A Survey on Retrieval-Augmented Generation for Large Language Models

引言：大语言模型的机遇与挑战

Introduction: Opportunities and Challenges of Large Language Models

大语言模型（LLMs）在自然语言理解和生成任务上展现出了令人瞩目的能力，已成为人工智能领域的核心驱动力。然而，在实际部署和应用中，它们也暴露出一些固有的局限性。这些挑战主要包括：模型可能产生看似合理但不符合事实的“幻觉”内容；其知识受限于训练数据的时间点，难以获取最新信息；以及其内部推理过程如同一个“黑箱”，缺乏透明度和可追溯性。为了解决这些问题，检索增强生成（RAG）技术应运而生，它通过将外部知识库与LLMs的生成能力相结合，旨在提升模型输出的准确性、时效性和可信度。

Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language understanding and generation tasks, becoming a core driving force in the field of artificial intelligence. However, in practical deployment and application, they also reveal some inherent limitations. These challenges mainly include: the model may generate plausible but factually incorrect "hallucinations"; its knowledge is constrained by the timestamp of its training data, making it difficult to access the latest information; and its internal reasoning process resembles a "black box," lacking transparency and traceability. To address these issues, Retrieval-Augmented Generation (RAG) technology has emerged. It aims to enhance the accuracy, timeliness, and credibility of model outputs by combining external knowledge bases with the generative capabilities of LLMs.

什么是检索增强生成（RAG）？

What is Retrieval-Augmented Generation (RAG)?

检索增强生成是一种将信息检索与文本生成相结合的架构范式。其核心思想是：在LLMs生成回答或内容之前，先从外部知识源（如数据库、文档集合、互联网）中检索出与当前查询最相关的信息片段。然后，将这些检索到的上下文与用户的原始查询一起输入给LLMs，从而引导模型基于更准确、更具体的事实依据进行生成。这种方法不仅有效利用了LLMs强大的语言理解和生成能力，还通过引入外部知识弥补了其内在知识的不足。

Retrieval-Augmented Generation is an architectural paradigm that combines information retrieval with text generation. Its core idea is: before an LLM generates an answer or content, it first retrieves the most relevant information snippets related to the current query from external knowledge sources (such as databases, document collections, the internet). Then, these retrieved contexts, along with the user's original query, are fed into the LLM, thereby guiding the model to generate content based on more accurate and specific factual evidence. This method not only effectively leverages the powerful language understanding and generation capabilities of LLMs but also compensates for the limitations of their intrinsic knowledge by introducing external information.

RAG的主要优势体现在以下几个方面：

• 提升事实准确性：通过提供相关证据，减少模型“捏造”事实的可能性。
• 实现知识更新：无需重新训练整个大模型，仅通过更新外部知识库即可让模型获取最新信息。
• 增强可解释性与可信度：生成的结果可以关联到具体的检索来源，使推理过程更具可追溯性。
• 支持领域专业化：可以轻松接入特定领域的专业知识库，使通用LLMs快速适应垂直场景。

The main advantages of RAG are reflected in the following aspects:

• Improves Factual Accuracy: By providing relevant evidence, it reduces the likelihood of the model "fabricating" facts.
• Enables Knowledge Updates: Allows the model to acquire the latest information by simply updating the external knowledge base, without the need to retrain the entire large model.
• Enhances Explainability and Trustworthiness: Generated results can be linked to specific retrieval sources, making the reasoning process more traceable.
• Supports Domain Specialization: Can easily integrate specialized knowledge bases from specific domains, enabling general-purpose LLMs to quickly adapt to vertical scenarios.

RAG范式的演进

The Evolution of RAG Paradigms

根据该综述论文的梳理，RAG技术主要经历了三个阶段的演进：朴素RAG、高级RAG和模块化RAG。

According to the survey paper, RAG technology has primarily evolved through three stages: Naive RAG, Advanced RAG, and Modular RAG.

1. 朴素RAG

1. Naive RAG

这是RAG最基础的形态，通常遵循“检索-读取”的流水线。流程包括：

1. 索引：将文档语料库进行分块并向量化，构建可检索的索引。
2. 检索：将用户查询向量化，并从索引中检索出Top-K个最相似的文本块。
3. 生成：将检索到的文本块与原始查询拼接，作为提示词输入LLM，生成最终答案。

This is the most basic form of RAG, typically following a "retrieve-then-read" pipeline. The process includes:

1. Indexing: Chunking and vectorizing the document corpus to build a searchable index.
2. Retrieval: Vectorizing the user query and retrieving the Top-K most similar text chunks from the index.
3. Generation: Concatenating the retrieved text chunks with the original query as a prompt input to the LLM to generate the final answer.

这种模式简单直接，但存在明显缺陷，例如检索结果可能不精确、包含无关信息，或者无法处理需要多步推理的复杂问题。

This model is straightforward but has obvious drawbacks, such as potentially imprecise retrieval results containing irrelevant information, or an inability to handle complex problems requiring multi-step reasoning.

