A new method replaces associative dense-vector similarity with a causal graph based on Reichenbach's common cause principle to improve RAG. By treating retrieved documents as colliders, the model uses an attention mechanism to distinguish between shared keywords and actual causal relevance to the query.
HOW THIS AFFECTS YOU
●
builderYou can potentially improve RAG precision without retraining by applying causal-based reranking to filter out keyword-only matches.
●
researcherThis introduces a training-free way to move RAG from associative similarity to causal grounding.