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
|
// Copyright Vespa.ai. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
#include "nearest_neighbor_blueprint.h"
#include "emptysearch.h"
#include "nearest_neighbor_iterator.h"
#include "nns_index_iterator.h"
#include <vespa/searchlib/fef/termfieldmatchdataarray.h>
#include <vespa/searchlib/tensor/dense_tensor_attribute.h>
#include <vespa/searchlib/tensor/distance_function_factory.h>
#include <vespa/vespalib/objects/objectvisitor.h>
#include <vespa/log/log.h>
LOG_SETUP(".searchlib.queryeval.nearest_neighbor_blueprint");
using vespalib::eval::Value;
namespace vespalib { class Doom; }
namespace search::queryeval {
namespace {
vespalib::string
to_string(NearestNeighborBlueprint::Algorithm algorithm)
{
using NNBA = NearestNeighborBlueprint::Algorithm;
switch (algorithm) {
case NNBA::EXACT: return "exact";
case NNBA::EXACT_FALLBACK: return "exact fallback";
case NNBA::INDEX_TOP_K: return "index top k";
case NNBA::INDEX_TOP_K_WITH_FILTER: return "index top k using filter";
}
return "unknown";
}
} // namespace <unnamed>
NearestNeighborBlueprint::NearestNeighborBlueprint(const queryeval::FieldSpec& field,
std::unique_ptr<search::tensor::DistanceCalculator> distance_calc,
uint32_t target_hits,
bool approximate,
uint32_t explore_additional_hits,
double distance_threshold,
double global_filter_lower_limit,
double global_filter_upper_limit,
double target_hits_max_adjustment_factor,
const vespalib::Doom& doom)
: ComplexLeafBlueprint(field),
_distance_calc(std::move(distance_calc)),
_attr_tensor(_distance_calc->attribute_tensor()),
_query_tensor(_distance_calc->query_tensor()),
_target_hits(target_hits),
_adjusted_target_hits(target_hits),
_approximate(approximate),
_explore_additional_hits(explore_additional_hits),
_distance_threshold(std::numeric_limits<double>::max()),
_global_filter_lower_limit(global_filter_lower_limit),
_global_filter_upper_limit(global_filter_upper_limit),
_target_hits_max_adjustment_factor(target_hits_max_adjustment_factor),
_distance_heap(target_hits),
_found_hits(),
_algorithm(Algorithm::EXACT),
_global_filter(GlobalFilter::create()),
_global_filter_set(false),
_global_filter_hits(),
_global_filter_hit_ratio(),
_doom(doom)
{
if (distance_threshold < std::numeric_limits<double>::max()) {
_distance_threshold = _distance_calc->function().convert_threshold(distance_threshold);
_distance_heap.set_distance_threshold(_distance_threshold);
}
uint32_t est_hits = _attr_tensor.get_num_docs();
setEstimate(HitEstimate(est_hits, false));
auto nns_index = _attr_tensor.nearest_neighbor_index();
set_want_global_filter(nns_index && _approximate);
}
NearestNeighborBlueprint::~NearestNeighborBlueprint() = default;
void
NearestNeighborBlueprint::set_global_filter(const GlobalFilter &global_filter, double estimated_hit_ratio)
{
_global_filter = global_filter.shared_from_this();
_global_filter_set = true;
auto nns_index = _attr_tensor.nearest_neighbor_index();
if (_approximate && nns_index) {
uint32_t est_hits = _attr_tensor.get_num_docs();
if (_global_filter->is_active()) { // pre-filtering case
_global_filter_hits = _global_filter->count();
_global_filter_hit_ratio = static_cast<double>(_global_filter_hits.value()) / est_hits;
if (_global_filter_hit_ratio.value() < _global_filter_lower_limit) {
_algorithm = Algorithm::EXACT_FALLBACK;
} else {
est_hits = std::min(est_hits, _global_filter_hits.value());
}
} else { // post-filtering case
// The goal is to expose 'targetHits' hits to first-phase ranking.
