Skip to content

Commit

Permalink
Revert "Auto-normalize dot_product vectors at index & query (#98944)" (…
Browse files Browse the repository at this point in the history 
…#99421)

This reverts commit 7b9c367.
  • Loading branch information
@benwtrent
benwtrent committed Sep 11, 2023
1 parent 79bae69 commit 83b70e3
Show file tree
Hide file tree
Showing 9 changed files with 54 additions and 223 deletions.
5 changes: 0 additions & 5 deletions docs/changelog/98944.yaml
View file

This file was deleted.

4 changes: 2 additions & 2 deletions docs/reference/how-to/knn-search.asciidoc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -21,8 +21,8 @@ options.
The `cosine` option accepts any float vector and computes the cosine
similarity. While this is convenient for testing, it's not the most efficient
approach. Instead, we recommend using the `dot_product` option to compute the
similarity. When using `dot_product`, all vectors are normalized during index to have
a magnitude of 1. The `dot_product` option is significantly faster, since it
similarity. To use `dot_product`, all vectors need to be normalized in advance
to have length 1. The `dot_product` option is significantly faster, since it
avoids performing extra vector length computations during the search.

[discrete]
Expand Down
10 changes: 5 additions & 5 deletions docs/reference/mapping/types/dense-vector.asciidoc
View file
Original file line number Diff line number Diff line change
Expand Up @@ -163,9 +163,9 @@ Computes the dot product of two vectors. This option provides an optimized way
to perform cosine similarity. The constraints and computed score are defined
by `element_type`.
+
When `element_type` is `float`, all vectors are automatically converted to unit length, including both
document and query vectors. Consequently, `dot_product` does not allow vectors with a zero magnitude.
The document `_score` is computed as `(1 + dot_product(query, vector)) / 2`.
When `element_type` is `float`, all vectors must be unit length, including both
document and query vectors. The document `_score` is computed as
`(1 + dot_product(query, vector)) / 2`.
+
When `element_type` is `byte`, all vectors must have the same
length including both document and query vectors or results will be inaccurate.
Expand All @@ -175,9 +175,9 @@ where `dims` is the number of dimensions per vector.
`cosine`:::
Computes the cosine similarity. Note that the most efficient way to perform
cosine similarity is to have all vectors normalized to unit length, and instead use
cosine similarity is to normalize all vectors to unit length, and instead use
`dot_product`. You should only use `cosine` if you need to preserve the
original vectors and cannot allow Elasticsearch to normalize them. The document `_score`
original vectors and cannot normalize them in advance. The document `_score`
is computed as `(1 + cosine(query, vector)) / 2`. The `cosine` similarity does
not allow vectors with zero magnitude, since cosine is not defined in this
case.
Expand Down
120 changes: 0 additions & 120 deletions rest-api-spec/src/yamlRestTest/resources/rest-api-spec/test/search.vectors/40_knn_search.yml
View file
Original file line number Diff line number Diff line change
Expand Up @@ -368,123 +368,3 @@ setup:
filter: {"term": {"name": "cow.jpg"}}

- length: {hits.hits: 0}
---
"kNN search with dot-product auto-normalized":
- skip:
features: close_to
version: ' - 8.10.99'
reason: 'dense_vector auto-normalized dot_product in 8.11'

- do:
indices.create:
index: test_dot_product
body:
mappings:
properties:
name:
type: keyword
dot_product_vector:
type: dense_vector
dims: 5
index: true
similarity: dot_product
cosine_vector:
type: dense_vector
dims: 5
index: true
similarity: cosine

- do:
index:
index: test_dot_product
id: "1"
body:
name: cow.jpg
dot_product_vector: [ 230.0, 300.33, -34.8988, 15.555, -200.0 ]
cosine_vector: [ 230.0, 300.33, -34.8988, 15.555, -200.0 ]

- do:
index:
index: test_dot_product
id: "2"
body:
name: moose.jpg
dot_product_vector: [ -0.5, 100.0, -13, 14.8, -156.0 ]
cosine_vector: [ -0.5, 100.0, -13, 14.8, -156.0 ]

- do:
index:
index: test_dot_product
id: "3"
body:
name: rabbit.jpg
dot_product_vector: [ 0.5, 111.3, -13.0, 14.8, -156.0 ]
cosine_vector: [ 0.5, 111.3, -13.0, 14.8, -156.0 ]

