This survey provides a comprehensive overview of Retrieval-Augmented Large Language Models (RA-LLMs), which integrate external knowledge to enhance the performance of large language models (LLMs). The survey covers three primary technical perspectives: architectures, training strategies, and applications. It begins with an introduction to LLMs and prompt learning, highlighting their capabilities and limitations. The core of the survey focuses on the retrieval process, including retriever types, retrieval granularity, pre- and post-retrieval enhancement, and database construction. It also discusses the generation process, detailing parameter-accessible and parameter-inaccessible generators, and various integration methods for retrieval and generation. The necessity and frequency of retrieval are explored, emphasizing the importance of selective incorporation of retrieved information. Finally, the survey reviews training methods for RA-LLMs, categorizing them into training-free and training-based approaches, and further dividing training-based methods into independent, sequential, and joint training. The survey aims to provide a systematic review of recent advances in RA-LLMs, addressing their technical foundations, practical applications, and future research directions.This survey provides a comprehensive overview of Retrieval-Augmented Large Language Models (RA-LLMs), which integrate external knowledge to enhance the performance of large language models (LLMs). The survey covers three primary technical perspectives: architectures, training strategies, and applications. It begins with an introduction to LLMs and prompt learning, highlighting their capabilities and limitations. The core of the survey focuses on the retrieval process, including retriever types, retrieval granularity, pre- and post-retrieval enhancement, and database construction. It also discusses the generation process, detailing parameter-accessible and parameter-inaccessible generators, and various integration methods for retrieval and generation. The necessity and frequency of retrieval are explored, emphasizing the importance of selective incorporation of retrieved information. Finally, the survey reviews training methods for RA-LLMs, categorizing them into training-free and training-based approaches, and further dividing training-based methods into independent, sequential, and joint training. The survey aims to provide a systematic review of recent advances in RA-LLMs, addressing their technical foundations, practical applications, and future research directions.