MatRAG: Hierarchical RAG Using Matryoshka Representation Learning

MatRAG reduces multi-hop QA costs by aligning a Directed Acyclic Graph of document clusters with nested Matryoshka dimensions. This enables efficient top-down traversal, lowering indexing and query-time computational overhead compared to traditional Knowledge Graph or LLM-driven approaches.