- do:
indices.refresh: { }

- do:
search:
index: test_dot_product
body:
fields: [ "name" ]
knn:
field: dot_product_vector
query_vector: [-0.5, 90.0, -10, 14.8, -156.0]
k: 2
num_candidates: 3

- match: {hits.total.value: 2}
- match: {hits.hits.0._id: "2"}
- set: { hits.hits.0._score: score_0 }
- match: {hits.hits.0.fields.name.0: "moose.jpg"}
- match: {hits.hits.1._id: "3"}
- set: { hits.hits.1._score: score_1 }
- match: {hits.hits.1.fields.name.0: "rabbit.jpg"}

- do:
search:
index: test_dot_product
body:
fields: [ "name" ]
knn:
field: cosine_vector
query_vector: [ -0.5, 90.0, -10, 14.8, -156.0 ]
k: 2
num_candidates: 3

- match: {hits.total.value: 2}
- match: {hits.hits.0._id: "2"}
- close_to: { hits.hits.0._score: { value: $score_0, error: 0.00001 } }
- match: {hits.hits.0.fields.name.0: "moose.jpg"}
- match: {hits.hits.1._id: "3"}
- close_to: { hits.hits.1._score: { value: $score_1, error: 0.00001 } }
- match: {hits.hits.1.fields.name.0: "rabbit.jpg"}
---
"kNN search fails with non-normalized dot-product in older versions":
- skip:
version: '8.10.99 - '
reason: 'dense_vector auto-normalized dot_product in 8.11'

- do:
indices.create:
index: test_failing_dot_product
body:
mappings:
properties:
dot_product_vector:
type: dense_vector
dims: 5
index: true
similarity: dot_product

- do:
catch: bad_request
index:
index: test_failing_dot_product
id: "1"
body:
name: cow.jpg
dot_product_vector: [ 230.0, 300.33, -34.8988, 15.555, -200.0 ]

98 changes: 19 additions & 79 deletions server/src/main/java/org/elasticsearch/index/mapper/vectors/DenseVectorFieldMapper.java
View file
Original file line number Diff line number Diff line change
Expand Up @@ -26,7 +26,6 @@
import org.apache.lucene.search.KnnFloatVectorQuery;
import org.apache.lucene.search.Query;
import org.apache.lucene.util.BytesRef;
import org.apache.lucene.util.VectorUtil;
import org.elasticsearch.common.xcontent.support.XContentMapValues;
import org.elasticsearch.index.IndexVersion;
import org.elasticsearch.index.fielddata.FieldDataContext;
Expand Down Expand Up @@ -57,7 +56,6 @@
import java.nio.ByteBuffer;
import java.nio.ByteOrder;
import java.time.ZoneId;
import java.util.Arrays;
import java.util.Locale;
import java.util.Map;
import java.util.Objects;
Expand All @@ -71,10 +69,8 @@
* A {@link FieldMapper} for indexing a dense vector of floats.
*/
public class DenseVectorFieldMapper extends FieldMapper {
private static final float EPS = 1e-4f;
public static final IndexVersion MAGNITUDE_STORED_INDEX_VERSION = IndexVersion.V_7_5_0;
public static final IndexVersion INDEXED_BY_DEFAULT_INDEX_VERSION = IndexVersion.V_8_11_0;
public static final IndexVersion DOT_PRODUCT_AUTO_NORMALIZED = IndexVersion.V_8_11_0;
public static final IndexVersion LITTLE_ENDIAN_FLOAT_STORED_INDEX_VERSION = IndexVersion.V_8_9_0;

public static final String CONTENT_TYPE = "dense_vector";
Expand Down Expand Up @@ -325,7 +321,6 @@ public void checkVectorBounds(float[] vector) {

@Override
void checkVectorMagnitude(
IndexVersion indexVersion,
VectorSimilarity similarity,
Function<StringBuilder, StringBuilder> appender,
float squaredMagnitude
Expand Down Expand Up @@ -388,12 +383,7 @@ public Field parseKnnVector(DocumentParserContext context, DenseVectorFieldMappe
squaredMagnitude += value * value;
}
fieldMapper.checkDimensionMatches(index, context);
checkVectorMagnitude(
fieldMapper.indexCreatedVersion,
fieldMapper.similarity,
errorByteElementsAppender(vector),
squaredMagnitude
);
checkVectorMagnitude(fieldMapper.similarity, errorByteElementsAppender(vector), squaredMagnitude);
return createKnnVectorField(fieldMapper.fieldType().name(), vector, fieldMapper.similarity.function);
}

