RAG (Retrieval-Augmented Generation)
Definition
RAG is a read-only retrieval mechanism that grounds language models in external, static knowledge bases (documents, wikis, manuals) by searching for semantically similar content and injecting it into the prompt context at inference time.
RAG（检索增强生成）是一种只读检索机制，通过在推理时搜索语义相似的内容并注入到提示上下文中，将语言模型锚定在外部静态知识库（文档、维基、手册）上。
Details
RAG addresses the knowledge cutoff problem by allowing models to access up-to-date information without retraining. The typical pipeline:
- Indexing: Documents are chunked and embedded into a vector database
- Retrieval: User query is embedded and used to search for similar chunks
- Augmentation: Retrieved chunks are prepended to the prompt
- Generation: Model generates response grounded in retrieved context
RAG 通过允许模型访问最新信息而无需重新训练来解决知识截止问题。典型流程包括：索引（文档分块并嵌入向量数据库）、检索（查询嵌入后搜索相似块）、增强（检索块前置到提示）、生成（基于检索上下文生成响应）。
Key characteristics:
- Stateless: Each query is independent; no memory of previous interactions
- Document-centric: Optimized for “What does this document say?” questions
- Similarity-based: Retrieval relies on semantic similarity, not temporal or relational reasoning
关键特征：无状态（每次查询独立，无先前交互记忆）、以文档为中心（优化"这个文档说了什么"类问题）、基于相似度（检索依赖语义相似性，而非时间或关系推理）。
Common pitfall: Using RAG as a substitute for Agent Memory by dumping conversation logs into a vector database. This causes “agent amnesia” because RAG lacks temporal reasoning and cannot maintain coherent state across sessions.
常见陷阱：将对话日志倾倒到向量数据库中，试图用 RAG 替代 Agent Memory。这会导致"智能体失忆"，因为 RAG 缺乏时间推理能力，无法跨会话维护连贯状态。
Connections
- Complements: Agent Memory (RAG for static knowledge, Memory for dynamic context)
- Related to: Harness (RAG is often a tool invoked by the harness)
- Mentioned in: Memory is the Harness
Sources added by Heal on 2026-04-06:
- Vectorize.io - RAG vs Memory · 2026-04
- Memori Labs - Agent Memory Architecture · 2026-04
- Mem0.ai - Memory for AI Agents · 2026-04