File size: 6,422 Bytes
dcc24f8 |
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 |
"""
FinEE Confidence - Scoring logic for extraction results.
Calculates confidence scores based on:
- Source of extraction (regex > rules > LLM)
- Completeness of fields
- Consistency between sources
"""
from typing import Dict, List, Optional
from .schema import ExtractionResult, Confidence, ExtractionSource, FieldMeta
# Field weights for confidence calculation
FIELD_WEIGHTS = {
'amount': 0.25, # Critical field
'type': 0.15, # Critical field
'date': 0.15,
'account': 0.10,
'reference': 0.10,
'merchant': 0.10,
'category': 0.10,
'vpa': 0.05,
}
# Source reliability scores
SOURCE_SCORES = {
ExtractionSource.REGEX: 0.95,
ExtractionSource.RULES: 0.85,
ExtractionSource.LLM: 0.70,
ExtractionSource.CACHE: 1.0, # Cached results are already validated
}
def calculate_confidence_score(result: ExtractionResult) -> float:
"""
Calculate overall confidence score (0.0 to 1.0).
Args:
result: Extraction result with metadata
Returns:
Confidence score between 0.0 and 1.0
"""
if not result:
return 0.0
total_weight = 0.0
weighted_score = 0.0
for field_name, weight in FIELD_WEIGHTS.items():
value = getattr(result, field_name, None)
if value is not None:
# Get source-based score
if field_name in result.meta:
source = result.meta[field_name].source
field_score = SOURCE_SCORES.get(source, 0.5)
# Apply field-specific confidence if available
if result.meta[field_name].confidence:
field_score *= result.meta[field_name].confidence
else:
# Default score for fields without metadata
field_score = 0.5
weighted_score += weight * field_score
total_weight += weight
if total_weight == 0:
return 0.0
return weighted_score / total_weight
def calculate_completeness(result: ExtractionResult,
required: List[str] = None,
desired: List[str] = None) -> float:
"""
Calculate field completeness score.
Args:
result: Extraction result
required: List of required field names
desired: List of desired field names
Returns:
Completeness score (0.0 to 1.0)
"""
if required is None:
required = ['amount', 'type']
if desired is None:
desired = ['merchant', 'category', 'date', 'reference']
required_score = 0.0
for field in required:
if getattr(result, field, None) is not None:
required_score += 1.0
required_score /= len(required) if required else 1.0
desired_score = 0.0
for field in desired:
if getattr(result, field, None) is not None:
desired_score += 1.0
desired_score /= len(desired) if desired else 1.0
# Required fields are weighted more heavily
return 0.7 * required_score + 0.3 * desired_score
def determine_confidence_level(score: float,
high_threshold: float = 0.9,
medium_threshold: float = 0.7) -> Confidence:
"""
Determine confidence level from score.
Args:
score: Confidence score (0.0 to 1.0)
high_threshold: Threshold for HIGH confidence
medium_threshold: Threshold for MEDIUM confidence
Returns:
Confidence enum value
"""
if score >= high_threshold:
return Confidence.HIGH
elif score >= medium_threshold:
return Confidence.MEDIUM
elif score > 0:
return Confidence.LOW
else:
return Confidence.FAILED
def update_result_confidence(result: ExtractionResult,
high_threshold: float = 0.9,
medium_threshold: float = 0.7) -> ExtractionResult:
"""
Update the confidence fields on an ExtractionResult.
Args:
result: Extraction result to update
high_threshold: Threshold for HIGH confidence
medium_threshold: Threshold for MEDIUM confidence
Returns:
Updated ExtractionResult
"""
# Calculate score
score = calculate_confidence_score(result)
# Factor in completeness
completeness = calculate_completeness(result)
combined_score = 0.7 * score + 0.3 * completeness
# Update result
result.confidence_score = combined_score
result.confidence = determine_confidence_level(
combined_score,
high_threshold,
medium_threshold
)
return result
def should_use_llm(result: ExtractionResult,
required: List[str] = None,
desired: List[str] = None) -> bool:
"""
Determine if LLM should be used for additional extraction.
Args:
result: Current extraction result
required: Required fields
desired: Desired fields
Returns:
True if LLM extraction is recommended
"""
missing = result.get_missing_fields(required, desired)
# Always use LLM if required fields are missing
if required:
for field in required:
if field in missing:
return True
# Use LLM if more than half of desired fields are missing
if desired:
missing_desired = [f for f in missing if f in desired]
if len(missing_desired) > len(desired) / 2:
return True
return False
def get_extraction_summary(result: ExtractionResult) -> Dict[str, str]:
"""
Get a summary of extraction sources for each field.
Args:
result: Extraction result
Returns:
Dict mapping field names to source descriptions
"""
summary = {}
for field_name in FIELD_WEIGHTS.keys():
value = getattr(result, field_name, None)
if value is not None:
if field_name in result.meta:
source = result.meta[field_name].source.value
conf = result.meta[field_name].confidence
summary[field_name] = f"{source} ({conf:.0%})"
else:
summary[field_name] = "unknown"
else:
summary[field_name] = "missing"
return summary
|