Download src/confidence.py from Zayeemk/DocuPurse: direct link, hf CLI and curl.
- Browser
- Download file 3.78 kB
-
https://huggingface.co/spaces/Zayeemk/DocuPurse/resolve/main/src/confidence.py
- Command line
-
hf download hf://spaces/Zayeemk/DocuPurse/src/confidence.py
-
curl -L -o confidence.py https://huggingface.co/spaces/Zayeemk/DocuPurse/resolve/main/src/confidence.py
3.78 kB
| """Confidence scoring for DocuParse AI extractions. | |
| Confidence is calculated from transparent heuristics - NOT from LLM self-reported confidence. | |
| Signals: | |
| - Field extraction completeness | |
| - Document classification confidence | |
| - Pydantic validation success | |
| - Arithmetic reconciliation results | |
| - Date validation success | |
| """ | |
| from __future__ import annotations | |
| from typing import Dict, Tuple | |
| from .schemas import DocumentExtraction | |
| _FIELD_WEIGHTS: Dict[str, float] = { | |
| "document_id": 0.10, | |
| "vendor": 0.10, | |
| "customer": 0.05, | |
| "document_date": 0.10, | |
| "due_date": 0.05, | |
| "currency": 0.05, | |
| "subtotal": 0.08, | |
| "tax": 0.07, | |
| "total": 0.10, | |
| "line_items": 0.15, | |
| "payment_method": 0.05, | |
| "discount": 0.03, | |
| "shipping": 0.03, | |
| "purchase_order_number": 0.04, | |
| } | |
| _BONUS_RECONCILIATION_PASS = 0.10 | |
| _PENALTY_RECONCILIATION_FAIL = 0.15 | |
| _BONUS_DATE_VALID = 0.05 | |
| _PENALTY_DATE_INVALID = 0.05 | |
| def _field_present(extraction: DocumentExtraction, field: str) -> bool: | |
| if field == "line_items": | |
| return len(extraction.line_items) > 0 | |
| value = getattr(extraction, field, None) | |
| return value is not None and str(value).strip() != "" | |
| def _field_confidence(extraction: DocumentExtraction, field: str) -> float: | |
| """Individual field confidence score.""" | |
| if not _field_present(extraction, field): | |
| return 0.0 | |
| base = 0.80 | |
| if field in ("subtotal", "tax", "total", "discount", "shipping"): | |
| decimal_attr = f"{field}_decimal" | |
| if getattr(extraction, decimal_attr, None) is not None: | |
| base = 0.95 | |
| elif field == "document_date": | |
| if extraction.document_date: | |
| from .normalizer import normalise_date | |
| if normalise_date(extraction.document_date): | |
| base = 0.99 | |
| elif field == "line_items": | |
| items = extraction.line_items | |
| all_have_total = all(item.line_total_decimal is not None for item in items) | |
| base = 0.95 if all_have_total else 0.80 | |
| return base | |
| def calculate_confidence( | |
| extraction: DocumentExtraction, | |
| classification_confidence: float, | |
| ) -> Tuple[float, Dict[str, float], str]: | |
| """ | |
| Calculate overall and per-field confidence. | |
| Returns (overall, field_scores, explanation). | |
| """ | |
| field_scores: Dict[str, float] = {} | |
| weighted_sum = 0.0 | |
| total_weight = sum(_FIELD_WEIGHTS.values()) | |
| for field, weight in _FIELD_WEIGHTS.items(): | |
| score = _field_confidence(extraction, field) | |
| field_scores[field] = score | |
| weighted_sum += score * weight | |
| base_confidence = weighted_sum / total_weight | |
| base_confidence = base_confidence * 0.8 + classification_confidence * 0.2 | |
| total_check = next( | |
| (c for c in extraction.validation.checks if c.name == "total_reconciliation"), | |
| None, | |
| ) | |
| if total_check is not None: | |
| if total_check.passed: | |
| base_confidence += _BONUS_RECONCILIATION_PASS | |
| else: | |
| base_confidence -= _PENALTY_RECONCILIATION_FAIL | |
| date_check = next( | |
| (c for c in extraction.validation.checks if c.name == "date_validation"), | |
| None, | |
| ) | |
| if date_check is not None: | |
| if date_check.passed: | |
| base_confidence += _BONUS_DATE_VALID | |
| else: | |
| base_confidence -= _PENALTY_DATE_INVALID | |
| overall = max(0.0, min(1.0, base_confidence)) | |
| fields_extracted = len([f for f in _FIELD_WEIGHTS if _field_present(extraction, f)]) | |
| explanation = ( | |
| f"Confidence calculated from: field completeness ({fields_extracted}/" | |
| f"{len(_FIELD_WEIGHTS)} fields extracted), classification confidence " | |
| f"({classification_confidence:.0%}), and deterministic validation results." | |
| ) | |
| return overall, field_scores, explanation | |