qdrant-edge 0.7.2

A lightweight, in-process vector search engine designed for embedded devices, autonomous systems, and mobile agents.
Documentation 
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

use std::sync::atomic::AtomicBool;

use crate::common::counter::hardware_accumulator::HwMeasurementAcc;
use crate::common::types::PointOffsetType;
use itertools::Itertools;

use crate::sparse::SearchScratch;
use crate::sparse::common::sparse_vector::RemappedSparseVector;
use crate::sparse::index::inverted_index::InvertedIndex;
use crate::sparse::index::search_context::SearchContext;
use crate::sparse::index::tests::common::{build_index, match_all};

fn do_search<I: InvertedIndex>(index: &I, query: RemappedSparseVector) -> HwMeasurementAcc {
    let is_stopped = AtomicBool::new(false);
    let accumulator = HwMeasurementAcc::new();
    let hardware_counter = accumulator.get_counter_cell();
    let top = 10;
    let mut scratch = SearchScratch::new_for_test();
    let mut search_context = SearchContext::new(
        query,
        top,
        index,
        &mut scratch,
        &is_stopped,
        &hardware_counter,
    )
    .unwrap();

    let result = search_context.search(&match_all);
    // there might be less than `top` result
    // happens if index contains less than `top` sparse vectors with indices overlapping the query indices
    assert!(result.len() <= top);

    accumulator
}

fn do_plain_search<I: InvertedIndex>(
    index: &I,
    query: RemappedSparseVector,
    docs: &[PointOffsetType],
) -> HwMeasurementAcc {
    let is_stopped = AtomicBool::new(false);
    let accumulator = HwMeasurementAcc::new();
    let hardware_counter = accumulator.get_counter_cell();
    let top = 10;
    let mut scratch = SearchScratch::new_for_test();
    let mut search_context = SearchContext::new(
        query,
        top,
        index,
        &mut scratch,
        &is_stopped,
        &hardware_counter,
    )
    .unwrap();

    let result = search_context.plain_search(docs);

    // there might be less than `top` result
    // happens if index contains less than `top` sparse vectors with indices overlapping the query indices
    assert!(result.len() <= top);

    accumulator
}

#[test]
fn test_hw_counter_for_sparse_search() {
    let count = 1000;
    let density = 32;
    let vocab1 = 32;
    let vocab2 = 512;

    // Expected posting length = count * density / vocab1 / 2 = 1000 * 32 / 32 / 2 = 500
    // Expected posting length = count * density / vocab2 / 2 = 1000 * 32 / 512 / 2 = 62.5

    let index_f32 = build_index::<f32>(count, density, vocab1, vocab2);
    let index_f16 = build_index::<half::f16>(count, density, vocab1, vocab2);
    let index_u8 = build_index::<u8>(count, density, vocab1, vocab2);

    let freq_query = RemappedSparseVector {
        indices: vec![1],
        values: vec![1.0],
    };

    let freq_query2 = RemappedSparseVector {
        indices: vec![0, 1, 2],
        values: vec![1.0, 1.0, 1.0],
    };

    let infrequent_query = RemappedSparseVector {
        indices: vec![100],
        values: vec![1.0],
    };

    let infrequent_query2 = RemappedSparseVector {
        indices: vec![0, 100, 200],
        values: vec![1.0, 1.0, 1.0],
    };

    let acc_f32_freq_1 = do_search(&index_f32.index, freq_query.clone());
    let acc_f16_freq_1 = do_search(&index_f16.index, freq_query.clone());
    let acc_u8_freq_1 = do_search(&index_u8.index, freq_query.clone());

    let acc_f32_freq_3 = do_search(&index_f32.index, freq_query2.clone());
    let acc_f16_freq_3 = do_search(&index_f16.index, freq_query2.clone());
    let acc_u8_freq_3 = do_search(&index_u8.index, freq_query2.clone());

    let acc_f32_infreq_1 = do_search(&index_f32.index, infrequent_query.clone());
    let acc_f16_infreq_1 = do_search(&index_f16.index, infrequent_query.clone());
    let acc_u8_infreq_1 = do_search(&index_u8.index, infrequent_query.clone());

    let acc_f32_infreq_3 = do_search(&index_f32.index, infrequent_query2.clone());
    let acc_f16_infreq_3 = do_search(&index_f16.index, infrequent_query2.clone());
    let acc_u8_infreq_3 = do_search(&index_u8.index, infrequent_query2.clone());

