1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
|
# The MIT License (MIT)
# Copyright (c) Microsoft Corporation. All rights reserved.
import os
import unittest
import uuid
import pytest
import azure.cosmos.cosmos_client as cosmos_client
import azure.cosmos.exceptions as exceptions
import test_config
import vector_test_data
from azure.cosmos import http_constants
from azure.cosmos.partition_key import PartitionKey
def verify_ordering(item_list, distance_function):
for i in range(len(item_list)):
assert item_list[i]["text"] == vector_test_data.get_ordered_item_texts()[i]
if distance_function == "euclidean":
for i in range(len(item_list) - 1):
assert item_list[i]["SimilarityScore"] <= item_list[i + 1]["SimilarityScore"]
else:
for i in range(len(item_list) - 1):
assert item_list[i]["SimilarityScore"] >= item_list[i + 1]["SimilarityScore"]
@pytest.mark.cosmosSearchQuery
class TestVectorSimilarityQuery(unittest.TestCase):
"""Test to check vector similarity queries behavior."""
client: cosmos_client.CosmosClient = None
config = test_config.TestConfig
host = config.host
masterKey = config.masterKey
connectionPolicy = config.connectionPolicy
TEST_DATABASE_ID = config.TEST_DATABASE_ID
TEST_CONTAINER_ID = "Vector Similarity Container " + str(uuid.uuid4())
@classmethod
def setUpClass(cls):
if (cls.masterKey == '[YOUR_KEY_HERE]' or
cls.host == '[YOUR_ENDPOINT_HERE]'):
raise Exception(
"You must specify your Azure Cosmos account values for "
"'masterKey' and 'host' at the top of this class to run the "
"tests.")
cls.client = cosmos_client.CosmosClient(cls.host, cls.masterKey)
cls.test_db = cls.client.create_database(str(uuid.uuid4()))
cls.created_quantized_cosine_container = cls.test_db.create_container(
id="quantized" + cls.TEST_CONTAINER_ID,
partition_key=PartitionKey(path="/pk"),
offer_throughput=test_config.TestConfig.THROUGHPUT_FOR_2_PARTITIONS,
indexing_policy=test_config.get_vector_indexing_policy(embedding_type="quantizedFlat"),
vector_embedding_policy=test_config.get_vector_embedding_policy(data_type="float32",
distance_function="cosine",
dimensions=128))
cls.created_flat_euclidean_container = cls.test_db.create_container(
id="flat" + cls.TEST_CONTAINER_ID,
partition_key=PartitionKey(path="/pk"),
offer_throughput=test_config.TestConfig.THROUGHPUT_FOR_2_PARTITIONS,
indexing_policy=test_config.get_vector_indexing_policy(embedding_type="flat"),
vector_embedding_policy=test_config.get_vector_embedding_policy(data_type="float32",
distance_function="euclidean",
dimensions=128))
cls.created_diskANN_dotproduct_container = cls.test_db.create_container(
id="diskANN" + cls.TEST_CONTAINER_ID,
partition_key=PartitionKey(path="/pk"),
offer_throughput=test_config.TestConfig.THROUGHPUT_FOR_2_PARTITIONS,
indexing_policy=test_config.get_vector_indexing_policy(embedding_type="diskANN"),
vector_embedding_policy=test_config.get_vector_embedding_policy(data_type="float32",
distance_function="dotproduct",
dimensions=128))
cls.created_large_container = cls.test_db.create_container(
id="large_container" + cls.TEST_CONTAINER_ID,
partition_key=PartitionKey(path="/pk"),
offer_throughput=test_config.TestConfig.THROUGHPUT_FOR_2_PARTITIONS,
indexing_policy=test_config.get_vector_indexing_policy(embedding_type="quantizedFlat"),
vector_embedding_policy=test_config.get_vector_embedding_policy(data_type="float32",
distance_function="cosine",
dimensions=2))
for item in vector_test_data.get_vector_items():
cls.created_quantized_cosine_container.create_item(item)
cls.created_flat_euclidean_container.create_item(item)
cls.created_diskANN_dotproduct_container.create_item(item)
@classmethod
def tearDownClass(cls):
try:
cls.test_db.delete_container("quantized" + cls.TEST_CONTAINER_ID)
cls.test_db.delete_container("flat" + cls.TEST_CONTAINER_ID)
cls.test_db.delete_container("diskANN" + cls.TEST_CONTAINER_ID)
cls.test_db.delete_container("large_container" + cls.TEST_CONTAINER_ID)
cls.client.delete_database(cls.test_db.id)
except exceptions.CosmosHttpResponseError:
pass
def test_wrong_vector_search_queries(self):
vector_string = vector_test_data.get_embedding_string("I am having a wonderful day.")
