use for this in some very high dimensional data sets -- I may want to quantize the original vectors at a high bit rate (4 or 8) bits and discard them, and also provide 1 or 2 bit quantization with centering that are used for graph navigation. In your hypothetical situation yes you would absolutely lose precision, how much depends on how aggressive the initial quantization is. ########## lucene/core/src/java/org/apache/lucene/codecs/lucene106/Lucene106ScalarQuantizedVectorsReader.java: ########## @@ -0,0 +1,738 @@ +/* + * Licensed to the Apache Software Foundation (ASF) under one or more + * contributor license agreements. See the NOTICE file distributed with + * this work for additional information regarding copyright ownership. + * The ASF licenses this file to You under the Apache License, Version 2.0 + * (the "License"); you may not use this file except in compliance with + * the License. You may obtain a copy of the License at + * + * http://www.apache.org/licenses/LICENSE-2.0 + * + * Unless required by applicable law or agreed to in writing, software + * distributed under the License is distributed on an "AS IS" BASIS, + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. + * See the License for the specific language governing permissions and + * limitations under the License. + */ +package org.apache.lucene.codecs.lucene106; + +import static org.apache.lucene.codecs.lucene106.Lucene106ScalarQuantizedVectorsFormat.VECTOR_DATA_EXTENSION; +import static org.apache.lucene.codecs.lucene99.Lucene99HnswVectorsReader.readSimilarityFunction; +import static org.apache.lucene.codecs.lucene99.Lucene99HnswVectorsReader.readVectorEncoding; +import static org.apache.lucene.search.DocIdSetIterator.NO_MORE_DOCS; +import static org.apache.lucene.util.quantization.OptimizedScalarQuantizer.transposeHalfByte; + +import java.io.IOException; +import java.util.HashMap; +import java.util.Map; +import java.util.Objects; +import java.util.stream.Stream; +import org.apache.lucene.codecs.CodecUtil; +import org.apache.lucene.codecs.KnnVectorsReader; +import org.apache.lucene.codecs.hnsw.FlatVectorsReader; +import org.apache.lucene.codecs.hnsw.FlatVectorsScorer; +import org.apache.lucene.codecs.lucene95.OrdToDocDISIReaderConfiguration; +import org.apache.lucene.index.ByteVectorValues; +import org.apache.lucene.index.CorruptIndexException; +import org.apache.lucene.index.DocsWithFieldSet; +import org.apache.lucene.index.FieldInfo; +import org.apache.lucene.index.FieldInfos; +import org.apache.lucene.index.Float16VectorValues; +import org.apache.lucene.index.FloatVectorValues; +import org.apache.lucene.index.IndexFileNames; +import org.apache.lucene.index.KnnVectorValues; +import org.apache.lucene.index.MergePolicy; +import org.apache.lucene.index.SegmentReadState; +import org.apache.lucene.index.SegmentWriteState; +import org.apache.lucene.index.VectorEncoding; +import org.apache.lucene.index.VectorSimilarityFunction; +import org.apache.lucene.search.AcceptDocs; +import org.apache.lucene.search.KnnCollector; +import org.apache.lucene.search.VectorScorer; +import org.apache.lucene.store.ChecksumIndexInput; +import org.apache.lucene.store.DataAccessHint; +import org.apache.lucene.store.FileDataHint; +import org.apache.lucene.store.FileTypeHint; +import org.apache.lucene.store.IOContext; +import org.apache.lucene.store.IndexInput; +import org.apache.lucene.store.IndexOutput; +import org.apache.lucene.util.Bits; +import org.apache.lucene.util.IOUtils; +import org.apache.lucene.util.RamUsageEstimator; +import org.apache.lucene.util.hnsw.CloseableRandomVectorScorerSupplier; +import org.apache.lucene.util.hnsw.RandomVectorScorer; +import org.apache.lucene.util.hnsw.RandomVectorScorerSupplier; +import