[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-llm-wikis-beat-raw-rag-knowledge-work-zh":3,"article-related-llm-wikis-beat-raw-rag-knowledge-work-zh":31,"series-ai-agent-cd51c43c-312b-4bcf-a6b2-b3217c4e05b7":76},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":23,"views":27,"created_at":28,"published_at":29,"topic_cluster_id":30},"cd51c43c-312b-4bcf-a6b2-b3217c4e05b7","llm-wikis-beat-raw-rag-knowledge-work-zh","LLM 維護的 wiki 比原始 RAG 更適合真正的知識工作","\u003Cp data-speakable=\"summary\">\u003Ca href=\"\u002Ftag\u002Fllm\">LLM\u003C\u002Fa> 維護的 wiki 比原始 \u003Ca href=\"\u002Ftag\u002Frag\">RAG\u003C\u002Fa> 更適合知識\u003Ca href=\"\u002Fnews\u002Fclaude-sonnet-46-sre-benchmark-rootly-zh\">工作\u003C\u002Fa>，因為它會累積知識而不是每次重算。\u003C\u002Fp>\u003Cp>我站在這一邊：真正有價值的知識系統，不是更大的\u003Ca href=\"\u002Fnews\u002Fpixelrag-screenshots-retrievable-context-zh\">檢索\u003C\u002Fa>索引，而是由 LLM 持續維護的 wiki。原始 RAG 只能在提問當下把相關片段撈出來，卻不會記住自己已經學到什麼、否決過什麼、上一輪哪個綜合結論被採納。把來源、查詢與修正持續寫回 wiki，才會把一次次工作轉成可累積的結構。這不是搜尋\u003Ca href=\"\u002Fnews\u002Fmcp-servers-ai-workflows-explained-zh\">工具\u003C\u002Fa>升級，而是系統本身開始有記憶。\u003C\u002Fp>\u003Ch2>第一個論點\u003C\u002Fh2>\u003Cp>原始 RAG 的上限是「召回」，不是「理解」。它擅長把對的 chunk 撈出來，卻不擅長把五篇論文、三份會議記錄和一串客訴訊息整合成可重用的結論。只要問題稍微複雜一點，模型每次都得從頭拼一次答案，結果是同一批資料被反覆閱讀、同一組關係被反覆重建。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782760669415-7e3l.png\" alt=\"LLM 維護的 wiki 比原始 RAG 更適合真正的知識工作\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>這種浪費在日常工具裡已經很明顯。像 NotebookLM、\u003Ca href=\"\u002Ftag\u002Fchatgpt\">ChatGPT\u003C\u002Fa> 檔案上傳這類系統，單次問答可以很驚艷，但下一次問相關問題時，仍要重新讀同樣的材料。若一個產品每次都要把已知事實重新算一遍，它就不是知識庫，而是重複運算費用表。知識工作要的是累積，不是反覆啟動。\u003C\u002Fp>\u003Ch2>第二個論點\u003C\u002Fh2>\u003Cp>更好的模型，是讓 LLM 一邊讀，一邊把知識寫回 wiki。每新增一份來源，就更新實體頁、強化主題摘要、標記矛盾、補上交叉連結。Karpathy 提到的核心直覺很重要：知識應該被編譯一次並持續維護，而不是每次查詢都重新推導。這使系統更像軟體，而不是聊天紀錄。\u003C\u002Fp>\u003Cp>實務上的好處是輸出可以沉澱成資產。今天模型產出一張比較表、一段產品論證或一份研究綜整，這些內容應該回寫進 wiki，否則最好的工作只會消失在對話視窗裡。若系統能把每次嚴肅提問都變成下一次更快的起點，研究、競品分析、長期產品探索就會開始複利。這種複利才是知識工作的真正效率來源。\u003C\u002Fp>\u003Ch2>反方可能怎麼說\u003C\u002Fh2>\u003Cp>最有力的反對意見是：由 LLM 維護的 wiki 可能比原始 RAG 更容易把錯誤固化。模型一旦誤讀來源、寫下錯摘要，還把摘要層層傳播到連結頁面，錯誤就會變得很黏。相較之下，原始檢索至少保留了原文，還能讓人回頭核對。對高風險領域來說，這不是小問題。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782760669748-wjd3.png\" alt=\"LLM 維護的 wiki 比原始 RAG 更適合真正的知識工作\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>另一個反對理由是規模。小團隊、單一專案時，一個索引檔加幾個 markdown 頁面就夠了；但資料一多，還是需要搜尋、來源追蹤與審核流程。wiki 不是 source of truth 的替代品，也不是人類判斷的替代品。它只是建立在原始語料上的一層整理，不是把現實壓縮成魔法。\u003C\u002Fp>\u003Cp>但這些限制不推翻這個方向，反而定義了正確做法：wiki 必須明確標示為衍生層、可修訂、可稽核。原始來源保持不可變，wiki 保持可編輯；錯誤因此可見、可回溯、可修正，而不是被藏在聊天歷史裡。風險不是理由去堅持 raw RAG，風險是理由把 wiki 設計成有來源、有版本、有審核的維護型資產。\u003C\u002Fp>\u003Ch2>你能做什麼\u003C\u002Fh2>\u003Cp>如果你是工程師、PM 或創辦人，別再把 LLM 當成一次性回答機器，而要把它當成知識庫的維護工。保留原始來源不動，另外建立一層 wiki 做綜整，再加一份決策與死路的日誌，避免系統重複犯同樣的錯。