# try to send a vector search query without limit filters
query = "SELECT c.text, VectorDistance(c.embedding, [{}]) AS " \
"SimilarityScore FROM c ORDER BY VectorDistance(c.embedding, [{}])".format(vector_string, vector_string)
try:
list(self.created_large_container.query_items(query=query, enable_cross_partition_query=True))
pytest.fail("Client should not allow queries without filters.")
except ValueError as e:
assert "Executing a vector search query without TOP or LIMIT can consume many RUs very fast and" \
" have long runtimes. Please ensure you are using one of the two filters with your" \
" vector search query." in e.args[0]
# try to send a vector search query specifying the ordering as ASC or DESC
query = "SELECT c.text, VectorDistance(c.embedding, [{}]) AS " \
"SimilarityScore FROM c ORDER BY VectorDistance(c.embedding, [{}]) ASC".format(vector_string,
vector_string)
try:
list(self.created_large_container.query_items(query=query, enable_cross_partition_query=True))
pytest.fail("Client should not allow queries with ASC/DESC.")
except exceptions.CosmosHttpResponseError as e:
assert e.status_code == http_constants.StatusCodes.BAD_REQUEST
# TODO: Seems like this error message differs depending on Ubuntu vs. Windows runs?
assert ("One of the input values is invalid." in e.message
or "Specifying a sorting order (ASC or DESC) with VectorDistance function is not supported." in e.message)
def test_vector_search_environment_variables(self):
vector_string = vector_test_data.get_embedding_string("I am having a wonderful day.")
query = "SELECT TOP 10 c.text, VectorDistance(c.embedding, [{}]) AS " \
"SimilarityScore FROM c ORDER BY VectorDistance(c.embedding, [{}])".format(vector_string, vector_string)
os.environ["AZURE_COSMOS_MAX_ITEM_BUFFER_VECTOR_SEARCH"] = "1"
try:
[item for item in self.created_large_container.query_items(query=query, enable_cross_partition_query=True)]
pytest.fail("Config was not set correctly.")
except ValueError as e:
assert e.args[0] == ("Executing a vector search query with more items than the max is not allowed. "
"Please ensure you are using a limit smaller than the max, or change the max.")
os.environ["AZURE_COSMOS_MAX_ITEM_BUFFER_VECTOR_SEARCH"] = "50000"
os.environ["AZURE_COSMOS_DISABLE_NON_STREAMING_ORDER_BY"] = "False"
[item for item in self.created_large_container.query_items(query=query, enable_cross_partition_query=True)]
def test_ordering_distances(self):
# Besides ordering distances, we also verify that the query text properly replaces any set embedding policies
# load up previously calculated embedding for the given string
vector_string = vector_test_data.get_embedding_string("I am having a wonderful day.")
