Scholarlink

<p>The proliferation of unstructured data has exposed the limitations of traditional retrieval systems in addressing complex queries, multi-hop reasoning, and context-aware synthesis. This study evaluates four Retrieval-Augmented Generation (RAG) systems–Naive RAG, Memory-Augmented Retrieval (MAR), Recursive Abstractive Processing for Tree-Organized Retrieval (RAPTOR), and a novel Hybrid Model–using the BioASQ biomedical dataset. MAR leverages memory persistence to excel in static queries but struggles with dynamic adaptability. RAPTOR employs hierarchical abstraction to enhance multi-hop reasoning but lacks persistent memory capabilities. The Hybrid Model integrates MAR’s memory-driven retrieval with RAPTOR’s hierarchical reasoning to balance performance across query types, achieving superior results in both static and dynamic tasks. Key findings emphasize the critical role of embedding strategies in retrieval performance and highlight avenues for advancing query refinement, adaptive retrieval techniques, and multimodal integration. These innovations aim to address scalability and contextual understanding challenges, positioning the Hybrid Model as a robust framework for context-aware, scalable information retrieval.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Optimizing information retrieval: a hybrid model leveraging MAR and RAPTOR frameworks

  • 期刊论文

  • 发表日期：

    2026年3月24日

查看全文
  • Timur Abdygulov,
  • Meet Daxini,
  • Amir H. Jafari

摘要

The proliferation of unstructured data has exposed the limitations of traditional retrieval systems in addressing complex queries, multi-hop reasoning, and context-aware synthesis. This study evaluates four Retrieval-Augmented Generation (RAG) systems–Naive RAG, Memory-Augmented Retrieval (MAR), Recursive Abstractive Processing for Tree-Organized Retrieval (RAPTOR), and a novel Hybrid Model–using the BioASQ biomedical dataset. MAR leverages memory persistence to excel in static queries but struggles with dynamic adaptability. RAPTOR employs hierarchical abstraction to enhance multi-hop reasoning but lacks persistent memory capabilities. The Hybrid Model integrates MAR’s memory-driven retrieval with RAPTOR’s hierarchical reasoning to balance performance across query types, achieving superior results in both static and dynamic tasks. Key findings emphasize the critical role of embedding strategies in retrieval performance and highlight avenues for advancing query refinement, adaptive retrieval techniques, and multimodal integration. These innovations aim to address scalability and contextual understanding challenges, positioning the Hybrid Model as a robust framework for context-aware, scalable information retrieval.