    // Higher precision floats cost more CPU and IO
    assert!(acc_f32_freq_1.get_cpu() > acc_f16_freq_1.get_cpu());
    assert!(acc_f16_freq_1.get_cpu() > acc_u8_freq_1.get_cpu());

    assert!(acc_f32_freq_1.get_vector_io_read() > acc_f16_freq_1.get_vector_io_read());
    assert!(acc_f16_freq_1.get_vector_io_read() > acc_u8_freq_1.get_vector_io_read());

    // More indices are more expensive than less indices
    assert!(acc_f32_freq_3.get_cpu() > acc_f32_freq_1.get_cpu());
    assert!(acc_f16_freq_3.get_cpu() > acc_f16_freq_1.get_cpu());
    assert!(acc_u8_freq_3.get_cpu() > acc_u8_freq_1.get_cpu());

    assert!(acc_f32_freq_3.get_vector_io_read() > acc_f32_freq_1.get_vector_io_read());
    assert!(acc_f16_freq_3.get_vector_io_read() > acc_f16_freq_1.get_vector_io_read());
    assert!(acc_u8_freq_3.get_vector_io_read() > acc_u8_freq_1.get_vector_io_read());

    // Frequent terms are more expensive than infrequent terms

    assert!(acc_f32_freq_1.get_cpu() > acc_f32_infreq_1.get_cpu());
    assert!(acc_f16_freq_1.get_cpu() > acc_f16_infreq_1.get_cpu());
    assert!(acc_u8_freq_1.get_cpu() > acc_u8_infreq_1.get_cpu());

    assert!(acc_f32_freq_1.get_vector_io_read() > acc_f32_infreq_1.get_vector_io_read());
    assert!(acc_f16_freq_1.get_vector_io_read() > acc_f16_infreq_1.get_vector_io_read());
    assert!(acc_u8_freq_1.get_vector_io_read() > acc_u8_infreq_1.get_vector_io_read());

    // More indices are more expensive than less indices

    assert!(acc_f32_infreq_3.get_cpu() > acc_f32_infreq_1.get_cpu());
    assert!(acc_f16_infreq_3.get_cpu() > acc_f16_infreq_1.get_cpu());
    assert!(acc_u8_infreq_3.get_cpu() > acc_u8_infreq_1.get_cpu());

    assert!(acc_f32_infreq_3.get_vector_io_read() > acc_f32_infreq_1.get_vector_io_read());
    assert!(acc_f16_infreq_3.get_vector_io_read() > acc_f16_infreq_1.get_vector_io_read());
    assert!(acc_u8_infreq_3.get_vector_io_read() > acc_u8_infreq_1.get_vector_io_read());
}

#[test]
fn test_hw_counter_for_plain_sparse_search() {
    let count = 1000;
    let density = 32;
    let vocab1 = 32;
    let vocab2 = 512;

    // Expected posting length = count * density / vocab1 / 2 = 1000 * 32 / 32 / 2 = 500
    // Expected posting length = count * density / vocab2 / 2 = 1000 * 32 / 512 / 2 = 62.5

    let index_f32 = build_index::<f32>(count, density, vocab1, vocab2);
    let index_f16 = build_index::<half::f16>(count, density, vocab1, vocab2);
    let index_u8 = build_index::<u8>(count, density, vocab1, vocab2);

    let documents_to_score = 500;

    // Plain search requires list of document ids
    // Generate random document ids
    let document_ids: Vec<PointOffsetType> = (0..documents_to_score)
        .map(|_| rand::random::<PointOffsetType>() % count as PointOffsetType)
        .unique()
        .collect();

    let freq_query = RemappedSparseVector {
        indices: vec![1],
        values: vec![1.0],
    };

    let freq_query2 = RemappedSparseVector {
        indices: vec![0, 1, 2],
        values: vec![1.0, 1.0, 1.0],
    };

    let infreq_query = RemappedSparseVector {
        indices: vec![100],
        values: vec![1.0],
    };