Expand Down Expand Up @@ -485,31 +475,20 @@ public void checkVectorBounds(float[] vector) {

@Override
void checkVectorMagnitude(
IndexVersion indexVersion,
VectorSimilarity similarity,
Function<StringBuilder, StringBuilder> appender,
float squaredMagnitude
) {
StringBuilder errorBuilder = null;

if (indexVersion.before(DOT_PRODUCT_AUTO_NORMALIZED)) {
if (similarity == VectorSimilarity.DOT_PRODUCT && Math.abs(squaredMagnitude - 1.0f) > EPS) {
errorBuilder = new StringBuilder(
"The [" + VectorSimilarity.DOT_PRODUCT + "] similarity can only be used with unit-length vectors."
);
}
if (similarity == VectorSimilarity.COSINE && Math.sqrt(squaredMagnitude) == 0.0f) {
errorBuilder = new StringBuilder(
"The [" + similarity + "] similarity does not support vectors with zero magnitude."
);
}
} else {
if ((similarity == VectorSimilarity.COSINE || similarity == VectorSimilarity.DOT_PRODUCT)
&& Math.sqrt(squaredMagnitude) == 0.0f) {
errorBuilder = new StringBuilder(
"The [" + similarity + "] similarity does not support vectors with zero magnitude."
);
}
if (similarity == VectorSimilarity.DOT_PRODUCT && Math.abs(squaredMagnitude - 1.0f) > 1e-4f) {
errorBuilder = new StringBuilder(
"The [" + VectorSimilarity.DOT_PRODUCT + "] similarity can only be used with unit-length vectors."
);
} else if (similarity == VectorSimilarity.COSINE && Math.sqrt(squaredMagnitude) == 0.0f) {
errorBuilder = new StringBuilder(
"The [" + VectorSimilarity.COSINE + "] similarity does not support vectors with zero magnitude."
);
}

if (errorBuilder != null) {
Expand All @@ -532,15 +511,7 @@ public Field parseKnnVector(DocumentParserContext context, DenseVectorFieldMappe
}
fieldMapper.checkDimensionMatches(index, context);
checkVectorBounds(vector);
checkVectorMagnitude(
fieldMapper.indexCreatedVersion,
fieldMapper.similarity,
errorFloatElementsAppender(vector),
squaredMagnitude
);
if (fieldMapper.indexCreatedVersion.onOrAfter(DOT_PRODUCT_AUTO_NORMALIZED)) {
fieldMapper.similarity.floatPreprocessing(vector, squaredMagnitude);
}
checkVectorMagnitude(fieldMapper.similarity, errorFloatElementsAppender(vector), squaredMagnitude);
return createKnnVectorField(fieldMapper.fieldType().name(), vector, fieldMapper.similarity.function);
}

Expand Down Expand Up @@ -598,7 +569,6 @@ abstract double parseKnnVectorToByteBuffer(DocumentParserContext context, DenseV
public abstract void checkVectorBounds(float[] vector);

abstract void checkVectorMagnitude(
IndexVersion indexVersion,
VectorSimilarity similarity,
Function<StringBuilder, StringBuilder> errorElementsAppender,
float squaredMagnitude
Expand Down Expand Up @@ -717,21 +687,6 @@ float score(float similarity, ElementType elementType, int dim) {
case FLOAT -> (1 + similarity) / 2f;
};
}

@Override
void floatPreprocessing(float[] vector, float squareSum) {
if (squareSum == 0) {
throw new IllegalArgumentException("Cannot normalize a zero-length vector");
}
// Vector already has a magnitude have `1`
if (Math.abs(squareSum - 1.0f) < EPS) {
return;
}
float length = (float) Math.sqrt(squareSum);
for (int i = 0; i < vector.length; i++) {
vector[i] /= length;
}
}
};

public final VectorSimilarityFunction function;
Expand All @@ -746,8 +701,6 @@ public final String toString() {
}

abstract float score(float similarity, ElementType elementType, int dim);

void floatPreprocessing(float[] vector, float squareSum) {}
}

private abstract static class IndexOptions implements ToXContent {
Expand Down Expand Up @@ -906,13 +859,11 @@ public Query createKnnQuery(byte[] queryVector, int numCands, Query filter, Floa
}

