RAG and RAU: A Survey on Retrieval-Augmented

What I found interesting was the fifth section that describes the strategies to improve RAG performance. Basically:

1. Quality control. What documents to include?

2. Timing. When to query?

3. Pre & post processing. Improve LLM outputs based on retrieved data.

4. End to end training. Expensive, and data intensive but possibly the best long-term approach.

5. Controller. An interesting idea with similarities to Reinforcement Learning.

I wonder what are your thoughts. Which one is most promising? What has been your experience when building RAG apps? Also, is RAG the leading architecture for building applications?