Understanding Confidence Scores
Claro calculates confidence using four key factors: 1. Source Reliability (40% weight) - How trustworthy are the information sources?- First-party documents and authoritative sites score higher
- Unknown or low-trust sources reduce confidence
- Knowledge Base content typically scores higher than web sources
- Consistent information across sources increases confidence
- Conflicting data from different sources lowers the score
- Single-source answers receive moderate confidence
- Low response variance indicates higher model certainty
- High variance suggests the model was uncertain
- Based on internal probability distributions
- Strong semantic match between query and sources
- Comprehensive coverage of the question topic
- Clear, well-structured source content
Confidence Badges
| Badge | Score Range | Color | What It Means |
|---|---|---|---|
| High | 85–100 | 🟢 Green | Highly reliable, ready to use |
| Medium | 60–84 | 🟡 Yellow | Generally good, worth spot-checking |
| Low | 0–59 | 🔴 Red | Needs human review or re-processing |
Working with Citations
Viewing Citations:- Hover over confidence badges for a preview
- Click badges to see full source list and relevant passages
- Follow links to verify information in original sources
- Knowledge Base: Page numbers and document names
- Web sources: URLs and relevant text snippets
- Geolocation: Map services and data providers
- File extraction: Document sections and page references
Optimizing Confidence Scores
To increase confidence:- Add relevant Knowledge Bases with authoritative information
- Use specific prompts that clearly define what you need
- Provide examples of the desired output format in the prompts
- Cross-reference multiple sources when possible
- Refine your prompt to be more specific
- Add relevant reference documents to the Knowledge Bases
- Re-run individual rows with adjusted parameters
- Consider manual review for critical data points
- Sort tables by confidence after large runs to prioritize review
- Set confidence thresholds in Column Settings to auto-filter results
- Use feedback loops (Dedicated plan) to train models on edge cases