[{"data":1,"prerenderedAt":133},["ShallowReactive",2],{"article-other\u002Fneeds-rag":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"tags":11,"body":13,"_type":127,"_id":128,"_source":129,"_file":130,"_stem":131,"_extension":132},"\u002Farticles\u002Fother\u002Fneeds-rag","other",false,"","RAG 是否还需要：从热度降温到 AI 制药场景的价值重估","分析 RAG 热度下降的多重原因，探讨在 AI 制药等高合规 ToB 场景下 RAG 不可替代的价值——零幻觉、知识时效性与异构数据整合。","2026-08-09",[12],"随笔",{"type":14,"children":15,"toc":117},"root",[16,24,40,46,53,78,84,89,107,112],{"type":17,"tag":18,"props":19,"children":21},"element","h2",{"id":20},"起因",[22],{"type":23,"value":20},"text",{"type":17,"tag":25,"props":26,"children":27},"p",{},[28,30,38],{"type":23,"value":29},"看到了一篇文章(",{"type":17,"tag":31,"props":32,"children":36},"a",{"href":33,"rel":34},"https:\u002F\u002Fjuejin.cn\u002Fpost\u002F7631495035477983242",[35],"nofollow",[37],{"type":23,"value":33},{"type":23,"value":39},")\n，上面说了为什么RAG现在没有之前讨论度高了。然后就想结合AI制药的这个业务场景搞明白这件事情-是否需要RAG，在什么情况下使用RAG。",{"type":17,"tag":18,"props":41,"children":43},{"id":42},"现在的理解2026-08-08",[44],{"type":23,"value":45},"现在的理解(2026-08-08)",{"type":17,"tag":47,"props":48,"children":50},"h3",{"id":49},"rag热度下降的原因",[51],{"type":23,"value":52},"RAG热度下降的原因",{"type":17,"tag":54,"props":55,"children":56},"ol",{},[57,63,68,73],{"type":17,"tag":58,"props":59,"children":60},"li",{},[61],{"type":23,"value":62},"Skill + tool 足够应对大部分的ToC产品，像Claude Code、Codex这种Agent。大部分某个技能的专业知识本质就是结构化的逻辑判断、经验规则，而这些东西天然就是skill的主战场，不需要用RAG。只有面对海量、无法规则化的细节数据时才真正需要 RAG。",{"type":17,"tag":58,"props":64,"children":65},{},[66],{"type":23,"value":67},"综合成本高。",{"type":17,"tag":58,"props":69,"children":70},{},[71],{"type":23,"value":72},"LLM本身的大模型不断升级，context windows从早期的4K到现在200K+，可以塞下大部分内容。",{"type":17,"tag":58,"props":74,"children":75},{},[76],{"type":23,"value":77},"现在大量的Agent应用都是ToC，自然RAG用的就不多。",{"type":17,"tag":47,"props":79,"children":81},{"id":80},"在ai制药的应用场景下rag的用武之地",[82],{"type":23,"value":83},"在AI制药的应用场景下，RAG的用武之地",{"type":17,"tag":25,"props":85,"children":86},{},[87],{"type":23,"value":88},"RAG在AI制药场景下用处非常大。",{"type":17,"tag":54,"props":90,"children":91},{},[92,97,102],{"type":17,"tag":58,"props":93,"children":94},{},[95],{"type":23,"value":96},"零幻觉容忍：毕竟制药人命关天，大模型随机乱编是致命的，RAG强制让大模型基于检索到的事实回答问题，且每一句回答都可以溯源到具体的文献、专利或实验ID。",{"type":17,"tag":58,"props":98,"children":99},{},[100],{"type":23,"value":101},"知识的时效性与私密性：RAG无需重新训练大模型，即可动态接入每日更新的外部文献和高壁垒的私有数据。",{"type":17,"tag":58,"props":103,"children":104},{},[105],{"type":23,"value":106},"结构化和非结构化的桥梁：制药数据涵盖文本、分子 SMILES、蛋白质序列、表格、知识图谱等，RAG（尤其是结合了 GraphRAG 的多模态检索）能够很好地将这些异构数据串联起来。",{"type":17,"tag":18,"props":108,"children":110},{"id":109},"总结",[111],{"type":23,"value":109},{"type":17,"tag":25,"props":113,"children":114},{},[115],{"type":23,"value":116},"RAG在ToC的产品中没那么广泛的应用场景，但是在数据量大、精确度要求高、商业价值极高且强调合规的 toB 垂直领域（如 AI 制药），RAG 是不可替代的技术底座。",{"title":7,"searchDepth":118,"depth":118,"links":119},2,[120,121,126],{"id":20,"depth":118,"text":20},{"id":42,"depth":118,"text":45,"children":122},[123,125],{"id":49,"depth":124,"text":52},3,{"id":80,"depth":124,"text":83},{"id":109,"depth":118,"text":109},"markdown","content:articles:other:needs-RAG.md","content","articles\u002Fother\u002Fneeds-RAG.md","articles\u002Fother\u002Fneeds-RAG","md",1789465115004]