RAG (Retrieval-Augmented Generation)

检索增强生成。在推理时从外部知识库检索相关文档喂给大模型，从而把生成内容"锚定"在可验证、最新、可追溯的信息上。

English Definition

"A technique that augments an LLM's generation by retrieving relevant external documents at inference time, grounding the output in verifiable, up-to-date knowledge sources."

用法说明 · 针对中文母语者

RAG 不是万能解：它能显著降低幻觉但无法完全消除；论文中常需配合 faithfulness 等指标一起评估。

常见错误

·把 RAG 当作单纯的向量数据库检索，忽略 retrieval quality 本身也会带来新偏差
·以为 RAG 可以让模型"记住"所有知识——它本质上是即时检索，不改变模型参数

真实用例 · 4 条

"employs a large language model (LLM) with retrieval-augmented generation (RAG) to translate clinical assessments into an understandable interface"

论文Pipeline extracted — review needed·系统架构描述

"Retrieval-augmented generation (RAG) mitigates this challenge by connecting LLMs to external knowledge sources"

论文Pipeline extracted — review needed·技术原理

"while RAG significantly reduces hallucinations, it does not eliminate them entirely"

论文Pipeline extracted — review needed·指出 RAG 的局限性

"reliance on RAG still entails residual risks. Research shows that LLMs may ignore or distort retrieved context, particularly when key information is buried within long inputs"

论文Pipeline extracted — review needed·RAG 失败模式

由 pipeline 自动采集，待人工 review