Design AtlasNot every answer a model gives carries the same certainty. An answer directly quoting a source document and one guessed from a pattern with no grounding differ in reliability, yet both render as equally confident-sounding sentences on screen. This indicator makes that difference visible so a low-confidence answer isn't taken at face value.
A category like high/medium/low, or a short bar, is safer than a raw percentage ("87% confident") — a model's internal probability and its actual chance of being right are different things, and showing a number invites people to read it as precise statistics it isn't.
When confidence is low, don't stop at the badge — pairing it with why ("this topic has little recent information") or a next step (double-check, view sources) turns the indicator from a source of anxiety into something actually useful.
When to use
Use it in high-stakes domains — medical, legal, financial — or for answers grounded in retrieved documents. If the confidence estimate itself is unreliable (models are often bad at knowing what they don't know), it's better left out than bolted on.