We Measured Our Chatbot's Hallucination Rate. Here's the Real Number, and the One Answer It Got Wrong
We ran a hallucination evaluation on our own knowledge base: 65 questions, 1 unsupported claim in 49 answered (2.0%), and 15 of 15 should-refuse questions correctly refused. Here's the method, the follow-up question it got wrong, and the caveats that come with a sample this size.



















