I think the next phase of LLM products is LLM + RAG but the retrieval system is personalized per user and behaves differently based on user preferences, social graph, etc. What are products doing this?

Permissionless protocols like farcaster and bountycaster get really interesting in this context too.

In a RAG + LLM, RAG makes the major part. Getting them right is a set of tasks various Recommender Systems have been working on more than decade. Furthermore in KGAG, KG takes the larger chunk. We're building an agent that has all the web3 (domain + apps) context and making llm frame the answer is the lightest task.