// We try to achieve this by adjusting targetHits based on the estimated hit ratio of the query before post-filtering.
// However, this is bound by 'target-hits-max-adjustment-factor' to limit the cost of searching the HNSW index.
if (estimated_hit_ratio > 0.0) {
_adjusted_target_hits = std::min(static_cast<double>(_target_hits) / estimated_hit_ratio,
static_cast<double>(_target_hits) * _target_hits_max_adjustment_factor);
}
}
if (_algorithm != Algorithm::EXACT_FALLBACK) {
est_hits = std::min(est_hits, _adjusted_target_hits);
setEstimate(HitEstimate(est_hits, false));
perform_top_k(nns_index);
}
}
}
void
NearestNeighborBlueprint::perform_top_k(const search::tensor::NearestNeighborIndex* nns_index)
{
uint32_t k = _adjusted_target_hits;
const auto &df = _distance_calc->function();
if (_global_filter->is_active()) {
_found_hits = nns_index->find_top_k_with_filter(k, df, *_global_filter, k + _explore_additional_hits, _doom, _distance_threshold);
_algorithm = Algorithm::INDEX_TOP_K_WITH_FILTER;
} else {
_found_hits = nns_index->find_top_k(k, df, k + _explore_additional_hits, _doom, _distance_threshold);
_algorithm = Algorithm::INDEX_TOP_K;
}
}
void
NearestNeighborBlueprint::sort(InFlow in_flow)
{
strict(in_flow.strict());
}
std::unique_ptr<SearchIterator>
NearestNeighborBlueprint::createLeafSearch(const search::fef::TermFieldMatchDataArray& tfmda) const
{
assert(tfmda.size() == 1);
fef::TermFieldMatchData &tfmd = *tfmda[0]; // always search in only one field
switch (_algorithm) {
case Algorithm::INDEX_TOP_K_WITH_FILTER:
case Algorithm::INDEX_TOP_K:
return NnsIndexIterator::create(tfmd, _found_hits, _distance_calc->function());
default:
;
}
return NearestNeighborIterator::create(strict(), tfmd,
std::make_unique<search::tensor::DistanceCalculator>(_attr_tensor, _query_tensor),
_distance_heap, *_global_filter);
}
void
NearestNeighborBlueprint::visitMembers(vespalib::ObjectVisitor& visitor) const
{
ComplexLeafBlueprint::visitMembers(visitor);
visitor.visitString("attribute_tensor", _attr_tensor.getTensorType().to_spec());
visitor.visitString("query_tensor", _query_tensor.type().to_spec());
visitor.visitInt("target_hits", _target_hits);
visitor.visitInt("adjusted_target_hits", _adjusted_target_hits);
visitor.visitInt("explore_additional_hits", _explore_additional_hits);
visitor.visitBool("wanted_approximate", _approximate);
visitor.visitBool("has_index", _attr_tensor.nearest_neighbor_index());
visitor.visitString("algorithm", to_string(_algorithm));
visitor.visitInt("top_k_hits", _found_hits.size());
visitor.openStruct("global_filter", "GlobalFilter");
visitor.visitBool("wanted", getState().want_global_filter());
visitor.visitBool("set", _global_filter_set);
visitor.visitBool("calculated", _global_filter->is_active());
visitor.visitFloat("lower_limit", _global_filter_lower_limit);
visitor.visitFloat("upper_limit", _global_filter_upper_limit);
if (_global_filter_hits.has_value()) {
visitor.visitInt("hits", _global_filter_hits.value());
}
if (_global_filter_hit_ratio.has_value()) {
visitor.visitFloat("hit_ratio", _global_filter_hit_ratio.value());
}
visitor.closeStruct();
}
bool
NearestNeighborBlueprint::always_needs_unpack() const
{
return true;
}
std::ostream&
operator<<(std::ostream& out, NearestNeighborBlueprint::Algorithm algorithm)
{
out << to_string(algorithm);
return out;
}
}
|
