做智能体应用的团队会发现传统RAG的痛点被精准戳中，Agentic RAG的改进方案直接可用，建议点开看看具体架构设计。
Milvus团队指出，传统RAG在智能体工作流中表现不佳，存在单次检索遗漏上下文、相似性不等于相关性、缺乏检索质量检查、单一策略不适用所有查询等问题。但RAG并未死亡，而是进化成了Agentic RAG，通过查询路由、混合检索、检索评估（如Corrective RAG）和多步检索来解决上述问题。生产中的教训是：检索层必须匹配工作负载，架构越复杂越难维护。文章提供了更深入的架构建议。
A lot of the "RAG is dead" arguments have some truth: traditional RAG is a poor fit for agentic work...
A lot of the "RAG is dead" arguments have some truth: traditional RAG is a poor fit for agentic workloads. 𝗛𝗼𝘄𝗲𝘃𝗲𝗿, 𝗥𝗔𝗚 𝗶𝘀𝗻'𝘁 𝗱𝗲𝗮𝗱; 𝗮𝗴𝗲𝗻𝘁𝗶𝗰 𝗥𝗔𝗚 𝗶𝘀 𝗿𝗲𝗹𝗲𝘃𝗮𝗻𝘁 𝗮𝘀 𝗲𝘃𝗲𝗿, 𝗮𝗻𝗱 𝗵𝗲𝗿𝗲'𝘀 𝘄𝗵𝘆. 𝗧𝗿𝗮𝗱𝗶𝘁𝗶𝗼𝗻𝗮𝗹 𝗥𝗔𝗚 𝗱𝗼𝗲𝘀 𝘀𝘁𝗿𝘂𝗴𝗴𝗹𝗲 𝘄𝗶𝘁𝗵 𝗮𝗴𝗲𝗻𝘁𝗶𝗰 𝘄𝗼𝗿𝗸𝗹𝗼𝗮𝗱𝘀: • Single-pass retrieval often misses context needed for multi-step tasks. • Similarity is not always the same as relevance. • Naive pipelines often have no check for bad retrieval before generation. • One retrieval strategy does not fit every query type. 𝗔𝗴𝗲𝗻𝘁𝗶𝗰 𝗥𝗔𝗚 𝗮𝗱𝗱𝗿𝗲𝘀𝘀𝗲𝘀 𝘁𝗵𝗲𝘀𝗲 𝗴𝗮𝗽𝘀 𝘄𝗶𝘁𝗵: • Query routing, so different queries can use different retrieval paths • Hybrid retrieval, so dense and sparse search can work together when needed • Retrieval evaluation, including Corrective RAG-style checks, to flag weak context before generation • Multi-step retrieval, so agents can gather context across a reasoning chain 𝗧𝗵𝗲 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝗹𝗲𝘀𝘀𝗼𝗻 𝗶𝘀 𝘀𝗶𝗺𝗽𝗹𝗲: 𝗺𝗮𝘁𝗰𝗵 𝘁𝗵𝗲 𝗿𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹 𝗹𝗮𝘆𝗲𝗿 𝘁𝗼 𝘁𝗵𝗲 𝘄𝗼𝗿𝗸𝗹𝗼𝗮𝗱. More architecture d milvus.io/blog/build-sma… ean better results. The more complex the architecture, the harder it is to maintain and debug. 𝗜𝗳 𝘆𝗼𝘂 𝘄𝗮𝗻𝘁 𝘁𝗼 𝗱𝗶𝗴 𝗱𝗲𝗲𝗽𝗲𝗿: → Smarter RAG with routing and hybrid retrieval: https://t.co/IR3lCHapku 💬 1 🔄 0 ❤️ 1 👀 26 📊 1 ⚡