Boosting Fields and Queries: Controlling Relevance

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

import math
from dataclasses import dataclass
from typing import List, Dict, Optional
from enum import Enum
 
class BoostMode(Enum):
    """How to combine boost with base score."""
    MULTIPLY = "multiply"    # score * boost (most common)
    ADD = "add"              # score + boost
    REPLACE = "replace"      # boost (ignore base score)
    MAX = "max"              # max(score, boost)
    MIN = "min"              # min(score, boost)
    AVG = "avg"              # (score + boost) / 2
 
class ScoreMode(Enum):
    """How to combine multiple boosts."""
    MULTIPLY = "multiply"    # boost1 * boost2 * ...
    SUM = "sum"              # boost1 + boost2 + ...
    AVG = "avg"              # average of boosts
    MAX = "max"              # maximum boost
    MIN = "min"              # minimum boost
    FIRST = "first"          # first matching boost
 
@dataclass
class BoostConfig:
    """Configuration for a single boost factor."""
    name: str
    value: float
    mode: BoostMode = BoostMode.MULTIPLY
    
    # Normalization parameters
    min_value: Optional[float] = None  # Clamp boost to minimum
    max_value: Optional[float] = None  # Clamp boost to maximum
    
    def normalize(self, boost: float) -> float:
        """Apply normalization constraints to boost value."""
        if self.min_value is not None:
            boost = max(self.min_value, boost)
        if self.max_value is not None:
            boost = min(self.max_value, boost)
        return boost
 
class BoostCalculator:
    """
    Demonstrates different boosting strategies and their effects.
    """
    
    @staticmethod
    def multiplicative_boost(base_score: float, boost_factor: float) -> float:
        """
        Standard multiplicative boost.
        Preserves relative ordering within boosted set.
        """
        return base_score * boost_factor
    
    @staticmethod
    def additive_boost(base_score: float, boost_amount: float) -> float:
        """
        Additive boost, Can flip rankings.
        Useful when you want absolute priority for certain signals.
        """
        return base_score + boost_amount
    
    @staticmethod
    def logarithmic_boost(base_score: float, factor: float, 
                          log_base: float = math.e) -> float:
        """
        Log-dampened boost. Prevents extreme boosts from dominating.
        Useful for popularity signals that span orders of magnitude.
        
        Example: page views range from 10 to 10,000,000
                 Raw multiplicative boost would make high-view pages
                 dominate regardless of relevance.
                 Log dampening: log(10) = 2.3, log(10M) = 16.1
        """
        return base_score * (1 + factor * math.log(1 + base_score, log_base))
    
    @staticmethod
    def saturation_boost(base_score: float, boost_value: float,
                        pivot: float = 1.0) -> float:
        """
        Sigmoid-like saturation boost.
        Provides diminishing returns for very large boost values.
        
        Formula: score * (boost / (boost + pivot))
        
        When boost >> pivot: factor approaches 1.0
        When boost = pivot: factor = 0.5
        When boost << pivot: factor approaches 0
        """
        boost_factor = boost_value / (boost_value + pivot)
        return base_score * boost_factor
 
 
def combine_multiple_boosts(base_score: float, 
                            boosts: List[BoostConfig],
                            score_mode: ScoreMode = ScoreMode.MULTIPLY) -> float:
    """
    Combine multiple boost factors according to score mode.
    
    Example use case: E-commerce search with multiple boosts
    - Field boost: title match = 2.0
    - Freshness boost: 0.8 - 1.2 based on age
    - Popularity boost: 0.9 - 1.5 based on sales
    - Quality boost: 0.5 - 1.0 based on rating
    """
    if not boosts:
        return base_score
    
    boost_values = [b.normalize(b.value) for b in boosts]
    
    if score_mode == ScoreMode.MULTIPLY:
        combined = 1.0
        for bv in boost_values:
            combined *= bv
            
    elif score_mode == ScoreMode.SUM:
        combined = sum(boost_values)
        
    elif score_mode == ScoreMode.AVG:
        combined = sum(boost_values) / len(boost_values)
        
    elif score_mode == ScoreMode.MAX:
        combined = max(boost_values)
        
    elif score_mode == ScoreMode.MIN:
        combined = min(boost_values)
        
    elif score_mode == ScoreMode.FIRST:
        combined = boost_values[0] if boost_values else 1.0
        
    else:
        combined = 1.0
    
    return base_score * combined
 
 
def boost_impact_demonstration():
    """
    Demonstrate how different boost strategies affect ranking.
    """
    # Scenario: 3 documents with different base scores
    documents = [
        {"id": "A", "base_score": 10.0, "popularity": 1000, "freshness": 0.9},
        {"id": "B", "base_score": 8.0, "popularity": 5000, "freshness": 0.6},
        {"id": "C", "base_score": 5.0, "popularity": 50000, "freshness": 1.0},
    ]
    
    print("Base ranking: A (10.0) > B (8.0) > C (5.0)")
    print()
    
