Module org.elasticsearch.simdvec

Package org.elasticsearch.simdvec.internal

Class Int4VectorScorer

java.lang.Object

org.apache.lucene.util.hnsw.RandomVectorScorer.AbstractRandomVectorScorer

org.elasticsearch.simdvec.internal.Int4VectorScorer

All Implemented Interfaces:

org.apache.lucene.util.hnsw.HasKnnVectorValues, org.apache.lucene.util.hnsw.RandomVectorScorer

public final class Int4VectorScorer
extends org.apache.lucene.util.hnsw.RandomVectorScorer.AbstractRandomVectorScorer

Int4 packed-nibble query-time scorer. The float query is quantized externally and passed in as unpacked bytes (one byte per dimension, 0-15 range) along with corrective terms. Each stored vector is dims/2 packed bytes followed by corrective terms (3 floats + 1 int).

Nested Class Summary

Nested classes/interfaces inherited from interface org.apache.lucene.util.hnsw.RandomVectorScorer

org.apache.lucene.util.hnsw.RandomVectorScorer.AbstractRandomVectorScorer

Method Summary

Modifier and Type	Method	Description
float	bulkScore(int[] ordinals, float[] scores, int numNodes)
static Optional<org.apache.lucene.util.hnsw.RandomVectorScorer>	create(org.apache.lucene.index.VectorSimilarityFunction sim, org.apache.lucene.codecs.lucene104.QuantizedByteVectorValues values, byte[] unpackedQuery, float lowerInterval, float upperInterval, float additionalCorrection, int quantizedComponentSum)	Creates an int4 query-time scorer if the input supports efficient access.
float	score(int node)

Methods inherited from class org.apache.lucene.util.hnsw.RandomVectorScorer.AbstractRandomVectorScorer

getAcceptOrds, maxOrd, ordToDoc, values

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Method Details

create

public static Optional<org.apache.lucene.util.hnsw.RandomVectorScorer> create(org.apache.lucene.index.VectorSimilarityFunction sim,
 org.apache.lucene.codecs.lucene104.QuantizedByteVectorValues values,
 byte[] unpackedQuery,
 float lowerInterval,
 float upperInterval,
 float additionalCorrection,
 int quantizedComponentSum)

Creates an int4 query-time scorer if the input supports efficient access.

Parameters:

sim - the similarity function

values - the quantized vector values

unpackedQuery - the quantized query (dims bytes, one per dimension, 0-15)

lowerInterval - query corrective term

upperInterval - query corrective term

additionalCorrection - query corrective term

quantizedComponentSum - query corrective term

Returns:

an optional scorer or empty if the input doesn't support native access

score

public float score(int node)
            throws IOException

Throws:

IOException

bulkScore

public float bulkScore(int[] ordinals,
 float[] scores,
 int numNodes)
                throws IOException

Throws:

IOException