Skip to content

Commit

Permalink
kNN vector rescoring for quantized vectors (elastic#116663)
Browse files Browse the repository at this point in the history
(cherry picked from commit 5996772)

# Conflicts:
#	server/src/main/java/org/elasticsearch/search/vectors/KnnSearchBuilder.java
#	x-pack/plugin/rank-rrf/src/main/java/org/elasticsearch/xpack/rank/rrf/RRFRankBuilder.java
  • Loading branch information
carlosdelest committed Dec 11, 2024
1 parent 3491312 commit 81837fb
Show file tree
Hide file tree
Showing 66 changed files with 2,365 additions and 296 deletions.
5 changes: 5 additions & 0 deletions docs/changelog/116663.yaml
Original file line number Diff line number Diff line change
@@ -0,0 +1,5 @@
pr: 116663
summary: KNN vector rescoring for quantized vectors
area: Vector Search
type: feature
issues: []
Original file line number Diff line number Diff line change
Expand Up @@ -1359,7 +1359,7 @@ public void testKnnQueryNotSupportedInPercolator() throws IOException {
""");
indicesAdmin().prepareCreate("index1").setMapping(mappings).get();
ensureGreen();
QueryBuilder knnVectorQueryBuilder = new KnnVectorQueryBuilder("my_vector", new float[] { 1, 1, 1, 1, 1 }, 10, 10, null);
QueryBuilder knnVectorQueryBuilder = new KnnVectorQueryBuilder("my_vector", new float[] { 1, 1, 1, 1, 1 }, 10, 10, null, null);

IndexRequestBuilder indexRequestBuilder = prepareIndex("index1").setId("knn_query1")
.setSource(jsonBuilder().startObject().field("my_query", knnVectorQueryBuilder).endObject());
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -18,7 +18,7 @@ setup:
dims: 5
index: true
index_options:
type: hnsw
type: int8_hnsw
similarity: l2_norm

- do:
Expand Down Expand Up @@ -73,3 +73,59 @@ setup:
- match: {hits.total.value: 1}
- match: {hits.hits.0._id: "3"}
- match: {hits.hits.0.fields.name.0: "rabbit.jpg"}

---
"Vector rescoring has no effect for non-quantized vectors and provides same results as non-rescored knn":
- requires:
reason: 'Quantized vector rescoring is required'
test_runner_features: [capabilities]
capabilities:
- method: GET
path: /_search
capabilities: [knn_quantized_vector_rescore]
- skip:
features: "headers"

# Rescore
- do:
headers:
Content-Type: application/json
search:
rest_total_hits_as_int: true
index: index1
body:
knn:
field: vector
query_vector: [2, 2, 2, 2, 3]
k: 3
num_candidates: 3
rescore_vector:
num_candidates_factor: 1.5

# Get rescoring scores - hit ordering may change depending on how things are distributed
- match: { hits.total: 3 }
- set: { hits.hits.0._score: rescore_score0 }
- set: { hits.hits.1._score: rescore_score1 }
- set: { hits.hits.2._score: rescore_score2 }

# Exact knn via script score
- do:
headers:
Content-Type: application/json
search:
rest_total_hits_as_int: true
index: index1
body:
query:
script_score:
query: {match_all: {} }
script:
source: "1.0 / (1.0 + Math.pow(l2norm(params.query_vector, 'vector'), 2.0))"
params:
query_vector: [2, 2, 2, 2, 3]

# Compare scores as hit IDs may change depending on how things are distributed
- match: { hits.total: 3 }
- match: { hits.hits.0._score: $rescore_score0 }
- match: { hits.hits.1._score: $rescore_score1 }
- match: { hits.hits.2._score: $rescore_score2 }
Original file line number Diff line number Diff line change
@@ -0,0 +1,137 @@
setup:
- requires:
reason: 'Quantized vector rescoring is required'
test_runner_features: [ capabilities ]
capabilities:
- method: GET
path: /_search
capabilities: [ knn_quantized_vector_rescore ]
- skip:
features: "headers"

- do:
indices.create:
index: bbq_hnsw
body:
settings:
index:
number_of_shards: 1
mappings:
properties:
vector:
type: dense_vector
dims: 64
index: true
similarity: max_inner_product
index_options:
type: bbq_hnsw

- do:
index:
index: bbq_hnsw
id: "1"
body:
vector: [0.077, 0.32 , -0.205, 0.63 , 0.032, 0.201, 0.167, -0.313,
0.176, 0.531, -0.375, 0.334, -0.046, 0.078, -0.349, 0.272,
0.307, -0.083, 0.504, 0.255, -0.404, 0.289, -0.226, -0.132,
-0.216, 0.49 , 0.039, 0.507, -0.307, 0.107, 0.09 , -0.265,
-0.285, 0.336, -0.272, 0.369, -0.282, 0.086, -0.132, 0.475,
-0.224, 0.203, 0.439, 0.064, 0.246, -0.396, 0.297, 0.242,
-0.028, 0.321, -0.022, -0.009, -0.001 , 0.031, -0.533, 0.45,
-0.683, 1.331, 0.194, -0.157, -0.1 , -0.279, -0.098, -0.176]
# Flush in order to provoke a merge later
- do:
indices.flush:
index: bbq_hnsw

