You're building an LLM-powered feature and notice that the model sometimes gives confident but incorrect answers. You need a practical plan to reduce hallucinations before launch.
What is hallucination in the context of LLMs, and what are 3 concrete techniques you would use to reduce it?