Research and Analysis on the Mechanism of Suppressing Large Model Hallucination Based on Modular RAG Architecture

Large Language Models (LLMs) perform remarkably well in knowledge-intensive tasks, yet they are difficult to deploy in high-stakes scenarios due to 'hallucinations'—outputs that contradict factual information. Although Retrieval-Augmented Generation (RAG) is widely regarded as a mainstream approach to mitigate hallucinations, most existing studies treat it as a black box and rarely analyze the heterogeneous functions of its internal modules. To address this, we propose a 'functionally decomposed' modular RAG taxonomy, dividing the entire process into three stages: retrieval, refinement, and generation, from which three technical pathways are derived: Direct Injection RAG (DI-RAG), Relevance-Focused RAG (RF-RAG), and Fact-Checking RAG (FC-RAG). Utilizing the large-scale real-world Q&A dataset MS MARCO, we constructed four benchmarks simulating high-risk scenarios such as information noise, knowledge conflicts, and outdated knowledge. Using Qwen1.5-7B-Chat as the generative backbone, we systematically evaluate the marginal benefits of the three architectures in suppressing hallucinations and quantify the contributions of components like re-rankers and multi-query verifiers in specific hallucination scenarios, providing actionable empirical guidance and optimization pathways for building high-fidelity RAG systems.