Vector Search As Nearest Neighbor Matching: RAG-based Policy Learning in Causal Inference