Retrieval systems have become a foundational infrastructure component in modern Web services, supporting applications such as content recommendation, advertising targeting, and API discovery. In large-scale industrial environments, retrieval is increasingly deployed as an independent service layer, commonly referred to as Retrieval-as-a-Service (RaaS). This paper presents a system-oriented survey of industrial retrieval pipelines, focusing on architectural design and deployment trade-offs under real-world constraints. Unlike prior surveys that emphasize algorithmic developments, we analyze retrieval systems from an infrastructure perspective, highlighting how latency requirements, scalability constraints, and resource limitations shape system design in production environments. We introduce a unified RaaS pipeline abstraction that models retrieval as a multi-stage service, including high-efficiency candidate generation, embedding-based semantic matching, and resource-aware re-ranking. We further examine the integration of Large Language Model (LLM)-based retrieval mechanisms and analyze their impact on semantic performance, latency, and computational overhead. The results provide a system-level understanding of retrieval as a service-oriented infrastructure and offer practical guidelines for designing scalable, efficient, and QoS-aware retrieval architectures in large-scale Web systems.

翻译：检索系统已成为现代Web服务中基础性的基础设施组件，支持内容推荐、广告定向投放和API发现等应用。在大型工业环境中，检索越来越多地被部署为独立服务层，通常称为检索即服务（RaaS）。本文对工业检索管道进行了系统导向的调研，重点关注实际约束下的架构设计与部署权衡。与以往侧重算法发展的调研不同，我们从基础设施视角分析检索系统，揭示延迟需求、可扩展性约束和资源限制如何影响生产环境中的系统设计。我们引入统一的RaaS管道抽象，将检索建模为多阶段服务，包括高效候选生成、基于嵌入的语义匹配及资源感知重排序。我们进一步考察了大语言模型（LLM）驱动的检索机制的集成，并分析其对语义性能、延迟和计算开销的影响。研究结果为检索作为服务导向型基础设施提供了系统级理解，并为在大型Web系统中设计可扩展、高效且支持服务质量的检索架构提供了实用指南。