    # Apply different boost strategies
    strategies = [
        ("Raw multiplicative popularity", lambda d: d["base_score"] * d["popularity"]),
        ("Log popularity", lambda d: d["base_score"] * math.log10(d["popularity"])),
        ("Sqrt popularity", lambda d: d["base_score"] * math.sqrt(d["popularity"])),
        ("Saturation popularity", lambda d: BoostCalculator.saturation_boost(
            d["base_score"], d["popularity"], pivot=10000)),
        ("Freshness only", lambda d: d["base_score"] * d["freshness"]),
        ("Combined (log pop × freshness)", lambda d: d["base_score"] * 
            math.log10(d["popularity"]) * d["freshness"]),
    ]
    
    for name, scorer in strategies:
        scored = [(d["id"], scorer(d)) for d in documents]
        scored.sort(key=lambda x: -x[1])
        ranking = " > ".join([f'{d[0]} ({d[1]:.1f})' for d in scored])
        print(f"{name}:")
        print(f"  {ranking}")
        print()

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

{
  "query": {
    "multi_match": {
      "query": "wireless headphones",
      "type": "best_fields",
      "fields": [
        "sku^50",
        "product_name^5",
        "brand^3",
        "category^2",
        "description^1",
        "reviews^0.5"
      ],
      "tie_breaker": 0.3
    }
  }
}
 
// Explanation of parameters:
//
// "type": "best_fields"
//   - Uses the MAXIMUM score from any single field
//   - Good when you expect the match to be in ONE relevant field
//   - Other types:
//     - "most_fields": SUM scores from all matching fields
//     - "cross_fields": Treats fields as one big field
//     - "phrase": Phrase match across fields
//     - "phrase_prefix": Typeahead-style matching
//
// Field boost syntax: "field_name^boost_value"
//   - sku^50: SKU matches are 50x more important than description
//   - reviews^0.5: Review matches are LESS important than description
//
// "tie_breaker": 0.3
//   - When using best_fields, secondary field matches contribute
//     30% of their score (prevents ignoring partial matches)
//   - 0.0 = only best field counts
//   - 1.0 = sum all field scores (defeats purpose of best_fields)

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

{
  "query": {
    "boosting": {
      "positive": {
        "multi_match": {
          "query": "winter jacket",
          "fields": ["product_name^3", "description"]
        }
      },
      "negative": {
        "term": {
          "out_of_stock": true
        }
      },
      "negative_boost": 0.2
    }
  }
}
 
// The boosting query has two parts:
// 1. "positive": Documents matching this get normal scores
// 2. "negative": Documents ALSO matching this get penalized
//
// "negative_boost": 0.2 means:
//   - Out-of-stock items have their scores multiplied by 0.2
//   - They're not excluded, just demoted
//   - Good for soft filtering (prefer in-stock, show out-of-stock last)

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

from dataclasses import dataclass, field
from typing import List, Dict, Any, Optional
from datetime import datetime
import json
 
@dataclass
class UserContext:
    """User context for personalized boosting."""
    user_id: Optional[str] = None
    location: Optional[str] = None
    preferred_brands: List[str] = field(default_factory=list)
    purchase_history_categories: List[str] = field(default_factory=list)
    price_sensitivity: str = "medium"  # low, medium, high
    is_premium_member: bool = False
 
@dataclass
class SearchContext:
    """Search-time context for dynamic boosting."""
    current_time: datetime = field(default_factory=datetime.now)
    is_holiday_season: bool = False
    active_promotions: List[str] = field(default_factory=list)
    inventory_priority: str = "balanced"  # in_stock_first, balanced, include_all
 
class DynamicBoostBuilder:
    """
    Constructs dynamic query-time boosts based on context.
    
    This pattern separates boost logic from query construction,
    making it testable and configurable.
    """
    
    def __init__(self, user_ctx: UserContext, search_ctx: SearchContext):
        self.user = user_ctx
        self.search = search_ctx
        self.should_clauses: List[Dict] = []
    
    def add_brand_affinity_boosts(self) -> 'DynamicBoostBuilder':
        """Boost brands the user has shown preference for."""
        for brand in self.user.preferred_brands:
            self.should_clauses.append({
                "term": {
                    "brand.keyword": {
                        "value": brand,
                        "boost": 1.8  # Significant but not overwhelming
                    }
                }
            })
        return self
    
