In recent years, Retrieval-Augmented Generation (RAG) technology has effectively mitigated the “hallucination” phenomenon and knowledge obsolescence issues in Large Language Models (LLMs) caused by reliance on static training data through dynamic integration of external knowledge bases. This paper systematically reviews the core advancements and challenges in RAG tecnology. First, we clarify the fundamental theoretical framework of RAG, which consists of collaborative retrieval and generation modules: The retrieval module dynamically extracts relevant knowledge from heterogeneous data sources through semantic mapping, query optimization, and multi-strategy retrieval, while the generation module leverages pre-trained language models combined with re-ranking, context compression, and fusion strategies to produce high-quality text based on retrieval results. Second, we conduct in-depth analyses of critical technical challenges and solutions in the retrieval module, including data chunking techniques, query optimization methods, and retrieval strategies. Concurrently, we explore recent progress in retrieval result optimization and fusion mechanisms within the generation module. Finally, we identify current challenges in RAG technology, including insufficient retrieval-generation synergy, limited multimodal support, and privacy biases. Future research directions should focus on end-to-end joint optimization, multimodal expansion, and robustness enhancement. This paper provides theoretical references and practical guidance for systematic research and applications of RAG technology.