How can one's knowledge be selectively exposed through a conversational interface?

If one is wrapping a conversational interface around their knowledge base, but they also store highly sensitive information in there, then how can they selectively share the non-sensitive creative output? A low-tech approach would be to use keywords for restricting outputs. Additionally, semantic distance to such keywords could be used as an indicator of sensitivity. However, natural language inference can be used to evaluate whether a given response contains sensitive information, similar to descriptive search.