OmiG, this makes perfect sense and is, in fact, a known and useful technique. I also agree that the "caricature" of RAG (single vector retrieval) can and should be augmented with a variety of approaches. Including the one you are suggesting. In general, this falls into the pattern of chunk expansion. In indexing time, for every vector in the DB, you save an explicit set of related chunks. And, at retrieval time, whenever you get a match, you also explicitly fetch all the related chunks by id.  Related chunks can be anything you want. Examples could be three chunks before and after your chunk in the text, the document's title/abstract/summary, the file name, definitions of entities in the text, you name it. If I understand your approach, you are suggesting linking each chunk with a "causal chain" which sounds very promising.