Currently, Eddy's AI workflow sends all user queries through the Moderation API to detect harmful or sensitive content before generating responses. However, certain generic yet legitimate queries like "How to unlock a user" are incorrectly flagged as illicit by the Moderation API due to their ambiguity. These are not passed to the answer generation model, resulting in no results found, despite relevant content being available in the knowledge base.

This creates confusion for end users and prevents them from receiving accurate responses. Please review these scenarios so we can implement a better approach to handling such false positives.