Skip to main content

Reference: confidence levels

Every Chat answer (and every query log entry) carries one of four confidence levels.

LevelBadgeSystem's interpretationRecommended handling
HighgreenAnswered confidently from indexed documents.Rely on it; spot-check citations when stakes are high.
MediumaccentAnswered, but with some uncertainty.Read the cited excerpts before acting.
LowamberAn answer was found but confidence is low — review recommended.Treat as a pointer; verify in the source document.
Insufficient DataredCould not answer from the available documents.No answer is fabricated. Follow the suggested next steps; the question is logged in Gap analysis.

What drives confidence

Confidence reflects how well retrieved passages actually support the answer — how relevant the matched chunks were and how completely they cover the question. It is not a measure of how fluent the answer sounds; it's a measure of evidence.

Two related notices can accompany an answer:

  • Degraded search — retrieval ran in a reduced mode; consider re-asking.
  • Unverified sources — some model references couldn't be matched to real passages and are flagged rather than cited.

Improving confidence

Low confidence is nearly always a content problem, not a phrasing problem:

  1. Ask more specifically (helps retrieval find the right passage).
  2. Ensure the covering document exists, is current, and is INDEXED.
  3. Watch Gap analysis for recurring weak spots and upload what's missing.

High confidence with citations