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
|
// Copyright Yahoo. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
package com.yahoo.search.ranking;
import ai.vespa.models.evaluation.FunctionEvaluator;
import com.yahoo.tensor.Tensor;
import com.yahoo.tensor.TensorType;
import com.yahoo.vespa.config.search.RankProfilesConfig;
import java.util.*;
import java.util.function.Supplier;
class GlobalPhaseSetup {
final FunEvalSpec globalPhaseEvalSpec;
final int rerankCount;
final Collection<String> matchFeaturesToHide;
final List<NormalizerSetup> normalizers;
final Map<String, Tensor> defaultValues;
GlobalPhaseSetup(FunEvalSpec globalPhaseEvalSpec,
final int rerankCount,
Collection<String> matchFeaturesToHide,
List<NormalizerSetup> normalizers,
Map<String, Tensor> defaultValues)
{
this.globalPhaseEvalSpec = globalPhaseEvalSpec;
this.rerankCount = rerankCount;
this.matchFeaturesToHide = matchFeaturesToHide;
this.normalizers = normalizers;
this.defaultValues = defaultValues;
}
static class DefaultQueryFeatureExtractor {
final String baseName;
final String qfName;
TensorType type = null;
Tensor value = null;
DefaultQueryFeatureExtractor(String unwrappedQueryFeature) {
baseName = unwrappedQueryFeature;
qfName = "query(" + baseName + ")";
}
List<String> lookingFor() {
return List.of(qfName, "vespa.type.query." + baseName);
}
void accept(String key, String propValue) {
if (key.equals(qfName)) {
this.value = Tensor.from(propValue);
} else {
this.type = TensorType.fromSpec(propValue);
}
}
Tensor extract() {
if (value != null) {
return value;
}
if (type != null) {
return Tensor.Builder.of(type).build();
}
return Tensor.from(0.0);
}
}
static private Map<String, Tensor> extraDefaultQueryFeatureValues(RankProfilesConfig.Rankprofile rp,
List<String> fromQuery,
List<NormalizerSetup> normalizers)
{
Map<String, DefaultQueryFeatureExtractor> extractors = new HashMap<>();
for (String fn : fromQuery) {
extractors.put(fn, new DefaultQueryFeatureExtractor(fn));
}
for (var n : normalizers) {
for (String fn : n.inputEvalSpec().fromQuery()) {
extractors.put(fn, new DefaultQueryFeatureExtractor(fn));
}
}
Map<String, DefaultQueryFeatureExtractor> targets = new HashMap<>();
for (var extractor : extractors.values()) {
for (String key : extractor.lookingFor()) {
var old = targets.put(key, extractor);
if (old != null) {
throw new IllegalStateException("Multiple targets for key: " + key);
}
}
}
for (var prop : rp.fef().property()) {
var extractor = targets.get(prop.name());
if (extractor != null) {
extractor.accept(prop.name(), prop.value());
}
}
Map<String, Tensor> defaultValues = new HashMap<>();
for (var extractor : extractors.values()) {
defaultValues.put(extractor.qfName, extractor.extract());
}
return defaultValues;
}
static GlobalPhaseSetup maybeMakeSetup(RankProfilesConfig.Rankprofile rp, RankProfilesEvaluator modelEvaluator) {
var model = modelEvaluator.modelForRankProfile(rp.name());
Map<String, RankProfilesConfig.Rankprofile.Normalizer> availableNormalizers = new HashMap<>();
for (var n : rp.normalizer()) {
availableNormalizers.put(n.name(), n);
}
Supplier<FunctionEvaluator> functionEvaluatorSource = null;
int rerankCount = -1;
Set<String> namesToHide = new HashSet<>();
Set<String> matchFeatures = new HashSet<>();
Map<String, String> renameFeatures = new HashMap<>();
String toRename = null;
for (var prop : rp.fef().property()) {
if (prop.name().equals("vespa.globalphase.rerankcount")) {
rerankCount = Integer.valueOf(prop.value());
}
if (prop.name().equals("vespa.rank.globalphase")) {
functionEvaluatorSource = () -> model.evaluatorOf("globalphase");