org.apache.lucene.util.quantization.OptimizedScalarQuantizer; +import org.apache.lucene.util.quantization.QuantizedByteVectorValues; +import org.apache.lucene.util.quantization.QuantizedByteVectorValues.ScalarEncoding; +import org.apache.lucene.util.quantization.QuantizedVectorsReader; +import org.apache.lucene.util.quantization.ScalarQuantizer; + +/** + * Reader for scalar quantized vectors in the Lucene 10.5 format. + * + * @lucene.experimental + */ +public class Lucene106ScalarQuantizedVectorsReader extends FlatVectorsReader + implements QuantizedVectorsReader { + + private static final long SHALLOW_SIZE = + RamUsageEstimator.shallowSizeOfInstance(Lucene106ScalarQuantizedVectorsReader.class); + + private final Map<String, FieldEntry> fields = new HashMap<>(); + private final IndexInput quantizedVectorData; + private final FlatVectorsReader rawVectorsReader; + private final Lucene106ScalarQuantizedVectorScorer vectorScorer; + public static final int EXHAUSTIVE_BULK_SCORE_ORDS = 64; + + public Lucene106ScalarQuantizedVectorsReader( + SegmentReadState state, + FlatVectorsReader rawVectorsReader, + Lucene106ScalarQuantizedVectorScorer vectorsScorer) + throws IOException { + // Quantized vectors are accessed randomly from their node ID stored in the HNSW + // graph. + this(state, rawVectorsReader, vectorsScorer, DataAccessHint.RANDOM); + } + + public Lucene106ScalarQuantizedVectorsReader( + SegmentReadState state, + FlatVectorsReader rawVectorsReader, + Lucene106ScalarQuantizedVectorScorer vectorsScorer, + DataAccessHint accessHint) + throws IOException { + this.vectorScorer = vectorsScorer; + this.rawVectorsReader = rawVectorsReader; + int versionMeta = -1; + String metaFileName = + IndexFileNames.segmentFileName( + state.segmentInfo.name, + state.segmentSuffix, + Lucene106ScalarQuantizedVectorsFormat.META_EXTENSION); + try (ChecksumIndexInput meta = state.directory.openChecksumInput(metaFileName)) { + Throwable priorE = null; + try { + versionMeta = + CodecUtil.checkIndexHeader( + meta, + Lucene106ScalarQuantizedVectorsFormat.META_CODEC_NAME, + Lucene106ScalarQuantizedVectorsFormat.VERSION_START, + Lucene106ScalarQuantizedVectorsFormat.VERSION_CURRENT, + state.segmentInfo.getId(), + state.segmentSuffix); + readFields(meta, state.fieldInfos); + } catch (Throwable exception) { + priorE = exception; + } finally { + CodecUtil.checkFooter(meta, priorE); + } + + final IOContext.FileOpenHint[] hints = + Stream.of(FileTypeHint.DATA, FileDataHint.KNN_VECTORS, accessHint) + .filter(Objects::nonNull) + .toArray(IOContext.FileOpenHint[]::new); + quantizedVectorData = + openDataInput( + state, + versionMeta, + VECTOR_DATA_EXTENSION, + Lucene106ScalarQuantizedVectorsFormat.VECTOR_DATA_CODEC_NAME, + state.context.withHints(hints)); + } catch (Throwable t) { + IOUtils.closeWhileSuppressingExceptions(t, this); + throw t; + } + } + + private void readFields(ChecksumIndexInput meta, FieldInfos infos) throws IOException { + for (int fieldNumber = meta.readInt(); fieldNumber != -1; fieldNumber = meta.readInt()) { + FieldInfo info = infos.fieldInfo(fieldNumber); + if (info == null) { + throw new CorruptIndexException("Invalid field number: " + fieldNumber, meta); + } + FieldEntry fieldEntry = readField(meta, info); + validateFieldEntry(info, fieldEntry); + fields.put(info.name, fieldEntry); + } + } + + static void validateFieldEntry(FieldInfo info, FieldEntry fieldEntry) { + int dimension = info.getVectorDimension(); + if (dimension != fieldEntry.dimension) { + throw new IllegalStateException( + "Inconsistent vector dimension for field=\"" + + info.name + + "\"; " + + dimension + + " != " + + fieldEntry.dimension); + } + + long numQuantizedVectorBytes = + Math.multiplyExact( + (fieldEntry.scalarEncoding.getDocPackedLength(dimension) + + (Float.BYTES * 3) + + Integer.BYTES), + (long) fieldEntry.size); + if (numQuantizedVectorBytes != fieldEntry.vectorDataLength) { + throw new IllegalStateException( + "vector data length " + + fieldEntry.vectorDataLength + + " not matching size = " + + fieldEntry.size + + " * (dims=" + + dimension + + " + 16" + + ") = " + + numQuantizedVectorBytes); + } + } + + @Override + public FlatVectorsScorer getFlatVectorScorer(String field) throws IOException { + return vectorScorer; + } + + @Override + public RandomVectorScorer getRandomVectorScorer(String field, float[] target) throws IOException { + FieldEntry fi = fields.get(field); + if (fi == null) { + return null; + } + return vectorScorer.getRandomVectorScorer( + fi.similarityFunction, + OffHeapScalarQuantizedVectorValues.load( + fi.ordToDocDISIReaderConfiguration, + fi.dimension, + fi.size, + new OptimizedScalarQuantizer(fi.similarityFunction), + fi.scalarEncoding, + fi.similarityFunction, + vectorScorer, + fi.centroid, + fi.centroidDP, + fi.vectorDataOffset, + fi.vectorDataLength, + quantizedVectorData), + target); + } + + @Override + public RandomVectorScorer getRandomVectorScorer(String field, short[] target) throws IOException { + return rawVectorsReader.getRandomVectorScorer(field, target); + } + + @Override + public RandomVectorScorer getRandomVectorScorer(String field, byte[] target) throws IOException { + return rawVectorsReader.getRandomVectorScorer(field, target); + } + + @Override + public void checkIntegrity(MergePolicy.OneMerge merge) throws IOException { + rawVectorsReader.checkIntegrity(merge); + CodecUtil.checksumEntireFile(quantizedVectorData); + } + + @Override + public FloatVectorValues getFloatVectorValues(String field) throws IOException { + FieldEntry fi = fields.get(field); + if (fi == null) { + return null; + } + if (fi.vectorEncoding != VectorEncoding.FLOAT32) { + throw new IllegalArgumentException( + "field=\"" + + field + + "\" is encoded as: " + + fi.vectorEncoding + + " expected: " + + VectorEncoding.FLOAT32); + } + + FloatVectorValues rawFloatVectorValues = + fi.isDataBlind() ? null : rawVectorsReader.getFloatVectorValues(field); + + if (rawFloatVectorValues == null) { + // Data-blind mode: full-precision float vectors were never stored. Reconstruct floats from + // the quantized data. + return OffHeapScalarQuantizedFloatVectorValues.load( + fi.ordToDocDISIReaderConfiguration, + fi.dimension, + fi.size, + fi.scalarEncoding, + fi.similarityFunction, + vectorScorer, + fi.centroid, + fi.vectorDataOffset, + fi.vectorDataLength, + quantizedVectorData); + } + + OffHeapScalarQuantizedVectorValues sqvv = + OffHeapScalarQuantizedVectorValues.load( + fi.ordToDocDISIReaderConfiguration, + fi.dimension, + fi.size, + new OptimizedScalarQuantizer(fi.similarityFunction), + fi.scalarEncoding, + fi.similarityFunction, + vectorScorer, + fi.centroid, + fi.centroidDP, + fi.vectorDataOffset, + fi.vectorDataLength, + quantizedVectorData); + + if (rawFloatVectorValues.size() == 0) { + // Full-precision vectors were dropped after writing. Wrap the dequantizing read view with + // sqvv so scorer() stays quantized (as in the branch where raw vectors are present) while + // vectorValue()/rescorer() dequantize. Passing the dequantized view straight to a scorer + // would misroute to the non-quantized flat scorer over the quantized slice. + FloatVectorValues dequantizedRawVectorValues = + OffHeapScalarQuantizedFloatVectorValues.load( + fi.ordToDocDISIReaderConfiguration, + fi.dimension, + fi.size, + fi.scalarEncoding, + fi.similarityFunction, + vectorScorer, + fi.centroid, + fi.vectorDataOffset, + fi.vectorDataLength, + quantizedVectorData); + return new