Confidence Score
ConceptsA confidence score indicates how reliable a result is. DocLD provides confidence for extraction field values and for retrieval passages. Confidence helps prioritize manual review and identify uncertain results. Low-confidence extractions can be corrected; corrections feed into ground truth for accuracy measurement.
Extraction Confidence
| Score | Meaning | Action |
|---|---|---|
| 0.9+ | High confidence | Usually accurate; spot check occasionally |
| 0.7–0.9 | Medium confidence | Spot check; verify against source |
| Below 0.7 | Low confidence | Manual review required |
Per-field and overall confidence help prioritize review. When a field has low confidence, check the citation to see the source passage and verify or correct the value. Corrections can be promoted to ground truth for future accuracy measurement.
Interpreting Confidence by Domain
Confidence thresholds may vary by document type and use case:
| Domain | High threshold | Notes |
|---|---|---|
| Invoices | 0.9+ | Amounts and dates are critical; require high confidence |
| Contracts | 0.85+ | Legal terms may be nuanced; medium-high acceptable |
| Resumes | 0.8+ | Name and contact info are critical; skills may vary |
| General text | 0.7+ | For retrieval passages, medium confidence often acceptable |
Adjust your review workflow based on the criticality of each field. Use schema instructions to guide the model for edge cases and improve confidence over time.
Citation Confidence
Citations in chat and extraction may include confidence (e.g., "high", "medium", "low") indicating how well the passage matched the query or supported the extracted value. In chat, citation confidence helps users assess whether an answer is well-grounded. In extraction, citation confidence shows how strongly the source passage supports each field value.
Best Practices
- Set thresholds — Define minimum confidence per field type; flag or route low-confidence results for review
- Use citations — When confidence is low, inspect the citation to understand why; refine schema instructions if patterns emerge
- Track trends — Monitor confidence over time; declining confidence may indicate new document types or schema drift
- Combine with ground truth — Use ground truth to measure accuracy; compare confidence to actual correctness
Related Concepts
Confidence scores are produced by extraction and retrieval. Citations show where values or answers came from. Ground truth uses corrections to measure extraction accuracy over time.