The best fit is Reinforcement learning that uses customer ratings as reward signals.

Reinforcement learning that uses customer ratings as reward signals optimizes a decision policy by maximizing cumulative reward so thumbs up and thumbs down directly guide the chatbot to prefer responses that earn positive feedback. New product guides and community posts can be included as additional context or state features and the policy can learn when to consult or cite those sources during conversations. Over time the agent adapts based on real user reactions rather than relying only on static labeled examples.

Supervised learning with a labeled dataset of high-quality and low-quality responses is useful for initial training because it teaches mapping from inputs to desired outputs. It does not, however, provide a built-in mechanism to continuously learn from live thumbs up and thumbs down unless you add a separate feedback loop.

Unsupervised learning to cluster similar support questions helps organize and discover patterns in incoming queries and it can support routing or template creation. Clustering does not choose or optimize actions based on customer ratings so it is not sufficient for continuous improvement driven by feedback.

Supervised learning that retrains when the FAQ and knowledge articles are updated can incorporate new content into models during periodic retraining and that improves coverage after updates. It still lacks an intrinsic, reward-driven update path from each conversation so it will not continuously improve from per-chat ratings in the same way reinforcement learning does.

Full AWS Practitioner Certification Question