LanguagesServicesIntentsSynonymsCTAsTemplatesLead FieldsPick ListsThresholdsFast Path TesterBulk Import

Thresholds & Policy

Install-scoped tunables for matchers, evidence gate, no-answer policy, agent handoff.

⚠
Changes apply to live traffic
Existing in-flight sessions are unaffected; new sessions and turns use the new thresholds. Saved changes appear in the audit log.
A

Matcher thresholds

Lower = looser matching (more catches, more false positives). Higher = stricter.
T_service
Service+Synonym matcher confidence.
0.82
T_intent
Intent matcher confidence (after embedding rerank).
0.86
T_faq
FAQ retrieval similarity.
0.88
T_evidence
Top-K chunk score after reranker. Below this → no-answer.
0.55
B

No-answer policy

Fires when RAG retrieval scored below T_evidence — "I asked the bot something, but the bot has no approved knowledge for it."
No-answer text per language*
Shown when evidence gate fails.
Show callback CTA
Lead capture variant: name + phone + preferred time.
Show handoff CTA
Connect to live agent if available.
C

Safe fallback

Fires when Knowledge Question Check classifies the visitor message as "not a question" — e.g. greetings, small talk, off-topic. Different from no-answer: visitor didn't ask a question, so we soft-suggest topics instead of escalating.
Safe fallback text per language*
Shown when visitor message isn't a question. Tone should be welcoming, not apologetic.
Suggestion chips (en)
3–6 quick-pick suggestions shown under the fallback message. Comma-separated.
D

Agent policy

Agent idle timeout (seconds)
Auto-away after this much inactivity.
Preview widget