Confidence, citations and evidence
Every answer carries a confidence score and the sources it drew on. This article explains what the score means, how to read citations, when an agent declines, and what you can change to get better answers.
The confidence badge
Each answer in chat shows a badge with a three-bar meter:
High: strong source agreement; citations cover the whole answer.
Medium: sources mostly agree, with some gaps.
Low: limited or conflicting sources.
No sources: nothing scored above the confidence floor. Treat the answer as unverified.
Click the badge to open What went into this score. It shows the percentage and what went into it: Source quality, Retrieval, Agreement, Freshness and Completeness. It also notes overrides ("Pinned high by a reviewer") and conflicts, such as "Sources disagreed; the verified one was used."
Post-answer checks can lower the score. A problem found after the answer is written reduces it and can cap the band at Medium or Low.
In Flexible mode, an answer with no sources shows "general knowledge" instead of a badge.
Citations and excerpts
The Sources pill at the end of the row under an answer shows how many sources it drew on. Open it for the Sources panel, which lists every source with its type (Document, Note, Website, PDF, Structured, Web, Business fact and so on), a trust badge, and an open or original source link.
Inline citation pills sit in the text. Hover to preview the cited excerpt and page through sources. You can turn these off in Account settings → Preferences with Show inline citations.
Cited images and figures appear in a strip labeled "From the cited sources".
In the app, each source also shows its share of the answer's confidence. Website widget visitors don't see this.
The panel's footer reminds you that only sources the agent can read were searched.
How the agent uses evidence
Three rules shape every answer:
Irrelevant evidence is ignored silently. The agent never lists or describes retrieved material that does not bear on the question, and never volunteers figures for things you did not ask about.
Record values beat prose. Prices, availability, stock and other attribute values are quoted only from structured record facts. A number that appears in document prose is not used as a price, and a similar product's value is never offered in place of the one you asked about.
Identifiers carry forward. Part numbers, order numbers and similar identifiers already established in the conversation are looked up exactly on follow-up questions.
When an agent declines
The agent's How it answers setting decides what happens when sources do not cover a question:
Strict answers only from your sources and declines when they don't cover it.
Balanced is grounded first; clearly marked general knowledge fills small gaps.
Flexible blends your sources with the model's general knowledge.
A decline is a plain sentence, for example "I couldn't find that in our sources. You can reach the team at…" followed by your contact details on customer channels, or "I don't have enough information in the organization's sources to answer this confidently" internally. Answer rules on the Boundaries step add other declines, such as off-topic questions or unverified promotions.
When a hand-off is suggested
Confidence thresholds are channel-specific:
Website widget: the Low confidence scenario fires when confidence drops below its Sensitivity: Low (below 40%), Medium (below 60%, the default) or High (below 75%). The visitor sees "Not quite what you needed? Talk to a human", and the scenario's action runs. A declined answer usually triggers the Outside policy rules scenario instead; a clarifying question, or a reply that explains something the agent can't do there, does not. See Scenarios and handoff rules.
Email inboxes on Auto-send: Minimum confidence to auto-send (40 to 95, default 75). Below it, or with no grounded sources, the reply is left as a draft for a person.
Internal chat shows the badge, and Slack answers end with a numbered sources list. Neither hands off automatically.
In the agent test sandbox, "Would hand off" shows when an answer would have triggered a suggestion.
Improving answers
Set trust levels. Each source and item has a Trust level: Low ("Unvetted or informal"), Medium ("Normal knowledge") or High ("The official answer"). Reviewing, approving or verifying an item counts as trust too. Confidence is capped by what backs an answer: sources that nobody has reviewed or marked as the official answer top out at Medium, a single trusted source can reach High, and two or more trusted sources, or a structured record, raise the cap further.
Fix source quality. Thumbs down on an answer ("Wrong or unhelpful — flags a knowledge gap") asks what was wrong: Incorrect, Outdated, or Missing source. Add the missing document, update the outdated one, or mark the correct source as the official answer.
Tune the scenario. On the widget, raise the Low confidence sensitivity to hand off earlier, or lower it if too many good answers are being second-guessed.
Narrow the scope. An agent that reads only the right sources agrees with itself more often. See Knowledge scope and audiences.