# test euclidean distance
for i in range(1, 11):
# we define queries with and without specs to directly use the embeddings in our container policies
vanilla_query = "SELECT TOP {} c.text, VectorDistance(c.embedding, [{}]) AS " \
"SimilarityScore FROM c ORDER BY VectorDistance(c.embedding, [{}])".format(str(i),
vector_string,
vector_string)
specs_query = "SELECT TOP {} c.text, VectorDistance(c.embedding, [{}], false, {{'distanceFunction': 'euclidean'}}) AS " \
"SimilarityScore FROM c ORDER BY VectorDistance(c.embedding, [{}], false, {{'distanceFunction': 'euclidean'}})" \
.format(str(i), vector_string, vector_string)
flat_list = list(self.created_flat_euclidean_container.query_items(query=vanilla_query,
enable_cross_partition_query=True))
verify_ordering(flat_list, "euclidean")
quantized_list = list(
self.created_quantized_cosine_container.query_items(query=specs_query,
enable_cross_partition_query=True))
verify_ordering(quantized_list, "euclidean")
disk_ann_list = list(
self.created_diskANN_dotproduct_container.query_items(query=specs_query,
enable_cross_partition_query=True))
verify_ordering(disk_ann_list, "euclidean")
# test cosine distance
for i in range(1, 11):
vanilla_query = "SELECT TOP {} c.text, VectorDistance(c.embedding, [{}]) AS " \
"SimilarityScore FROM c ORDER BY VectorDistance(c.embedding, [{}])".format(str(i),
vector_string,
vector_string)
specs_query = "SELECT TOP {} c.text, VectorDistance(c.embedding, [{}], false, {{'distanceFunction': 'cosine'}}) AS " \
"SimilarityScore FROM c ORDER BY VectorDistance(c.embedding, [{}], false, {{'distanceFunction': 'cosine'}})" \
.format(str(i), vector_string, vector_string)
flat_list = list(
self.created_flat_euclidean_container.query_items(query=specs_query, enable_cross_partition_query=True))
verify_ordering(flat_list, "cosine")
quantized_list = list(
self.created_quantized_cosine_container.query_items(query=vanilla_query,
enable_cross_partition_query=True))
verify_ordering(quantized_list, "cosine")
disk_ann_list = list(
self.created_diskANN_dotproduct_container.query_items(query=specs_query,
enable_cross_partition_query=True))
verify_ordering(disk_ann_list, "cosine")
# test dot product distance
for i in range(1, 11):
vanilla_query = "SELECT TOP {} c.text, VectorDistance(c.embedding, [{}]) AS " \
"SimilarityScore FROM c ORDER BY VectorDistance(c.embedding, [{}])".format(str(i),
vector_string,
vector_string)
specs_query = "SELECT TOP {} c.text, VectorDistance(c.embedding, [{}], false, {{'distanceFunction': 'dotproduct'}}) AS " \
"SimilarityScore FROM c ORDER BY VectorDistance(c.embedding, [{}], false, {{'distanceFunction': 'dotproduct'}})" \
.format(str(i), vector_string, vector_string)
flat_list = list(
self.created_flat_euclidean_container.query_items(query=specs_query, enable_cross_partition_query=True))
verify_ordering(flat_list, "dotproduct")
quantized_list = list(
self.created_quantized_cosine_container.query_items(query=specs_query,
enable_cross_partition_query=True))
verify_ordering(quantized_list, "dotproduct")
disk_ann_list = list(
self.created_diskANN_dotproduct_container.query_items(query=vanilla_query,
enable_cross_partition_query=True))
verify_ordering(disk_ann_list, "dotproduct")
def test_vector_query_pagination(self):
# load up previously calculated embedding for the given string
vector_string = vector_test_data.get_embedding_string("I am having a wonderful day.")
query = "SELECT TOP 8 c.text, VectorDistance(c.embedding, [{}], false, {{'distanceFunction': 'cosine'}}) AS " \
"SimilarityScore FROM c ORDER BY VectorDistance(c.embedding, [{}], false, {{'distanceFunction': " \
"'cosine'}})".format(vector_string, vector_string)
query_iterable = self.created_quantized_cosine_container.query_items(query=query,
enable_cross_partition_query=True,
max_item_count=3)
all_fetched_res = []
count = 0
item_pages = query_iterable.by_page()
for items in item_pages:
count += 1
all_fetched_res.extend(list(items))
assert count == 3
assert len(all_fetched_res) == 8
verify_ordering(all_fetched_res, "cosine")
def test_vector_query_large_data(self):
# test different limit queries on a larger data set
embedding_value = 0.0001
for i in range(2000):
item = {"id": str(i), "pk": i % 2, "embedding": [embedding_value, embedding_value]}
