Study Finds Explicit Document Relationships Boost RAG Reasoning Without Better Retrieval

A developer ran a controlled experiment to test whether explicitly providing relational context between documents improves reasoning in retrieval-augmented generation (RAG) systems. A synthetic corpus of 254 business documents was built, covering billing, DevOps, and HR topics, with 60 multi-hop reasoning cases manually annotated with ground-truth relationships. Both test conditions received identical retrieved documents via a simple BM25-style retrieval baseline, with the only difference being whether the model also received structured relationship metadata such as supersession or contradiction links. The experiment found that adding explicit relational context meaningfully improved the model's ability to answer questions whose answers exist not within individual documents but in the connections between them. The findings suggest that relationship-aware context, even without retrieval improvements, can significantly enhance multi-hop reasoning in RAG pipelines.