2. 高级RAG

2. Advanced RAG

为了克服朴素RAG的不足，高级RAG在检索前后和生成阶段引入了多种优化策略：

• 检索前优化：包括对用户查询进行重写或扩展，以提升其与文档的匹配度；以及对索引文档进行更好的清洗、结构化或摘要。
• 检索中优化：采用更精细的检索策略，如混合检索（结合稠密向量检索和稀疏关键词检索）、重新排序（对初步检索结果进行精排）等。
• 检索后优化：对检索到的内容进行压缩、筛选或总结，只将最精华、最相关的部分输入给LLM，以减少噪声和上下文长度压力。

To overcome the shortcomings of Naive RAG, Advanced RAG introduces various optimization strategies before, during, and after retrieval, as well as during the generation phase:

• Pre-Retrieval Optimization: Includes rewriting or expanding user queries to improve their match with documents, and better cleaning, structuring, or summarizing of indexed documents.
• Retrieval Optimization: Employs more refined retrieval strategies, such as hybrid retrieval (combining dense vector retrieval and sparse keyword retrieval) and re-ranking (fine-ranking of initial retrieval results).
• Post-Retrieval Optimization: Compresses, filters, or summarizes the retrieved content, feeding only the most essential and relevant parts to the LLM to reduce noise and context length pressure.

3. 模块化RAG

3. Modular RAG

这是目前最灵活、最具可扩展性的范式。它将RAG系统解耦为多个功能独立的模块（如查询理解模块、检索器模块、重排序模块、答案生成模块、验证模块等）。这些模块可以像乐高积木一样被组合、替换或调整顺序，甚至引入新的模块（如用于多跳推理的“搜索引擎”模块或用于逻辑判断的“推理”模块）。模块化设计使得系统能够针对不同的任务需求进行定制，并便于集成最新的算法和技术。

This is currently the most flexible and scalable paradigm. It decouples the RAG system into multiple functionally independent modules (such as query understanding module, retriever module, re-ranker module, answer generation module, verification module, etc.). These modules can be combined, replaced, or their order adjusted like Lego bricks, and even new modules can be introduced (such as a "search engine" module for multi-hop reasoning or a "reasoning" module for logical judgment). The modular design allows the system to be customized for different task requirements and facilitates the integration of the latest algorithms and technologies.

RAG的核心技术组件

Core Technical Components of RAG

一个典型的RAG框架建立在三个核心组件之上：检索、增强和生成。

A typical RAG framework is built upon three core components: Retrieval, Augmentation, and Generation.

1. 检索

1. Retrieval

检索组件负责从海量数据中快速、准确地找到相关信息。关键技术包括：

• 向量检索：利用文本嵌入模型将文本转换为高维向量，通过计算向量相似度（如余弦相似度）进行检索。这是当前的主流方法。
• 混合检索：结合基于语义的向量检索和基于关键词匹配的传统检索（如BM25），兼顾召回率与精确率。
• 图检索：当知识以图结构（如知识图谱）组织时，可以基于图查询来检索实体和关系。
• 检索策略：包括多向量表示、查询扩展、上下文学习（ICL）增强检索等。

The retrieval component is responsible for quickly and accurately finding relevant information from massive datasets. Key technologies include:

• Vector Retrieval: Uses text embedding models to convert text into high-dimensional vectors and retrieves information by calculating vector similarity (e.g., cosine similarity). This is the current mainstream method.
• Hybrid Retrieval: Combines semantic-based vector retrieval with traditional keyword-matching retrieval (e.g., BM25), balancing recall and precision.
• Graph Retrieval: When knowledge is organized in a graph structure (e.g., knowledge graph), entity and relationship retrieval can be performed based on graph queries.
• Retrieval Strategies: Include multi-vector representation, query expansion, in-context learning (ICL) enhanced retrieval, etc.

2. 增强

2. Augmentation

增强组件负责对检索到的原始信息进行处理和整合，使其更适合LLM消化并用于生成。常见技术有：

• 提示工程：设计高效的提示词模板，将检索到的上下文、用户查询和生成指令有机结合起来。
• 上下文压缩/摘要：当检索到的文档过长时，对其进行提炼，保留核心信息，以节省宝贵的上下文窗口。
• 多文档融合：当从多个来源检索到信息时，需要去重、解决冲突并合成一致的上下文。

The augmentation component is responsible for processing and integrating the retrieved raw information, making it more suitable for the LLM to digest and use for generation. Common techniques include:

• Prompt Engineering: Designing efficient prompt templates to organically combine the retrieved context, user query, and generation instructions.
• Context Compression/Summarization: When retrieved documents are too long, they are distilled to retain core information, saving valuable context window space.
• Multi-Document Fusion: When information is retrieved from multiple sources, deduplication, conflict resolution, and synthesis into a consistent context are necessary.

