Scholarlink

To address the challenge of acquiring specialized knowledge in the field of ceramics, this study proposes an intelligent question-answering system based on the Retrieval-Augmented Generation (RAG) framework. First, a high-quality ceramic domain knowledge base was constructed by collecting and preprocessing domain-related data from multiple sources, using a recursive chunking strategy and text vectorization techniques. Then, the system employs a hybrid retrieval strategy that integrates BM25 keyword indexing with FAISS vector indexing, and leverages the Llama-Index and LangChain frameworks to enable efficient back-end retrieval and response generation. To evaluate the system’s performance, the RAGAs toolkit was utilized, in combination with the CRITIC-weighted TOPSIS multi-criteria decision-making method. A comparative experiment was conducted on 13 mainstream open-source LLMs that can run on consumer-grade GPUs. The results indicate that the Chinese-optimized QWQ-32B model demonstrated the best overall performance in the ceramic-domain question-answering tasks.

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

Research on the Construction of a Retrieval-Augmented Generation-Based Intelligent Question-Answering System in the Field of Ceramics

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Xiaoyan Wang,
  • Yu Huang,
  • Tao Liu,
  • Shenglan Peng

摘要

To address the challenge of acquiring specialized knowledge in the field of ceramics, this study proposes an intelligent question-answering system based on the Retrieval-Augmented Generation (RAG) framework. First, a high-quality ceramic domain knowledge base was constructed by collecting and preprocessing domain-related data from multiple sources, using a recursive chunking strategy and text vectorization techniques. Then, the system employs a hybrid retrieval strategy that integrates BM25 keyword indexing with FAISS vector indexing, and leverages the Llama-Index and LangChain frameworks to enable efficient back-end retrieval and response generation. To evaluate the system’s performance, the RAGAs toolkit was utilized, in combination with the CRITIC-weighted TOPSIS multi-criteria decision-making method. A comparative experiment was conducted on 13 mainstream open-source LLMs that can run on consumer-grade GPUs. The results indicate that the Chinese-optimized QWQ-32B model demonstrated the best overall performance in the ceramic-domain question-answering tasks.