Improving Retrieval Accuracy with SQL Hyperedges

Summary & Methodology Analysis

The SAG framework addresses the limitations of standard dense-retrieval methods, which struggle with structured constraints and multi-hop reasoning. Instead of building a static global graph, which is difficult to maintain, SAG uses an offline indexing strategy that extracts one event and a set of entities per document chunk. These chunks form a latent hyperedge index. At query time, the system performs parallel retrieval by using entity-guided structured recall via SQL joins and direct vector embedding similarity for event recall. This approach allows the system to traverse connections between documents dynamically without the overhead of a pre-computed graph structure.

For the final response, SAG employs a dual-path selection strategy. The query-time expansion phase uses reverse SQL joins on shared entities to aggregate associated events into a candidate pool. These candidates undergo a coarse ranking based on embedding similarity before an LLM selects the top five events to provide structural evidence. This selected structural context is then merged with semantic retrieval results to populate the final context window for generation. This combination ensures that the model has access to both explicit relationships discovered through SQL and broader semantic knowledge retrieved through vector similarity.

Despite its strong performance, SAG faces specific operational limitations. The current system lacks a mechanism to resolve entity aliases, which can lead to missed logical connections between documents. Furthermore, the index is append-only, meaning there is no defined process for revising or retiring outdated information. The paper does not specify the computational latency or the memory footprint required to maintain this hyperedge index.