Question
EY
ES
Last activity: 2 Oct 2025 22:55 EDT
Chatbot fallback behavior with Prediction Studio NLP – show low-confidence topics as suggestions
I am implementing a chatbot channel integrated with a text prediction model in Prediction Studio. My requirement is:
When the user enters a free-text question and the best predicted topic does not reach the defined confidence threshold, we want to trigger the fallback response.
In that fallback message, we would like to either:
Reuse the default Pega fallback response configured in the channel, or Create a new topic/intent dedicated to handling fallback cases.
Additionally, under the fallback message, we want to dynamically display the top N topics returned by the model (with the highest confidence, even if below the threshold) as suggested options for the user to choose from.
Question: Is there a product-supported way in Pega to:
Invoke the default fallback when the confidence threshold is not met, and at the same time expose the top predicted topics (with their scores) as suggested quick replies?
Or do I need to implement a custom flow/topic to handle this scenario?
Environment: Prediction Studio NLP model + Chatbot Channel.
HCA Healthcare
US
@DIEGO García López
Yes, use the channel’s confidence threshold to fire the built-in fallback, then add a tiny customization to surface the top N topics as quick replies. Set Minimum confidence in the Chatbot channel so low-confidence inputs hit the default fallback response. Create a lightweight “Fallback” topic (or use the channel fallback) that runs a Data Transform. In that DT, read the Prediction Studio output (the classifications list with confidence scores), sort by confidence, take top N, and build a suggestions list (labels like TopicName — 62%). Return a “Send message” response that shows your normal fallback text plus those suggestions as quick-reply buttons. When a user taps a suggestion, transition to that topic (or pass the predicted topic name as a parameter). Net: product-supported fallback is used as-is; you only add the DT that maps model scores to suggested replies.