Comment by overgard

From what I've heard, personas give a greater chance that the LLM will answer confidently.. and also a greater chance it'll hallucinate something when the data is sparse. Supposedly "grounding" the personas on real documents/web searches is the best approach. Anecdotal though.