SAG: SQL-Retrieval Augmented Generation with Query-Time Dynamic Hyperedges
SAG tries to get multi-hop RAG behavior from SQL-style joins over entities, without maintaining a full knowledge graph.
The paper indexes document chunks as complete events paired with their entities, treating each chunk as a latent hyperedge. At query time, shared entities become join keys that link relevant chunks into a query-scoped evidence neighborhood. The authors report top retrieval and end-to-end QA results across HotpotQA, 2WikiMultiHopQA, and MuSiQue, with MuSiQue Recall@5 at 80.36%, 11.52 points above the strongest baseline. ArXiv · AI/CL/LG's note
The paper indexes document chunks as complete events paired with their entities, treating each chunk as a latent hyperedge. At query time, shared entities become join keys that link relevant chunks into a query-scoped evidence neighborhood. The authors report top retrieval and end-to-end QA results across HotpotQA, 2WikiMultiHopQA, and MuSiQue, with MuSiQue Recall@5 at 80.36%, 11.52 points above the strongest baseline. ArXiv · AI/CL/LG's note
score 4