Split your documents into focused chunks and store them in a vector database for rag, so retrieval returns the right page instead of a average of everything.