ScalarQuantizedVectorValues(dequantizedRawVectorValues, sqvv); + } + + return new ScalarQuantizedVectorValues(rawFloatVectorValues, sqvv); + } + + @Override + public Float16VectorValues getFloat16VectorValues(String field) throws IOException { + return rawVectorsReader.getFloat16VectorValues(field); + } + + @Override + public ByteVectorValues getByteVectorValues(String field) throws IOException { + return rawVectorsReader.getByteVectorValues(field); + } + + @Override + public void search(String field, byte[] target, KnnCollector knnCollector, AcceptDocs acceptDocs) + throws IOException { + rawVectorsReader.search(field, target, knnCollector, acceptDocs); + } + + @Override + public void search(String field, short[] target, KnnCollector knnCollector, AcceptDocs acceptDocs) + throws IOException { + rawVectorsReader.search(field, target, knnCollector, acceptDocs); + } + + @Override + public void search(String field, float[] target, KnnCollector knnCollector, AcceptDocs acceptDocs) + throws IOException { + if (knnCollector.k() == 0) return; + final RandomVectorScorer scorer = getRandomVectorScorer(field, target); + if (scorer == null) return; + Bits acceptedOrds = scorer.getAcceptOrds(acceptDocs.bits()); + // if k is larger than the number of vectors we expect to visit in an HNSW search, + // we can just iterate over all vectors and collect them. + int[] ords = new int[EXHAUSTIVE_BULK_SCORE_ORDS]; + float[] scores = new float[EXHAUSTIVE_BULK_SCORE_ORDS]; + int numOrds = 0; + int numVectors = scorer.maxOrd(); + for (int i = 0; i < numVectors; i++) { + if (acceptedOrds == null || acceptedOrds.get(i)) { + if (knnCollector.earlyTerminated()) { + break; + } + ords[numOrds++] = i; + if (numOrds == ords.length) { + knnCollector.incVisitedCount(numOrds); + if (scorer.bulkScore(ords, scores, numOrds) > knnCollector.minCompetitiveSimilarity()) { + for (int j = 0; j < numOrds; j++) { + knnCollector.collect(scorer.ordToDoc(ords[j]), scores[j]); + } + } + numOrds = 0; + } + } + } + + if (numOrds > 0) { + knnCollector.incVisitedCount(numOrds); + if (scorer.bulkScore(ords, scores, numOrds) > knnCollector.minCompetitiveSimilarity()) { + for (int j = 0; j < numOrds; j++) { + knnCollector.collect(scorer.ordToDoc(ords[j]), scores[j]); + } + } + } + } + + @Override + public void close() throws IOException { + IOUtils.close(quantizedVectorData, rawVectorsReader); + } + + @Override + public long ramBytesUsed() { + long size = SHALLOW_SIZE; + size += + RamUsageEstimator.sizeOfMap( + fields, RamUsageEstimator.shallowSizeOfInstance(FieldEntry.class)); + size += rawVectorsReader.ramBytesUsed(); + return size; + } + + @Override + public Map<String, Long> getOffHeapByteSize(FieldInfo fieldInfo) { + Objects.requireNonNull(fieldInfo); + var raw = rawVectorsReader.getOffHeapByteSize(fieldInfo); + var fieldEntry = fields.get(fieldInfo.name); + if (fieldEntry == null) { + // Only FLOAT32 fields are scalar-quantized by this format; BYTE and FLOAT16 fields are + // stored raw by the delegate and therefore have no quantized field entry here. + assert fieldInfo.getVectorEncoding() == VectorEncoding.BYTE + || fieldInfo.getVectorEncoding() == VectorEncoding.FLOAT16; + return raw; + } + var quant = Map.of(VECTOR_DATA_EXTENSION, fieldEntry.vectorDataLength()); + return KnnVectorsReader.mergeOffHeapByteSizeMaps(raw, quant); + } + + public float[] getCentroid(String field) { + FieldEntry fieldEntry = fields.get(field); + if (fieldEntry != null) { + return fieldEntry.centroid; + } + return null; + } + + boolean hasRawFloatVectors(String field) throws IOException { + FieldEntry fi = fields.get(field); + if (fi == null || fi.isDataBlind()) { + return false; + } + FloatVectorValues raw = rawVectorsReader.getFloatVectorValues(field); + return raw != null && raw.size() > 0; + } + + private static IndexInput openDataInput( + SegmentReadState state, + int versionMeta, + String fileExtension, + String codecName, + IOContext context) + throws IOException { + String fileName = + IndexFileNames.segmentFileName(state.segmentInfo.name, state.segmentSuffix, fileExtension); + IndexInput in = state.directory.openInput(fileName, context); + try { + int versionVectorData = + CodecUtil.checkIndexHeader( + in, + codecName, + Lucene106ScalarQuantizedVectorsFormat.VERSION_START, + Lucene106ScalarQuantizedVectorsFormat.VERSION_CURRENT, + state.segmentInfo.getId(), + state.segmentSuffix); + if (versionMeta != versionVectorData) { + throw new CorruptIndexException( + "Format versions mismatch: meta=" + + versionMeta + + ", " + + codecName + + "=" + + versionVectorData, + in); + } + CodecUtil.retrieveChecksum(in); + return in; + } catch (Throwable t) { + IOUtils.closeWhileSuppressingExceptions(t, in); + throw t; + } + } + + private FieldEntry readField(IndexInput input, FieldInfo info) throws IOException { + VectorEncoding vectorEncoding = readVectorEncoding(input); + VectorSimilarityFunction similarityFunction = readSimilarityFunction(input); + if (similarityFunction != info.getVectorSimilarityFunction()) { + throw new IllegalStateException( + "Inconsistent vector similarity function for field=\"" + + info.name + + "\"; " + + similarityFunction + + " != " + + info.getVectorSimilarityFunction()); + } + return FieldEntry.create(input, vectorEncoding, info.getVectorSimilarityFunction()); + } + + @Override + public QuantizedByteVectorValues getQuantizedVectorValues(String field) throws IOException { + FieldEntry fi = fields.get(field); + if (fi == null) { + return null; + } + if (fi.vectorEncoding != VectorEncoding.FLOAT32) { + throw new IllegalArgumentException( + "field=\"" + + field + + "\" is encoded as: " + + fi.vectorEncoding + + " expected: " + + VectorEncoding.FLOAT32); + } + return OffHeapScalarQuantizedVectorValues.load( + fi.ordToDocDISIReaderConfiguration, + fi.dimension, + fi.size, + new OptimizedScalarQuantizer(fi.similarityFunction), + fi.scalarEncoding, + fi.similarityFunction, + vectorScorer, + fi.centroid, + fi.centroidDP, + fi.vectorDataOffset, + fi.vectorDataLength, + quantizedVectorData); + } + + @Override + public ScalarQuantizer getQuantizationState(String fieldName) { + return null; + } + + @Override + public CloseableRandomVectorScorerSupplier getRandomVectorScorerSupplierForMerge( + FieldInfo fieldInfo, SegmentWriteState segmentWriteState) throws IOException { + FieldEntry fi = fields.get(fieldInfo.name); + if (fi == null) { + return null; + } + QuantizedByteVectorValues vectorValues = getQuantizedVectorValues(fieldInfo.name); + if (fi.scalarEncoding.isAsymmetric() == false) { + RandomVectorScorerSupplier supplier = + vectorScorer.getRandomVectorScorerSupplier( + fieldInfo.getVectorSimilarityFunction(), vectorValues); + return CloseableRandomVectorScorerSupplier.create(supplier, vectorValues.size(), () -> {}); + } + FloatVectorValues floatVectorValues = getFloatVectorValues(fieldInfo.name); + OptimizedScalarQuantizer quantizer = + new OptimizedScalarQuantizer(fieldInfo.getVectorSimilarityFunction()); + String tempScoreQuantizedVectorName = null; + DocsWithFieldSet docsWithField; + try (IndexOutput tempScoreQuantizedVector = + segmentWriteState.directory.createTempOutput( + segmentWriteState.segmentInfo.name, "queries", segmentWriteState.context)) { + tempScoreQuantizedVectorName = tempScoreQuantizedVector.getName(); + docsWithField = + writeBinarizedQueryData( + vectorValues, + fi.scalarEncoding, + tempScoreQuantizedVector, + floatVectorValues, + quantizer); + CodecUtil.writeFooter(tempScoreQuantizedVector); + } catch (Throwable t) { + if (tempScoreQuantizedVectorName != null) { + IOUtils.deleteFilesSuppressingExceptions( + t, segmentWriteState.directory, tempScoreQuantizedVectorName); + } + throw t; + } + IndexInput quantizedScoreDataInput = + segmentWriteState.directory.openInput( + tempScoreQuantizedVectorName, segmentWriteState.context); + try { + OffHeapScalarQuantizedVectorValues scoreVectorValues = + new OffHeapScalarQuantizedVectorValues.DenseOffHeapVectorValues( + true, + fieldInfo.getVectorDimension(), + docsWithField.cardinality(), + vectorValues.getCentroid(), + vectorValues.getCentroidDP(), + quantizer, + fi.scalarEncoding, + fieldInfo.getVectorSimilarityFunction(), + vectorScorer, + quantizedScoreDataInput); + RandomVectorScorerSupplier scorerSupplier = + vectorScorer.getRandomVectorScorerSupplier( + fieldInfo.getVectorSimilarityFunction(), scoreVectorValues, vectorValues); + final String finalTempScoreQuantizedVectorName = tempScoreQuantizedVectorName; + return CloseableRandomVectorScorerSupplier.create( + scorerSupplier, + vectorValues.size(), + () -> { + IOUtils.close(quantizedScoreDataInput); + IOUtils.deleteFilesIgnoringExceptions( + segmentWriteState.directory, finalTempScoreQuantizedVectorName); + }); + } catch (Throwable t) { + IOUtils.closeWhileSuppressingExceptions(t, quantizedScoreDataInput); + throw t; + } + } + + static DocsWithFieldSet writeBinarizedQueryData( + QuantizedByteVectorValues quantizedByteVectorValues, + ScalarEncoding encoding, + IndexOutput binarizedQueryData, + FloatVectorValues floatVectorValues, + OptimizedScalarQuantizer binaryQuantizer) + throws IOException { + if (encoding.isAsymmetric() == false) { + throw new IllegalArgumentException("encoding and queryEncoding must be different"); + } + DocsWithFieldSet docsWithField = new DocsWithFieldSet(); + int discretizedDims = encoding.getDiscreteDimensions(floatVectorValues.dimension()); + byte[] quantizationScratch = new byte[discretizedDims]; + byte[] toQuery = new byte[encoding.getQueryPackedLength(discretizedDims)]; + KnnVectorValues.DocIndexIterator iterator = floatVectorValues.iterator(); + for (int docV = iterator.nextDoc(); docV != NO_MORE_DOCS; docV = iterator.nextDoc()) { + // write index vector + OptimizedScalarQuantizer.QuantizationResult r = + binaryQuantizer.scalarQuantize( + floatVectorValues.vectorValue(iterator.index()), + quantizationScratch, + encoding.getQueryBits(), + quantizedByteVectorValues.getCentroid()); + docsWithField.add(docV); + // pack and store the 4bit query vector + transposeHalfByte(quantizationScratch, toQuery); + binarizedQueryData.writeBytes(toQuery, toQuery.length); + binarizedQueryData.writeInt(Float.floatToIntBits(r.lowerInterval())); + binarizedQueryData.writeInt(Float.floatToIntBits(r.upperInterval())); + binarizedQueryData.writeInt(Float.floatToIntBits(r.additionalCorrection())); + binarizedQueryData.writeInt(r.quantizedComponentSum()); + } + return docsWithField; + } + + private record FieldEntry( + VectorSimilarityFunction similarityFunction, + VectorEncoding vectorEncoding, + int dimension, + long vectorDataOffset, + long vectorDataLength, + int size, + ScalarEncoding scalarEncoding, + float[] centroid, + float centroidDP, + OrdToDocDISIReaderConfiguration ordToDocDISIReaderConfiguration) { + + boolean isDataBlind() { + if (centroid == null) return false; + for (float v : centroid) { + if (v != 0f) return false; + } Review Comment: If the center is already a zero vector that is geometrically equivalent to our data blind configuration. Centering wouldn't subtract anything from the vector, and for angular similarity center_dot would be zero. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
