What strategies would you use to mitigate hallucinations in a RAG (Retrieval-Augmented Generation) pipeline at Google DeepMind?
