The danger is not that the model is confused. The danger is that it can be wrong in a polished voice. Teach A hallucination is generated content that is unsupported by the input or wrong against the facts. The problem is not limited to casual chat. Any system that generates fluent language can produce a statement that sounds reasonable while lacking evidence. RAG helps because it changes the input. Instead of asking the model to answer from memory, you provide retrieved source passages from a defined collection. The model can now anchor the answer in the passages. But two things…
Read the full lesson
Sign up free: one personalized lesson every day, matched to your role and goals.
Already have an account? Sign in