benwtrent commented on code in PR #12253: URL: https://github.com/apache/lucene/pull/12253#discussion_r1226688194
########## lucene/queries/src/java/org/apache/lucene/queries/function/valuesource/KnnVectorFieldSource.java: ########## @@ -0,0 +1,112 @@ +/* + * 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.queries.function.valuesource; + +import java.io.IOException; +import java.util.Map; +import org.apache.lucene.index.ByteVectorValues; +import org.apache.lucene.index.FloatVectorValues; +import org.apache.lucene.index.LeafReaderContext; +import org.apache.lucene.queries.function.FunctionValues; +import org.apache.lucene.queries.function.ValueSource; +import org.apache.lucene.search.DocIdSetIterator; + +/** An implementation for retrieving {@link FunctionValues} instances for knn vectors fields. */ +public class KnnVectorFieldSource extends ValueSource { + private final String fieldName; + + public KnnVectorFieldSource(String fieldName) { + this.fieldName = fieldName; + } + + @Override + public FunctionValues getValues(Map<Object, Object> context, LeafReaderContext readerContext) + throws IOException { + + final FloatVectorValues floatValues = readerContext.reader().getFloatVectorValues(fieldName); + final ByteVectorValues byteValues = readerContext.reader().getByteVectorValues(fieldName); + + if (floatValues == null && byteValues == null) { + throw new IllegalArgumentException( + "no vector value is indexed for field '" + fieldName + "'"); + } + return new FunctionValues() { + int lastDocID; + + @Override + public float[] floatVectorVal(int doc) throws IOException { + float[] floatVector = null; + if (floatValues != null && exists(floatValues, doc)) { + floatVector = floatValues.vectorValue(); + } else { + byte[] byteVector = byteVectorVal(doc); + if (byteVector != null) { + floatVector = new float[byteVector.length]; + for (int i = 0; i < byteVector.length; i++) { + floatVector[i] = byteVector[i]; + } + } + } + return floatVector; + } Review Comment: > Code duplication: Having separate implementations resulted in duplicate code for byte and float vectors. This is a non-issue. A consistent API is way better than implementing "DRY". > Difficulty in using float[] and byte[] vectors interchangeably: They should not be used interchangeably. `float[]` should return `float[]`, `byte[]` should return `byte[]`. Users stored them that way, that is the way they should be retrieved. > The knn vectors, whether byte or float, essentially represented the same field This is making a big assumption on the user. I don't understand this jump in logic. They are not the same field nor the same field type. > The difference lay in the underlying encoding and the precision used to store the vectors This is not true. `byte[]` means you stored them as `int8` values. They COULD be quantized `float[]` values, or the user just has `byte[]` values. Linear quantization could be construed as a "precision" thing, but that is simplifying what quantization is, and how its implemented. As an argument against "just precision" users should NEVER attempt to store something like `float[]{1.22348f, 0.2384f}` as a `byte` encoded vector without some quantization process in between creating the vector and storing them. Simply truncating is bad, will cause errors, and is a surprising result. > The decision to use the byte vector format over the float vector format was primarily driven by non-functional aspects such as space occupancy and performance. This is partially true. Space and performance are important, but casting between `float` and `byte` breaks an otherwise strong contract with the API calls. Additionally, I would say that generally, there are other valid reasons for `int8` storage other than "Hey I quantized the output of some floats". What if the user just wanted to store `int8` because it worked best for their data? > If there are compelling reasons for not proceeding with this approach, we could consider reverting to the point where the two implementations were separate. I would argue the onus is on the PR's authors to give compelling reasons for rejecting the clearly specified APIs of `byte` and `float` vectors. Not on reviewers to justify consistency with the rest of Lucene. -- 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: issues-unsubscr...@lucene.apache.org For queries about this service, please contact Infrastructure at: us...@infra.apache.org --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@lucene.apache.org For additional commands, e-mail: issues-h...@lucene.apache.org