    def add_category_affinity_boosts(self) -> 'DynamicBoostBuilder':
        """Boost categories from user's purchase history."""
        for category in self.user.purchase_history_categories[:5]:  # Limit
            self.should_clauses.append({
                "term": {
                    "category": {
                        "value": category,
                        "boost": 1.3
                    }
                }
            })
        return self
    
    def add_promotion_boosts(self) -> 'DynamicBoostBuilder':
        """Boost actively promoted products."""
        if self.search.active_promotions:
            self.should_clauses.append({
                "terms": {
                    "promotion_id": self.search.active_promotions,
                    "boost": 2.0
                }
            })
        return self
    
    def add_inventory_boosts(self) -> 'DynamicBoostBuilder':
        """Boost/penalize based on inventory status."""
        if self.search.inventory_priority == "in_stock_first":
            self.should_clauses.append({
                "term": {
                    "in_stock": {
                        "value": True,
                        "boost": 2.0
                    }
                }
            })
            # Optionally add negative boost for out of stock
        return self
    
    def add_freshness_boost(self, field: str = "created_at", 
                           scale: str = "30d",
                           boost: float = 1.5) -> 'DynamicBoostBuilder':
        """Add time-decay boost for recent items."""
        self.should_clauses.append({
            "function_score": {
                "functions": [{
                    "gauss": {
                        field: {
                            "origin": "now",
                            "scale": scale,
                            "decay": 0.5
                        }
                    },
                    "weight": boost
                }],
                "boost_mode": "multiply"
            }
        })
        return self
    
    def add_price_sensitivity_boost(self) -> 'DynamicBoostBuilder':
        """Boost based on user's price sensitivity."""
        if self.user.price_sensitivity == "high":
            # Price-sensitive users: boost lower prices
            self.should_clauses.append({
                "function_score": {
                    "functions": [{
                        "script_score": {
                            "script": {
                                "source": """
                                    double price = doc['price'].value;
                                    double maxPrice = 1000;
                                    return 1 + ((maxPrice - price) / maxPrice) * 0.5;
                                """
                            }
                        }
                    }],
                    "boost_mode": "multiply"
                }
            })
        return self
    
    def add_premium_member_boosts(self) -> 'DynamicBoostBuilder':
        """Premium members see premium products boosted."""
        if self.user.is_premium_member:
            self.should_clauses.append({
                "term": {
                    "is_premium_product": {
                        "value": True,
                        "boost": 1.4
                    }
                }
            })
        return self
    
    def build(self) -> List[Dict]:
        """Return the constructed boost clauses."""
        return self.should_clauses
    
    def build_complete_query(self, base_query: Dict) -> Dict:
        """Wrap base query with dynamic boosts."""
        if not self.should_clauses:
            return base_query
        
        return {
            "bool": {
                "must": base_query,
                "should": self.should_clauses
            }
        }
 
 
# Example usage
def create_personalized_search_query(
    query_text: str,
    user: UserContext,
    search: SearchContext
) -> Dict:
    """
    Build a fully personalized search query.
    """
    # Base text query
    base_query = {
        "multi_match": {
            "query": query_text,
            "fields": ["product_name^5", "brand^3", "description"],
            "type": "best_fields",
            "tie_breaker": 0.3
        }
    }
    
    # Build dynamic boosts
    boost_builder = DynamicBoostBuilder(user, search)
    complete_query = (
        boost_builder
        .add_brand_affinity_boosts()
        .add_category_affinity_boosts()
        .add_promotion_boosts()
        .add_inventory_boosts()
        .add_premium_member_boosts()
        .build_complete_query(base_query)
    )
    
    return {"query": complete_query}
 
 
# Demonstration
if __name__ == "__main__":
    user = UserContext(
        user_id="user_123",
        preferred_brands=["Sony", "Bose"],
        purchase_history_categories=["Electronics", "Headphones"],
        is_premium_member=True
    )
    
    search = SearchContext(
        active_promotions=["HOLIDAY2024"],
        inventory_priority="in_stock_first"
    )
    
    query = create_personalized_search_query("wireless headphones", user, search)
    print(json.dumps(query, indent=2))

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

{
  "query": {
    "function_score": {
      "query": {
        "multi_match": {
          "query": "running shoes",
          "fields": ["product_name^3", "description"]
        }
      },
      "functions": [
        {
          "field_value_factor": {
            "field": "popularity_score",
            "factor": 1.2,
            "modifier": "log1p",
            "missing": 1
          }
        }
      ],
      "boost_mode": "multiply",
      "score_mode": "multiply"
    }
  }
}
 
