Scholarlink

With the rapid advancement of intelligent retrieval technologies, Retrieval-Augmented Generation (RAG) question-answering systems have emerged as a mainstream solution. However, conventional RAG systems are constrained by the inherent limitations of localized knowledge bases, resulting in suboptimal knowledge timeliness. This shortcoming may compromise the reliability of answers generated by large language models (LLMs), potentially leading to misleading outputs. To address this challenge, we propose a self-correcting algorithm named Ir-Web-RAG, which leverages multi-round web retrieval. The algorithm first reconstructs the knowledge base architecture using a sentence-level RAG optimization strategy, where a fine-tuned similarity label generation model precisely matches queries with relevant knowledge snippets. Subsequently, a dynamic web retrieval routing mechanism is designed to adaptively determine retrieval rounds based on similarity labels, substantially improving the efficiency of online knowledge acquisition. Furthermore, the algorithm employs real-time knowledge retrieved from the web to perform targeted corrections on outdated domain-specific content in the local knowledge base. The refined high-quality knowledge is then integrated as input for LLMs to enhance answer accuracy and timeliness. Experiments were conducted on three baseline models (Qwen2.5: 7B, Qwen2.5: 14B and ChatGPT 3.5-Turbo) using three public benchmark datasets (SQuAD2.0, HotpotQA and PopQA) as well as specialized Chinese datasets in finance, legal, and medical domains. Results demonstrate that Ir-Web-RAG improves query-knowledge similarity by 20%–25%, and its integration with RAG systems significantly enhances the overall performance of intelligent QA systems.

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

16Ir-Web-RAG: Interleaving Web Retrieval with Chain-of-Thought Reasoning for Retrieval-Augmented Generation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Long Jiao,
  • Hongyong Leng,
  • Xuanchen Liu,
  • Lei Pan,
  • Fuyuan Tian

摘要

With the rapid advancement of intelligent retrieval technologies, Retrieval-Augmented Generation (RAG) question-answering systems have emerged as a mainstream solution. However, conventional RAG systems are constrained by the inherent limitations of localized knowledge bases, resulting in suboptimal knowledge timeliness. This shortcoming may compromise the reliability of answers generated by large language models (LLMs), potentially leading to misleading outputs. To address this challenge, we propose a self-correcting algorithm named Ir-Web-RAG, which leverages multi-round web retrieval. The algorithm first reconstructs the knowledge base architecture using a sentence-level RAG optimization strategy, where a fine-tuned similarity label generation model precisely matches queries with relevant knowledge snippets. Subsequently, a dynamic web retrieval routing mechanism is designed to adaptively determine retrieval rounds based on similarity labels, substantially improving the efficiency of online knowledge acquisition. Furthermore, the algorithm employs real-time knowledge retrieved from the web to perform targeted corrections on outdated domain-specific content in the local knowledge base. The refined high-quality knowledge is then integrated as input for LLMs to enhance answer accuracy and timeliness. Experiments were conducted on three baseline models (Qwen2.5: 7B, Qwen2.5: 14B and ChatGPT 3.5-Turbo) using three public benchmark datasets (SQuAD2.0, HotpotQA and PopQA) as well as specialized Chinese datasets in finance, legal, and medical domains. Results demonstrate that Ir-Web-RAG improves query-knowledge similarity by 20%–25%, and its integration with RAG systems significantly enhances the overall performance of intelligent QA systems.