    let infreq_query2 = RemappedSparseVector {
        indices: vec![101, 102, 103],
        values: vec![1.0, 1.0, 1.0],
    };

    let acc_f32_freq_1 = do_plain_search(&index_f32.index, freq_query.clone(), &document_ids);
    let acc_f16_freq_1 = do_plain_search(&index_f16.index, freq_query.clone(), &document_ids);
    let acc_u8_freq_1 = do_plain_search(&index_u8.index, freq_query.clone(), &document_ids);

    let acc_f32_freq_3 = do_plain_search(&index_f32.index, freq_query2.clone(), &document_ids);
    let acc_f16_freq_3 = do_plain_search(&index_f16.index, freq_query2.clone(), &document_ids);
    let acc_u8_freq_3 = do_plain_search(&index_u8.index, freq_query2.clone(), &document_ids);

    let acc_f32_infreq_1 = do_plain_search(&index_f32.index, infreq_query.clone(), &document_ids);
    let acc_f16_infreq_1 = do_plain_search(&index_f16.index, infreq_query.clone(), &document_ids);
    let acc_u8_infreq_1 = do_plain_search(&index_u8.index, infreq_query.clone(), &document_ids);

    let acc_f32_infreq_3 = do_plain_search(&index_f32.index, infreq_query2.clone(), &document_ids);
    let acc_f16_infreq_3 = do_plain_search(&index_f16.index, infreq_query2.clone(), &document_ids);
    let acc_u8_infreq_3 = do_plain_search(&index_u8.index, infreq_query2.clone(), &document_ids);

    // Higher precision floats cost more IO, but might have same CPU
    assert!(acc_f32_freq_1.get_cpu() >= acc_f16_freq_1.get_cpu());
    assert!(acc_f16_freq_1.get_cpu() >= acc_u8_freq_1.get_cpu());

    assert!(acc_f32_freq_1.get_vector_io_read() > acc_f16_freq_1.get_vector_io_read());
    assert!(acc_f16_freq_1.get_vector_io_read() > acc_u8_freq_1.get_vector_io_read());

    // More indices are more expensive than less indices
    assert!(acc_f32_freq_3.get_cpu() > acc_f32_freq_1.get_cpu());
    assert!(acc_f16_freq_3.get_cpu() > acc_f16_freq_1.get_cpu());
    assert!(acc_u8_freq_3.get_cpu() > acc_u8_freq_1.get_cpu());

    assert!(acc_f32_freq_3.get_vector_io_read() > acc_f32_freq_1.get_vector_io_read());
    assert!(acc_f16_freq_3.get_vector_io_read() > acc_f16_freq_1.get_vector_io_read());
    assert!(acc_u8_freq_3.get_vector_io_read() > acc_u8_freq_1.get_vector_io_read());

    // Frequent terms are more expensive than infrequent terms

    assert!(acc_f32_freq_1.get_cpu() > acc_f32_infreq_1.get_cpu());
    assert!(acc_f16_freq_1.get_cpu() > acc_f16_infreq_1.get_cpu());
    assert!(acc_u8_freq_1.get_cpu() > acc_u8_infreq_1.get_cpu());

    assert!(acc_f32_freq_1.get_vector_io_read() > acc_f32_infreq_1.get_vector_io_read());
    assert!(acc_f16_freq_1.get_vector_io_read() > acc_f16_infreq_1.get_vector_io_read());
    assert!(acc_u8_freq_1.get_vector_io_read() > acc_u8_infreq_1.get_vector_io_read());

    // More indices are more expensive than less indices

    assert!(acc_f32_infreq_3.get_cpu() > acc_f32_infreq_1.get_cpu());
    assert!(acc_f16_infreq_3.get_cpu() > acc_f16_infreq_1.get_cpu());
    assert!(acc_u8_infreq_3.get_cpu() > acc_u8_infreq_1.get_cpu());

    assert!(acc_f32_infreq_3.get_vector_io_read() > acc_f32_infreq_1.get_vector_io_read());
    assert!(acc_f16_infreq_3.get_vector_io_read() > acc_f16_infreq_1.get_vector_io_read());
    assert!(acc_u8_infreq_3.get_vector_io_read() > acc_u8_infreq_1.get_vector_io_read());
}