RAG overcomes the limitations of training data cutoff dates and reduces the probability of hallucinations by having language models first retrieve relevant content from an external knowledge base before generating answers.