- do:
index:
index: bbq_hnsw
id: "2"
body:
vector: [0.196, 0.514, 0.039, 0.555, -0.042, 0.242, 0.463, -0.348,
-0.08 , 0.442, -0.067, -0.05 , -0.001, 0.298, -0.377, 0.048,
0.307, 0.159, 0.278, 0.119, -0.057, 0.333, -0.289, -0.438,
-0.014, 0.361, -0.169, 0.292, -0.229, 0.123, 0.031, -0.138,
-0.139, 0.315, -0.216, 0.322, -0.445, -0.059, 0.071, 0.429,
-0.602, -0.142, 0.11 , 0.192, 0.259, -0.241, 0.181, -0.166,
0.082, 0.107, -0.05 , 0.155, 0.011, 0.161, -0.486, 0.569,
-0.489, 0.901, 0.208, 0.011, -0.209, -0.153, -0.27 , -0.013]
# Flush in order to provoke a merge later
- do:
indices.flush:
index: bbq_hnsw

- do:
index:
index: bbq_hnsw
id: "3"
body:
name: rabbit.jpg
vector: [0.139, 0.178, -0.117, 0.399, 0.014, -0.139, 0.347, -0.33 ,
0.139, 0.34 , -0.052, -0.052, -0.249, 0.327, -0.288, 0.049,
0.464, 0.338, 0.516, 0.247, -0.104, 0.259, -0.209, -0.246,
-0.11 , 0.323, 0.091, 0.442, -0.254, 0.195, -0.109, -0.058,
-0.279, 0.402, -0.107, 0.308, -0.273, 0.019, 0.082, 0.399,
-0.658, -0.03 , 0.276, 0.041, 0.187, -0.331, 0.165, 0.017,
0.171, -0.203, -0.198, 0.115, -0.007, 0.337, -0.444, 0.615,
-0.657, 1.285, 0.2 , -0.062, 0.038, 0.089, -0.068, -0.058]
# Flush in order to provoke a merge later
- do:
indices.flush:
index: bbq_hnsw

- do:
indices.forcemerge:
index: bbq_hnsw
max_num_segments: 1
---
"Profile rescored knn search":

- do:
search:
index: bbq_hnsw
body:
profile: true
knn:
field: vector
query_vector: [0.128, 0.067, -0.08 , 0.395, -0.11 , -0.259, 0.473, -0.393,
0.292, 0.571, -0.491, 0.444, -0.288, 0.198, -0.343, 0.015,
0.232, 0.088, 0.228, 0.151, -0.136, 0.236, -0.273, -0.259,
-0.217, 0.359, -0.207, 0.352, -0.142, 0.192, -0.061, -0.17 ,
-0.343, 0.189, -0.221, 0.32 , -0.301, -0.1 , 0.005, 0.232,
-0.344, 0.136, 0.252, 0.157, -0.13 , -0.244, 0.193, -0.034,
-0.12 , -0.193, -0.102, 0.252, -0.185, -0.167, -0.575, 0.582,
-0.426, 0.983, 0.212, 0.204, 0.03 , -0.276, -0.425, -0.158]
k: 3
num_candidates: 3
"rescore_vector":
"num_candidates_factor": 2.0

# We expect the knn search ops + rescoring num_cnaidates (for rescoring) per shard
- match: { profile.shards.0.dfs.knn.0.vector_operations_count: 6 }

# Search with similarity to check number of operations are propagated correctly
- do:
search:
index: bbq_hnsw
body:
profile: true
knn:
field: vector
query_vector: [0.128, 0.067, -0.08 , 0.395, -0.11 , -0.259, 0.473, -0.393,
0.292, 0.571, -0.491, 0.444, -0.288, 0.198, -0.343, 0.015,
0.232, 0.088, 0.228, 0.151, -0.136, 0.236, -0.273, -0.259,
-0.217, 0.359, -0.207, 0.352, -0.142, 0.192, -0.061, -0.17 ,
-0.343, 0.189, -0.221, 0.32 , -0.301, -0.1 , 0.005, 0.232,
-0.344, 0.136, 0.252, 0.157, -0.13 , -0.244, 0.193, -0.034,
-0.12 , -0.193, -0.102, 0.252, -0.185, -0.167, -0.575, 0.582,
-0.426, 0.983, 0.212, 0.204, 0.03 , -0.276, -0.425, -0.158]
k: 3
num_candidates: 3
similarity: 100000
"rescore_vector":
"num_candidates_factor": 2.0

# We expect the knn search ops + rescoring num_cnaidates (for rescoring) per shard
- match: { profile.shards.0.dfs.knn.0.vector_operations_count: 6 }
Original file line number Diff line number Diff line change
Expand Up @@ -541,3 +541,58 @@ setup:
num_candidates: 3

