Visibility: Public
What is a Confidence Score?
Each AI‑generated response is assigned a confidence score that indicates the system’s assessment of its reliability. The scoring pipeline evaluates multiple quality signals and applies structured deductions to a baseline score of High.
Major Signals That Impact the Confidence Score
The quality signals that are system evaluates comprises of the following categories:
- Citation Coverage & Precision
- Detects presence/absence of citations.
- Tracks citation density (percentage of sentences containing references).
- Measures context precision (alignment of response text with provided source context).
- Content Integrity Checks
- PII Detection: Screens for accidental disclosure of personally identifiable information.
- Tone Analysis: Flags negative, risky, or unsupported tonal patterns.
- Language Quality: Identifies non‑committal phrasing (e.g., “it depends,” “not available”) indicating reduced reliability.
- Knowledge Coverage
- If requested information is not retrievable within the provided materials, the system surfaces a structured placeholder:
INFORMATION MISSING IN KNOWLEDGE HUB (highlighted in yellow for visibility).
Scoring Levels & Criteria
- High: Source‑backed, precise, no integrity violations.
- Medium: Generally reliable but at least one cautionary signal (e.g., PII risk, vague language).
- Low: Insufficient support (few/no citations, weak alignment with source context). Recommended for human review before use.
Key Notes
- The scoring mechanic applies deductions only—scores always initialize at high and degrade based on signal strength.
- All scores are advisory heuristics, not authoritative truth judgments. Human review of both the response and source materials remains required in high‑stakes contexts.