KB Draft: How Yada AI Learns from Corrections

Draft status: KB draft created from the Refuge Bay onboarding questions. Review with Product/CS before publishing.

How Yada uses trained listing content, accuracy thresholds, human pauses, and retraining to prevent repeated incorrect AI responses.

How AI Answers Are Generated

Yada answers guest questions from the listing training set, listing metadata, guidebook/content cards, and any listing knowledge base text that has been indexed for the property.

The AI is expected to answer only the guest question, stay concise, and avoid making commitments that are not supported by the available context.

What Happens When an Answer Is Wrong

A human correction in the Inbox does not automatically rewrite the trained knowledge base by itself. The durable fix is to correct the underlying listing content, guidebook card, knowledge base text, or operating rule, then retrain the listing so future answers retrieve the corrected information.

Yada logs AI outputs and citations around autopilot sends, so the team can inspect what context was used and identify the source that needs correction.

Controls That Reduce Repeat Mistakes

• Accuracy threshold: higher settings block responses that include information outside known content.

• Approval controls: low-confidence or low-quality responses can be held for review before sending.

• Human intervention: if a human replies after an AI response is scheduled, Yada cancels the automated send.

• AI pause: a conversation can be paused for 15 minutes, 1 hour, 1 day, or indefinitely.

• Backoff time: after a human gets involved, Yada can wait before allowing AI to respond again.

Customization and Bot Names

Yada currently supports brand voice, tone, do-not-say rules, and signatures for outbound messaging. A separate named-bot persona such as “Larry” is not a first-class product setting in the current platform.

A practical customer-facing approach is to say that the AI assistant is still being trained, then provide the corrected answer and update the trained content so the same issue is less likely to repeat.