- match: { hits.total.value: 0 }
---
"Vector rescoring has no effect for non-quantized vectors and provides same results as non-rescored knn":
- requires:
reason: 'Quantized vector rescoring is required'
test_runner_features: [capabilities]
capabilities:
- method: GET
path: /_search
capabilities: [knn_quantized_vector_rescore]
- skip:
features: "headers"

# Non-rescored knn
- do:
headers:
Content-Type: application/json
search:
rest_total_hits_as_int: true
index: test
body:
fields: [ "name" ]
knn:
field: vector
query_vector: [-0.5, 90.0, -10, 14.8, -156.0]
k: 3
num_candidates: 3

# Get scores - hit ordering may change depending on how things are distributed
- match: { hits.total: 3 }
- set: { hits.hits.0._score: knn_score0 }
- set: { hits.hits.1._score: knn_score1 }
- set: { hits.hits.2._score: knn_score2 }

# Rescored knn
- do:
headers:
Content-Type: application/json
search:
rest_total_hits_as_int: true
index: test
body:
fields: [ "name" ]
knn:
field: vector
query_vector: [-0.5, 90.0, -10, 14.8, -156.0]
k: 3
num_candidates: 3
rescore_vector:
num_candidates_factor: 1.5

# Compare scores as hit IDs may change depending on how things are distributed
- match: { hits.total: 3 }
- match: { hits.hits.0._score: $knn_score0 }
- match: { hits.hits.1._score: $knn_score1 }
- match: { hits.hits.2._score: $knn_score2 }
Original file line number Diff line number Diff line change
Expand Up @@ -108,6 +108,75 @@ setup:
- match: { hits.hits.1._id: "3" }
- match: { hits.hits.2._id: "2" }
---
"Vector rescoring has same scoring as exact search for kNN section":
- requires:
reason: 'Quantized vector rescoring is required'
test_runner_features: [capabilities]
capabilities:
- method: GET
path: /_search
capabilities: [knn_quantized_vector_rescore]
- skip:
features: "headers"

# Rescore
- do:
headers:
Content-Type: application/json
search:
rest_total_hits_as_int: true
index: bbq_hnsw
body:
knn:
field: vector
query_vector: [0.128, 0.067, -0.08 , 0.395, -0.11 , -0.259, 0.473, -0.393,
0.292, 0.571, -0.491, 0.444, -0.288, 0.198, -0.343, 0.015,
0.232, 0.088, 0.228, 0.151, -0.136, 0.236, -0.273, -0.259,
-0.217, 0.359, -0.207, 0.352, -0.142, 0.192, -0.061, -0.17 ,
-0.343, 0.189, -0.221, 0.32 , -0.301, -0.1 , 0.005, 0.232,
-0.344, 0.136, 0.252, 0.157, -0.13 , -0.244, 0.193, -0.034,
-0.12 , -0.193, -0.102, 0.252, -0.185, -0.167, -0.575, 0.582,
-0.426, 0.983, 0.212, 0.204, 0.03 , -0.276, -0.425, -0.158]
k: 3
num_candidates: 3
rescore_vector:
num_candidates_factor: 1.5

# Get rescoring scores - hit ordering may change depending on how things are distributed
- match: { hits.total: 3 }
- set: { hits.hits.0._score: rescore_score0 }
- set: { hits.hits.1._score: rescore_score1 }
- set: { hits.hits.2._score: rescore_score2 }

# Exact knn via script score
- do:
headers:
Content-Type: application/json
search:
rest_total_hits_as_int: true
body:
query:
script_score:
query: {match_all: {} }
script:
source: "double similarity = dotProduct(params.query_vector, 'vector'); return similarity < 0 ? 1 / (1 + -1 * similarity) : similarity + 1"
params:
query_vector: [0.128, 0.067, -0.08 , 0.395, -0.11 , -0.259, 0.473, -0.393,
0.292, 0.571, -0.491, 0.444, -0.288, 0.198, -0.343, 0.015,
0.232, 0.088, 0.228, 0.151, -0.136, 0.236, -0.273, -0.259,
-0.217, 0.359, -0.207, 0.352, -0.142, 0.192, -0.061, -0.17 ,
-0.343, 0.189, -0.221, 0.32 , -0.301, -0.1 , 0.005, 0.232,
-0.344, 0.136, 0.252, 0.157, -0.13 , -0.244, 0.193, -0.034,
-0.12 , -0.193, -0.102, 0.252, -0.185, -0.167, -0.575, 0.582,
-0.426, 0.983, 0.212, 0.204, 0.03 , -0.276, -0.425, -0.158]

# Compare scores as hit IDs may change depending on how things are distributed
- match: { hits.total: 3 }
- match: { hits.hits.0._score: $rescore_score0 }
- match: { hits.hits.1._score: $rescore_score1 }
- match: { hits.hits.2._score: $rescore_score2 }

---
"Test bad quantization parameters":
- do:
catch: bad_request
Expand Down
Loading

0 comments on commit 81837fb

Please sign in to comment.