案例研究
企业级 RAG 检索与知识服务平台
建设统一知识检索与 RAG 服务底座，用多租户权限、安全裁剪和 OpenAI 兼容接口支撑多业务系统复用。
系统快照
架构、信号与证据会先于正文呈现。- Permission model
- 3-layer
- Retrieval
- Hybrid
- Reuse
- Multi-app
Problem
多个 AI 应用都需要知识检索，但如果每个应用各自搭建 RAG，会造成重复建设、接口不统一、权限控制不一致。企业级检索不仅要答得准，还要确保用户无法检索到权限之外的知识。
Constraints
- 语义检索不能绕过企业权限边界。
- 系统需要支持多租户隔离、配额管理和文档级 ACL。
- 业务应用需要标准接口，而不是直接理解 Embedding、向量库和重排器细节。
System Design
平台设计了三层权限模型：操作权限、知识库范围和文档 ACL。在检索阶段引入 Security Trimming，先过滤无权限片段，再进入生成阶段，避免语义召回绕过授权体系。
服务层封装 OpenAI 兼容接口，支持 Embeddings 与 Chat Completions，同时沉淀标准化知识入库流程。检索侧支持 Dense、BM25、Hybrid 等模式，并补齐可观测与审计机制。
Key Decisions
把 RAG 做成共享基础设施，而不是藏在每个应用里的功能模块。企业文档结构、术语和精确匹配要求差异很大，因此需要同时支持 Dense、BM25 和 Hybrid 检索。
Impact
- 形成公司统一知识检索与 RAG 服务底座，支撑多个 AI 应用稳定运行并跨系统复用。
- 建立企业级语义检索安全机制，解决大模型接入场景下的数据权限隔离问题。
- 降低各业务线重复建设成本，让 AI 能力从单点应用升级为可规模化复用的基础设施。
Reflection
后续重点会放在引用质量、评测集、审计链路和租户级运维能力上，让 RAG 平台真正成为可运营的基础设施。
Problem
Different AI applications needed knowledge retrieval, but repeated one-off RAG systems created duplicated cost and inconsistent permissions.
Enterprise retrieval must answer with the right knowledge while ensuring users cannot retrieve content outside their authorization scope.
Constraints
Semantic retrieval cannot bypass enterprise permission boundaries.
The system needed to support multi-tenant isolation, quota management, and document-level ACLs.
Application teams needed a standard API instead of integrating directly with embeddings, vector stores, and rerankers.
System Design
The platform uses a three-layer permission model: operation permission, knowledge-base scope, and document ACL.
Security Trimming is applied during retrieval so unauthorized chunks are filtered before they can reach generation.
The service exposes OpenAI-compatible embeddings and chat-completion interfaces, with standardized ingestion and retrieval workflows.
Key Decisions
RAG should be treated as shared infrastructure, not a feature hidden inside each application.
Dense, BM25, and Hybrid retrieval modes are all supported because enterprise documents differ in structure, vocabulary, and precision requirements.
Impact
The platform became a unified knowledge retrieval foundation for multiple AI systems and reduced duplicated build cost across teams.
It established a safer enterprise semantic retrieval mechanism for LLM applications.
Reflection
Future work should strengthen citation quality, evaluation sets, audit trails, and tenant-level operations.