📄 中文摘要
检索增强生成（RAG）被广泛宣传为解决人工智能幻觉的良方，通过引入搜索和向量数据库，模型能够减少虚假信息的生成。然而，现实情况更加复杂。大型语言模型（LLMs）在表述上常常显得自信，但却可能包含错误，尤其是在生成引用、网址和引述时，部分模型的虚假引用率高达18%至69%。RAG通过强制模型基于检索到的文档进行回答，而非单纯依赖记忆，从而显著降低了虚假引用的发生率，尤其在医学内容中效果明显。
检索增强生成如何有效防止人工智能幻觉
📄 English Summary
How Retrieval Augmented Generation Actually Prevents Ai Hallucinations
Retrieval Augmented Generation (RAG) is often marketed as a solution to AI hallucinations, claiming that by integrating search and a vector database, models can stop generating falsehoods. However, the reality is more nuanced. Large Language Models (LLMs) can sound convincingly accurate while being incorrect, fabricating citations, URLs, and quotes with high confidence. In some scenarios, the rate of generated false citations among popular models ranges from 18% to 69%, with medical content particularly affected. RAG addresses this issue by compelling models to base their responses on retrieved documents rather than relying solely on memory, which significantly reduces the incidence of ghost references.