}
if (prop.name().equals("vespa.hidden.matchfeature")) {
namesToHide.add(prop.value());
}
if (prop.name().equals("vespa.match.feature")) {
matchFeatures.add(prop.value());
}
if (prop.name().equals("vespa.feature.rename")) {
if (toRename == null) {
toRename = prop.value();
} else {
renameFeatures.put(toRename, prop.value());
toRename = null;
}
}
}
for (var entry : renameFeatures.entrySet()) {
String old = entry.getKey();
if (matchFeatures.contains(old)) {
matchFeatures.remove(old);
matchFeatures.add(entry.getValue());
}
}
if (rerankCount < 0) {
rerankCount = 100;
}
if (functionEvaluatorSource != null) {
var evaluator = functionEvaluatorSource.get();
var allInputs = List.copyOf(evaluator.function().arguments());
List<String> fromMF = new ArrayList<>();
List<String> fromQuery = new ArrayList<>();
List<NormalizerSetup> normalizers = new ArrayList<>();
for (var input : allInputs) {
String queryFeatureName = asQueryFeature(input);
if (queryFeatureName != null) {
fromQuery.add(queryFeatureName);
} else if (availableNormalizers.containsKey(input)) {
var cfg = availableNormalizers.get(input);
String normInput = cfg.input();
if (matchFeatures.contains(normInput)) {
Supplier<Evaluator> normSource = () -> new DummyEvaluator(normInput);
normalizers.add(makeNormalizerSetup(cfg, matchFeatures, normSource, List.of(normInput), rerankCount));
} else {
Supplier<FunctionEvaluator> normSource = () -> model.evaluatorOf(normInput);
var normInputs = List.copyOf(normSource.get().function().arguments());
var normSupplier = SimpleEvaluator.wrap(normSource);
normalizers.add(makeNormalizerSetup(cfg, matchFeatures, normSupplier, normInputs, rerankCount));
}
} else if (matchFeatures.contains(input) || matchFeatures.contains(WrappedHit.alternate(input))) {
fromMF.add(input);
} else {
throw new IllegalArgumentException("Bad config, missing global-phase input: " + input);
}
}
Supplier<Evaluator> supplier = SimpleEvaluator.wrap(functionEvaluatorSource);
var gfun = new FunEvalSpec(supplier, fromQuery, fromMF);
var defaultValues = extraDefaultQueryFeatureValues(rp, fromQuery, normalizers);
return new GlobalPhaseSetup(gfun, rerankCount, namesToHide, normalizers, defaultValues);
}
return null;
}
private static NormalizerSetup makeNormalizerSetup(RankProfilesConfig.Rankprofile.Normalizer cfg,
Set<String> matchFeatures,
Supplier<Evaluator> evalSupplier,
List<String> normInputs,
int rerankCount)
{
List<String> fromQuery = new ArrayList<>();
List<String> fromMF = new ArrayList<>();
for (var input : normInputs) {
String queryFeatureName = asQueryFeature(input);
if (queryFeatureName != null) {
fromQuery.add(queryFeatureName);
} else if (matchFeatures.contains(input) || matchFeatures.contains(WrappedHit.alternate(input))) {
fromMF.add(input);
} else {
throw new IllegalArgumentException("Bad config, missing normalizer input: " + input);
}
}
var fun = new FunEvalSpec(evalSupplier, fromQuery, fromMF);
return new NormalizerSetup(cfg.name(), makeNormalizerSupplier(cfg, rerankCount), fun);
}
private static Supplier<Normalizer> makeNormalizerSupplier(RankProfilesConfig.Rankprofile.Normalizer cfg, int rerankCount) {
return switch (cfg.algo()) {
case LINEAR -> () -> new LinearNormalizer(rerankCount);
case RRANK -> () -> new ReciprocalRankNormalizer(rerankCount, cfg.kparam());
};
}
static String asQueryFeature(String input) {
var optRef = com.yahoo.searchlib.rankingexpression.Reference.simple(input);
if (optRef.isPresent()) {
var ref = optRef.get();
if (ref.isSimple() && ref.name().equals("query")) {
return ref.simpleArgument().get();
}
}
return null;
}
}
|