3. 生成

3. Generation

生成组件即大语言模型本身，它接收经过增强的提示词，并输出最终的文本。在此环节的优化包括：

• 受控生成：通过约束解码等技术，确保生成内容严格遵循检索到的事实，或符合特定格式要求。
• 引用与溯源：在生成答案的同时，标注出所依据的源文档片段，提高可信度。
• 自我验证与反思：让LLM对自身生成的内容进行事实性检查或逻辑一致性验证，必要时触发重新检索或生成。

The generation component is the large language model itself, which receives the augmented prompt and outputs the final text. Optimizations at this stage include:

• Controlled Generation: Using techniques like constrained decoding to ensure the generated content strictly adheres to the retrieved facts or meets specific format requirements.
• Citation and Attribution: Annotating the generated answer with references to the source document fragments it is based on, enhancing credibility.
• Self-Verification and Reflection: Having the LLM perform factual checks or logical consistency verification on its own generated content, triggering re-retrieval or regeneration if necessary.

评估框架与未来挑战

Evaluation Frameworks and Future Challenges

随着RAG系统的复杂化，如何全面评估其性能变得至关重要。一个完善的RAG评估体系应涵盖多个维度：

• 检索质量：评估检索结果的相关性、召回率和新鲜度。
• 生成质量：评估生成答案的事实准确性、流畅性、相关性和信息量。
• 端到端效率：评估系统的响应延迟和吞吐量。
• 可追溯性与可信度：评估答案引用来源的准确性和支持度。

As RAG systems become more complex, comprehensively evaluating their performance becomes crucial. A robust RAG evaluation system should cover multiple dimensions:

• Retrieval Quality: Evaluates the relevance, recall, and freshness of retrieval results.
• Generation Quality: Evaluates the factual accuracy, fluency, relevance, and informativeness of generated answers.
• End-to-End Efficiency: Evaluates the system's response latency and throughput.
• Traceability and Trustworthiness: Evaluates the accuracy and supportiveness of the sources cited by the answers.

尽管RAG技术发展迅速，但仍面临诸多挑战：

• 检索的“最后一公里”问题：如何确保检索到的信息片段恰好包含生成答案所需的关键证据。
• 复杂推理与多跳检索：对于需要串联多个文档信息才能回答的问题，现有系统仍显吃力。
• 幻觉的根除：即使提供了正确上下文，LLM有时仍会忽略或曲解它，产生幻觉。
• 动态知识更新：如何实现外部知识库的低延迟、实时同步与增量更新。
• 评估基准的标准化：需要建立更全面、更具挑战性的基准测试来推动领域发展。

Despite the rapid development of RAG technology, it still faces numerous challenges:

• The "Last Mile" Problem of Retrieval: How to ensure the retrieved information snippets precisely contain the key evidence needed for generation.
• Complex Reasoning and Multi-Hop Retrieval: Existing systems still struggle with questions that require piecing together information from multiple documents.
• Eradicating Hallucinations: Even when provided with the correct context, LLMs sometimes ignore or misinterpret it, leading to hallucinations.
• Dynamic Knowledge Updates: How to achieve low-latency, real-time synchronization and incremental updates of external knowledge bases.
• Standardization of Evaluation Benchmarks: There is a need to establish more comprehensive and challenging benchmarks to drive progress in the field.

结论

Conclusion

检索增强生成技术通过巧妙地将大语言模型的生成能力与外部知识库的精确信息相结合，为解决LLMs的幻觉、知识过时和黑箱问题提供了一条行之有效的路径。从朴素RAG到模块化RAG的演进，体现了该领域从固定流水线向灵活、可组合智能系统的发展趋势。随着检索、增强、生成各组件技术的不断精进，以及评估体系的日益完善，RAG有望成为构建下一代可信、可靠、可追溯AI应用的关键基石。未来的研究将更聚焦于提升复杂推理能力、实现更精细的知识操控以及构建与人类价值观对齐的负责任RAG系统。

Retrieval-Augmented Generation technology, by ingeniously combining the generative capabilities of large language models with the precise information from external knowledge bases, provides an effective path to address the issues of hallucination, outdated knowledge, and the black-box nature of LLMs. The evolution from Naive RAG to Modular RAG reflects the field's trend from fixed pipelines towards flexible, composable intelligent systems. With continuous advancements in the technologies of the retrieval, augmentation, and generation components, along with the gradual improvement of evaluation systems, RAG is poised to become a key cornerstone for building the next generation of trustworthy, reliable, and traceable AI applications. Future research will focus more on enhancing complex reasoning capabilities, achieving finer-grained knowledge manipulation, and constructing responsible RAG systems aligned with human values.

常见问题（FAQ）

RAG技术具体是如何工作的？

RAG的工作流程通常分为三步：首先将文档分块并向量化建立索引；然后根据用户查询检索最相关的文本片段；最后将这些片段与查询一起输入大语言模型生成答案。

RAG相比传统大语言模型有哪些优势？

RAG能提升事实准确性，减少模型幻觉；无需重新训练即可更新知识；生成结果可追溯来源增强可信度；还能通过接入专业知识库快速适应垂直领域。

RAG技术经历了哪些发展阶段？

RAG技术主要经历了三个阶段演进：从最基础的朴素RAG（检索-读取流程），发展到高级RAG，再到更灵活的模块化RAG架构。