SAG: SQL-Retrieval Augmented Generation with Query-Time Dynamic Hyperedges

What the paper does

SAG (SQL-Retrieval Augmented Generation) is a structured RAG architecture that indexes each document chunk as a semantically complete event paired with its entities—forming a latent hyperedge that preserves n-ary relations without decomposing them into triples. At query time it treats shared entities as join keys, dynamically assembling a query-scoped neighborhood while keeping original chunks as evidence.

Key findings

• Dense RAG struggles with structured constraints and multi-hop reasoning; offline knowledge graphs help but fragment semantics and are costly to maintain.
• SAG avoids a global graph: event-entity indexing plus query-time joins yield multi-hop connectivity on demand.
• On HotpotQA, 2WikiMultiHopQA, and MuSiQue, SAG reports best retrieval and end-to-end QA, with gains widening as reasoning-chain complexity increases.

Why it matters

Offers a middle path between dense vector RAG and heavyweight KG construction—useful for agent memory and wiki-style retrieval where relations matter but graph ops are expensive.

Related

• mkg-rag-leveraging-multimodal-knowledge-graphs-in-retrieval-
• retrieval-augmented-code-generation-a-survey-with-focus-on-r

Source: https://arxiv.org/abs/2606.15971 | 2026-06-14