Does RAG Need Better Retrieval — or Better Relationships?

A journalist explored whether Retrieval-Augmented Generation (RAG) systems benefit more from improved retrieval or better relationships between documents. To isolate the impact of relationships, the researcher built an experiment with synthetic business documents and manually defined ground truth relationships, such as supersession, contradiction, and linkage, between documents.

The baseline retrieval was weak, with only 67% recall of oracle documents in the top five. When relationships were added to the retrieved documents, accuracy improved significantly, from 77% to 98% for both DeepSeek and GPT-4o-mini models, a 17-21 percentage point gain. This suggested that relationships contribute substantially to RAG performance.

However, the researcher then questioned what aspects of the relationships were most impactful. They found that the presentation of relationships mattered, as identical relational information presented in structured fields versus ordinary prose resulted in the same performance boost. This implied that how relationships are conveyed may be as crucial as the relationships themselves.

Interestingly, introducing incorrect relationships led to a sharp decline in accuracy, from 98% to 53% when 20% of relationship fields were corrupted, and further to 52% at 40% corruption. This indicated that the integrity of relationships is vital for RAG systems to leverage relational context effectively.