Retrieval-Augmented Generation (RAG) Survey: From Naive RAG to Modular RAG Evolution and Evaluation: Key Technology for Knowledge Enhancement in Large Language Models
This survey systematically reviews the evolution of retrieval-augmented generation (RAG) from naive RAG, advanced RAG to modular RAG, analyzes in detail the technical details of the three core components of retrieval, generation, and augmentation, and introduces the latest evaluation frameworks and benchmarks. By introducing external knowledge bases, RAG effectively alleviates issues such as hallucinations, outdated knowledge, and opaque reasoning in large language models, improving the accuracy and credibility of knowledge-intensive tasks.
This paper comprehensively reviews the development history of RAG technology, dividing the paradigm into three stages: naive RAG, advanced RAG, and modular RAG. It deeply analyzes the key technologies of retrieval, generation, and augmentation, and provides evaluation frameworks and benchmarks. By integrating the intrinsic knowledge of LLMs with external dynamic databases, RAG significantly improves the accuracy and traceability of generated content, providing a reliable solution for knowledge-intensive applications.
The RAG framework consists of three parts: retrieval, generation, and augmentation. Naive RAG adopts a simple retrieve-read process; advanced RAG introduces pre-retrieval, post-retrieval optimization, and multiple retrieval strategies; modular RAG supports pluggable components and process orchestration. In terms of evaluation, the paper introduces an evaluation framework covering dimensions such as answer accuracy, faithfulness, and relevance, as well as benchmarks like RGB and RECALL. The boundaries are that retrieval quality, computational overhead, and long-context processing remain challenges.
RAG technology has been widely applied in fields such as question-answering systems, conversational AI, and enterprise knowledge management, especially suitable for scenarios requiring real-time knowledge updates or handling private data. This survey provides a technical roadmap for the industry to choose RAG architectures, helping to reduce deployment costs and improve system reliability.
It is recommended that enterprises prioritize adopting modular RAG architectures, combining vector databases (e.g., Pinecone, Weaviate) and LLM APIs (e.g., GPT-4) to build knowledge-enhanced applications. They can refer to the evaluation framework in the paper to establish internal evaluation systems and focus on retrieval optimization techniques (e.g., HyDE, re-ranking) to improve effectiveness.
Future attention should be paid to the balance between retrieval efficiency and accuracy, the expansion of multimodal RAG, and the unification of evaluation standards. Specific verification signals include performance improvements on benchmarks such as KILT and Natural Questions, as well as real-world deployment cases of industrial-grade RAG systems.