self.created_large_container.create_item(item)
embedding_value += 0.0001
query = "SELECT c.id, VectorDistance(c.embedding, [0.0001, 0.0001], false," \
" {'distanceFunction': 'cosine'}) AS SimilarityScore FROM c ORDER BY" \
" VectorDistance(c.embedding, [0.0001, 0.0001], false, {'distanceFunction': 'cosine'})" \
"OFFSET 0 LIMIT 1000"
query_iterable = self.created_large_container.query_items(query=query,
enable_cross_partition_query=True)
result_list = list(query_iterable)
assert len(result_list) == 1000
query = "SELECT DISTINCT c.id, VectorDistance(c.embedding, [0.0001, 0.0001], false," \
" {'distanceFunction': 'cosine'}) AS SimilarityScore FROM c ORDER BY" \
" VectorDistance(c.embedding, [0.0001, 0.0001], false, {'distanceFunction': 'cosine'})" \
"OFFSET 0 LIMIT 1000"
query_iterable = self.created_large_container.query_items(query=query,
enable_cross_partition_query=True)
result_list = list(query_iterable)
assert len(result_list) == 1000
query = "SELECT TOP 750 c.id, VectorDistance(c.embedding, [0.0001, 0.0001], false," \
" {'distanceFunction': 'cosine'}) AS SimilarityScore FROM c ORDER BY" \
" VectorDistance(c.embedding, [0.0001, 0.0001], false, {'distanceFunction': 'cosine'})"
query_iterable = self.created_large_container.query_items(query=query,
enable_cross_partition_query=True)
result_list = list(query_iterable)
assert len(result_list) == 750
query = "SELECT DISTINCT TOP 750 c.id, VectorDistance(c.embedding, [0.0001, 0.0001], false," \
" {'distanceFunction': 'cosine'}) AS SimilarityScore FROM c ORDER BY" \
" VectorDistance(c.embedding, [0.0001, 0.0001], false, {'distanceFunction': 'cosine'})"
query_iterable = self.created_large_container.query_items(query=query,
enable_cross_partition_query=True)
result_list = list(query_iterable)
assert len(result_list) == 750
query = "SELECT c.id, VectorDistance(c.embedding, [0.0001, 0.0001], false," \
" {'distanceFunction': 'cosine'}) AS SimilarityScore FROM c ORDER BY" \
" VectorDistance(c.embedding, [0.0001, 0.0001], false, {'distanceFunction': 'cosine'})" \
" OFFSET 1000 LIMIT 500"
query_iterable = self.created_large_container.query_items(query=query,
enable_cross_partition_query=True)
result_list = list(query_iterable)
assert len(result_list) == 500
query = "SELECT DISTINCT c.id, VectorDistance(c.embedding, [0.0001, 0.0001], false," \
" {'distanceFunction': 'cosine'}) AS SimilarityScore FROM c ORDER BY" \
" VectorDistance(c.embedding, [0.0001, 0.0001], false, {'distanceFunction': 'cosine'})" \
" OFFSET 1000 LIMIT 500"
query_iterable = self.created_large_container.query_items(query=query,
enable_cross_partition_query=True)
result_list = list(query_iterable)
assert len(result_list) == 500
def test_vector_query_cross_partition_response_hook(self):
# load up previously calculated embedding for the given string
vector_string = vector_test_data.get_embedding_string("I am having a wonderful day.")
query = "SELECT TOP 5 c.text, VectorDistance(c.embedding, [{}], false, {{'distanceFunction': 'cosine'}}) AS " \
"SimilarityScore FROM c ORDER BY VectorDistance(c.embedding, [{}], false, {{'distanceFunction': " \
"'cosine'}})".format(vector_string, vector_string)
response_hook = test_config.ResponseHookCaller()
query_iterable = self.created_quantized_cosine_container.query_items(query=query,
enable_cross_partition_query=True,
response_hook=response_hook)
result_list = list(query_iterable)
assert len(result_list) == 5
assert response_hook.count == 2
def test_vector_query_partitioned_response_hook(self):
# load up previously calculated embedding for the given string
vector_string = vector_test_data.get_embedding_string("I am having a wonderful day.")
query = "SELECT TOP 4 c.text, VectorDistance(c.embedding, [{}], false, {{'distanceFunction': 'cosine'}}) AS " \
"SimilarityScore FROM c ORDER BY VectorDistance(c.embedding, [{}], false, {{'distanceFunction': " \
"'cosine'}})".format(vector_string, vector_string)
response_hook = test_config.ResponseHookCaller()
query_iterable = self.created_quantized_cosine_container.query_items(query=query,
partition_key='1',
response_hook=response_hook)
result_list = list(query_iterable)
assert len(result_list) == 4
assert response_hook.count == 1
if __name__ == "__main__":
unittest.main()
|
