[JPLeRouzic]

So is it an improved RAG, where RAG information is analyzed and compressed before being introduced in the LLM?

If so, the context window is still a problem, no?