The honest fallback
If a user asks a question that has no reliable match in your knowledge base, 0xDoktor does not invent an answer. It returns a configurable fallback message such as:
"I don't have a reliable answer for that in my knowledge base. You can contact our support team at support@example.com."
This behavior is intentional and is what makes 0xDoktor safe to deploy for support use cases. A chatbot that never hallucinates is more valuable than one that always sounds confident.
How confidence scoring works
Every answer 0xDoktor generates includes a confidence score between 0 and 1. The score reflects how well the retrieved documents actually match the user's question.
- High confidence (≥ 0.75) — the answer is well-supported by your content; shown normally
- Medium confidence (0.4–0.74) — the answer is shown with a disclaimer encouraging the user to verify
- Low confidence (< 0.4) — the fallback message is shown and the exchange is flagged for review
You can adjust these thresholds in your admin dashboard under Settings → Quality.
Human escalation
When confidence is low, 0xDoktor can show a human escalation prompt — a message that invites the user to contact a human agent, submit a ticket, or send an email. You configure the escalation copy and the contact link in the admin dashboard.
If you have a live chat system (Intercom, Crisp, HubSpot, etc.), the escalation message can include a link that opens that chat directly in the same browser session.
Using low-confidence questions to improve your KB
Every low-confidence exchange is recorded in your analytics dashboard (query-hashed, never storing the raw question). Clusters of unanswered questions reveal gaps in your knowledge base.
Go to Analytics → Unanswered Questions to see which topics users are asking about that you have not yet documented. Adding content for those topics is the most effective way to improve your chatbot's answer rate.
The wasResolved signal
After each answer, users can give a thumbs-up or thumbs-down. This feedback sets the wasResolved flag on the conversation. Over time, wasResolved rates per topic give you a ground-truth measure of chatbot effectiveness — more reliable than confidence scores alone.