Retrieval-augmented generation for multi-hop question answering faces a fundamental trade-off: better retrieval often means more computation. According to the paper, a significant portion of this cost is paid during indexing, when systems build expensive knowledge graphs to support later queries.
The authors present a Matryoshka-inspired hierarchical RAG design, suggesting a nested structure that organises retrieval at multiple levels of granularity. The approach appears intended to lower the reliance on costly knowledge graph indexing while still maintaining retrieval quality for complex, multi-step questions.
The abstract does not include experimental results, so the claimed efficiency gains remain to be verified. But the motivation is clear: making indexing cheaper without sacrificing retrieval accuracy is a key step toward practical multi-hop QA systems.