if (similarity == VectorSimilarity.DOT_PRODUCT || similarity == VectorSimilarity.COSINE) {
int squaredMagnitude = VectorUtil.dotProduct(queryVector, queryVector);
elementType.checkVectorMagnitude(
indexVersionCreated,
similarity,
elementType.errorByteElementsAppender(queryVector),
squaredMagnitude
);
float squaredMagnitude = 0.0f;
for (byte b : queryVector) {
squaredMagnitude += b * b;
}
elementType.checkVectorMagnitude(similarity, elementType.errorByteElementsAppender(queryVector), squaredMagnitude);
}
Query knnQuery = new KnnByteVectorQuery(name(), queryVector, numCands, filter);
if (similarityThreshold != null) {
Expand Down Expand Up @@ -940,22 +891,11 @@ public Query createKnnQuery(float[] queryVector, int numCands, Query filter, Flo
elementType.checkVectorBounds(queryVector);

if (similarity == VectorSimilarity.DOT_PRODUCT || similarity == VectorSimilarity.COSINE) {
float squaredMagnitude = VectorUtil.dotProduct(queryVector, queryVector);
elementType.checkVectorMagnitude(
indexVersionCreated,
similarity,
elementType.errorFloatElementsAppender(queryVector),
squaredMagnitude
);
// We don't want to normalize the original query vector.
// It mutates it in place and might cause down stream weirdness
// Instead we copy the value and then normalize that copy
if (similarity == VectorSimilarity.DOT_PRODUCT
&& elementType == ElementType.FLOAT
&& indexVersionCreated.onOrAfter(DOT_PRODUCT_AUTO_NORMALIZED)) {
queryVector = Arrays.copyOf(queryVector, queryVector.length);
similarity.floatPreprocessing(queryVector, squaredMagnitude);
float squaredMagnitude = 0.0f;
for (float e : queryVector) {
squaredMagnitude += e * e;
}
elementType.checkVectorMagnitude(similarity, elementType.errorFloatElementsAppender(queryVector), squaredMagnitude);
}
Query knnQuery = switch (elementType) {
case BYTE -> {
Expand Down
22 changes: 19 additions & 3 deletions server/src/test/java/org/elasticsearch/index/mapper/vectors/DenseVectorFieldMapperTests.java
View file
Original file line number Diff line number Diff line change
Expand Up @@ -320,7 +320,7 @@ public void testDotProductWithInvalidNorm() throws Exception {
b -> b.field("type", "dense_vector").field("dims", 3).field("index", true).field("similarity", VectorSimilarity.DOT_PRODUCT)
)
);
float[] vector = { 0f, 0f, 0f };
float[] vector = { -12.1f, 2.7f, -4 };
DocumentParsingException e = expectThrows(
DocumentParsingException.class,
() -> mapper.parse(source(b -> b.array("field", vector)))
Expand All @@ -329,7 +329,23 @@ public void testDotProductWithInvalidNorm() throws Exception {
assertThat(
e.getCause().getMessage(),
containsString(
"The [dot_product] similarity does not support vectors with zero magnitude. Preview of invalid vector: [0.0, 0.0, 0.0]"
"The [dot_product] similarity can only be used with unit-length vectors. Preview of invalid vector: [-12.1, 2.7, -4.0]"
)
);

DocumentMapper mapperWithLargerDim = createDocumentMapper(
fieldMapping(
b -> b.field("type", "dense_vector").field("dims", 6).field("index", true).field("similarity", VectorSimilarity.DOT_PRODUCT)
)
);
float[] largerVector = { -12.1f, 2.7f, -4, 1.05f, 10.0f, 29.9f };
e = expectThrows(DocumentParsingException.class, () -> mapperWithLargerDim.parse(source(b -> b.array("field", largerVector))));
assertNotNull(e.getCause());
assertThat(
e.getCause().getMessage(),
containsString(
"The [dot_product] similarity can only be used with unit-length vectors. "
+ "Preview of invalid vector: [-12.1, 2.7, -4.0, 1.05, 10.0, ...]"
)
);
}
Expand Down Expand Up @@ -499,7 +515,7 @@ public void testDefaultParamsBeforeIndexByDefault() throws Exception {
assertNull(denseVectorFieldType.getSimilarity());
}

public void testParamsBeforeIndexByDefault() throws Exception {
public void testtParamsBeforeIndexByDefault() throws Exception {
DocumentMapper documentMapper = createDocumentMapper(INDEXED_BY_DEFAULT_PREVIOUS_INDEX_VERSION, fieldMapping(b -> {
b.field("type", "dense_vector").field("dims", 3).field("index", true).field("similarity", "dot_product");
}));
Expand Down
Loading

0 comments on commit 83b70e3

Please sign in to comment.