// field_value_factor explained:
//
// "field": "popularity_score"
//   - The numeric field to use for boosting
//
// "factor": 1.2
//   - Multiplied with the field value before applying modifier
//
// "modifier": "log1p"
//   - How to transform the field value:
//   - none: Use raw value (dangerous for large values)
//   - log: log(value)
//   - log1p: log(1 + value) - handles 0 gracefully
//   - log2p: log(2 + value)
//   - ln: natural log
//   - ln1p: ln(1 + value)
//   - ln2p: ln(2 + value)
//   - square: value^2
//   - sqrt: sqrt(value)
//   - reciprocal: 1/value
//
// "missing": 1
//   - Value to use if field is missing (prevents scoring errors)
//
// For popularity_score=100:
// boost = log1p(1.2 * 100) = log1p(120) ≈ 4.8

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

from enum import Enum
from typing import List, Dict, Any
from dataclasses import dataclass
 
class ScoreMode(Enum):
    MULTIPLY = "multiply"
    SUM = "sum"
    AVG = "avg"
    MAX = "max"
    MIN = "min"
    FIRST = "first"
 
class BoostMode(Enum):
    MULTIPLY = "multiply"
    REPLACE = "replace"
    SUM = "sum"
    AVG = "avg"
    MAX = "max"
    MIN = "min"
 
@dataclass
class ScoringConfig:
    """Configuration for how to combine scores."""
    score_mode: ScoreMode
    boost_mode: BoostMode
    max_boost: float = 10.0  # Prevent runaway scores
 
def choose_scoring_strategy(use_case: str) -> ScoringConfig:
    """
    Select appropriate score/boost modes for common use cases.
    """
    strategies = {
        # E-commerce: All factors contribute, text relevance matters
        "ecommerce_default": ScoringConfig(
            score_mode=ScoreMode.MULTIPLY,
            boost_mode=BoostMode.MULTIPLY,
            max_boost=5.0
        ),
        
        # Recommendations: Signals are independent, add them up
        "recommendations": ScoringConfig(
            score_mode=ScoreMode.SUM,
            boost_mode=BoostMode.MULTIPLY,
            max_boost=10.0
        ),
        
        # Location-based: Proximity can dominate
        "local_search": ScoringConfig(
            score_mode=ScoreMode.MULTIPLY,
            boost_mode=BoostMode.MULTIPLY,
            max_boost=3.0  # Lower to prevent far-but-relevant being buried
        ),
        
        # News/Time-sensitive: Freshness is critical
        "news_search": ScoringConfig(
            score_mode=ScoreMode.MULTIPLY,
            boost_mode=BoostMode.MULTIPLY,
            max_boost=5.0
        ),
        
        # Trending: Popularity/velocity can override relevance
        "trending": ScoringConfig(
            score_mode=ScoreMode.MAX,
            boost_mode=BoostMode.REPLACE,
            max_boost=100.0  # Allow strong trending signals
        ),
        
        # Balanced personalization: Don't let personalization dominate
        "personalized": ScoringConfig(
            score_mode=ScoreMode.AVG,
            boost_mode=BoostMode.MULTIPLY,
            max_boost=2.0  # Conservative to maintain relevance
        ),
    }
    
    return strategies.get(use_case, strategies["ecommerce_default"])
 
 
def build_function_score_query(
    base_query: Dict,
    functions: List[Dict],
    config: ScoringConfig
) -> Dict:
    """
    Build a properly configured function_score query.
    """
    return {
        "function_score": {
            "query": base_query,
            "functions": functions,
            "score_mode": config.score_mode.value,
            "boost_mode": config.boost_mode.value,
            "max_boost": config.max_boost
        }
    }
 
 
# Example: E-commerce query with multiple boost functions
def ecommerce_search_example():
    base_query = {
        "multi_match": {
            "query": "wireless headphones",
            "fields": ["product_name^3", "brand^2", "description"],
            "type": "best_fields"
        }
    }
    
    functions = [
        # Popularity boost (log dampened)
        {
            "field_value_factor": {
                "field": "num_sold",
                "modifier": "log1p",
                "factor": 0.5,
                "missing": 1
            },
            "weight": 1
        },
        # Rating boost
        {
            "field_value_factor": {
                "field": "avg_rating",
                "modifier": "none",
                "missing": 3
            },
            "weight": 0.5
        },
        # Freshness boost (new arrivals)
        {
            "gauss": {
                "created_at": {
                    "origin": "now",
                    "scale": "30d",
                    "decay": 0.5
                }
            },
            "weight": 0.3
        },
        # In-stock boost
        {
            "filter": {
                "term": {"in_stock": True}
            },
            "weight": 1.2
        },
        # Promotion boost (current sales)
        {
            "filter": {
                "term": {"is_on_sale": True}
            },
            "weight": 1.3
        }
    ]
    
    config = choose_scoring_strategy("ecommerce_default")
    return build_function_score_query(base_query, functions, config)