先從一個領域、一個索引、一條 ingest 流程開始；只要模型能更新頁面、標記矛盾、保留推理脈絡，你就已經做出比聊天更有價值的東西：一個會複利的團隊記憶。","我主張，LLM 維護的 wiki 比原始 RAG 更適合知識工作，因為它能累積、保持更新，還能保留決策脈絡。","gist.github.com","https:\u002F\u002Fgist.github.com\u002Fkarpathy\u002F442a6bf555914893e9891c11519de94f",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782760669415-7e3l.png","ai-agent","zh","697af300-a6ed-47c9-93cc-4c3227a4d862",[17,18,19,20,21,22],"LLM wiki","RAG","知識工作","知識管理","決策記錄","可稽核性",[24,25,26],"原始 RAG 只擅長召回，LLM 維護的 wiki 才能累積知識。","把模型輸出回寫成 wiki，能讓每次研究與分析變成下一次的基礎。","正確做法不是捨棄原始來源，而是建立可追溯、可修訂的衍生知識層。",0,"2026-06-29T19:17:20.761542+00:00","2026-06-29T19:17:20.713+00:00","e3b68196-9e64-4c18-a3b6-a73e73bfb367",{"tags":32,"relatedLang":35,"relatedPosts":39},[33],{"name":18,"slug":34},"rag",{"id":15,"slug":36,"title":37,"language":38},"llm-wikis-beat-raw-rag-knowledge-work-en","LLM wikis beat raw RAG for real knowledge work","en",[40,46,52,58,64,70],{"id":41,"slug":42,"title":43,"cover_image":44,"image_url":44,"created_at":45,"category":13},"cde225a8-eb8e-4724-a089-77f36af0e8a6","mcps-new-primitives-make-agent-middleware-obsolete-zh","MCP 的新原語，正在淘汰自製 agent middleware","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782748974384-f5w9.png","2026-06-29T16:02:24.789168+00:00",{"id":47,"slug":48,"title":49,"cover_image":50,"image_url":50,"created_at":51,"category":13},"6e37d84c-aa27-4d4d-bbf1-81c47dc4522d","mcp-servers-ai-workflows-explained-zh","MCP Server 讓 AI 工具接上工作流","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782747180723-q3gs.png","2026-06-29T15:32:33.536175+00:00",{"id":53,"slug":54,"title":55,"cover_image":56,"image_url":56,"created_at":57,"category":13},"a5333ae2-bfd1-434a-92dd-575e824538c3","openmontage-open-source-ai-video-production-zh","OpenMontage 證明 AI 影片製作該由開源接管","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782685072512-v02f.png","2026-06-28T22:17:22.846394+00:00",{"id":59,"slug":60,"title":61,"cover_image":62,"image_url":62,"created_at":63,"category":13},"caea04da-9e30-4eb8-bb08-2ac3afc4f09e","gemini-35-flash-computer-use-safeguards-zh","Gemini 3.5 Flash 讓你寫電腦操作腳本","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782681504454-c1ly.png","2026-06-28T21:17:56.883563+00:00",{"id":65,"slug":66,"title":67,"cover_image":68,"image_url":68,"created_at":69,"category":13},"362a448e-b40e-437c-9529-94b0fd6a7689","design-md-bridge-taste-to-ui-scaffolds-zh","DESIGN.md 是把品味變成 UI 骨架的缺失橋樑","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782586072806-w93e.png","2026-06-27T18:47:23.886521+00:00",{"id":71,"slug":72,"title":73,"cover_image":74,"image_url":74,"created_at":75,"category":13},"4a77707d-c370-4584-bf40-e71c0414720e","openclaw-agent-control-layer-matters-zh","OpenClaw 證明：代理系統的勝負在控制層，不在模型","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782561770712-4ch7.png","2026-06-27T12:02:20.114404+00:00",[77,82,87,92,97,102,107,112,117,122],{"id":78,"slug":79,"title":80,"created_at":81},"4ae1e197-1d3d-4233-8733-eafe9cb6438b","claude-now-uses-your-pc-to-finish-tasks-zh","Claude 開始幫你操作電腦","2026-03-26T07:20:48.457387+00:00",{"id":83,"slug":84,"title":85,"created_at":86},"5bede67f-e21c-413d-9ab8-54a3c3d26227","googles-2026-ai-agent-report-decoded-zh","Google 2026 AI Agent 報告解讀","2026-03-26T11:15:22.651956+00:00",{"id":88,"slug":89,"title":90,"created_at":91},"2987d097-563f-46c7-b76f-b558d8ef7c2b","kimi-k25-review-stronger-still-not-legend-zh","Kimi K2.5 評測：更強，但還不是神作","2026-03-27T07:15:55.277513+00:00",{"id":93,"slug":94,"title":95,"created_at":96},"95c9053b-e3f4-4cb5-aace-5c54f4c9e044","claude-code-controls-mac-desktop-zh","Claude Code 也能操控 Mac 了","2026-03-28T03:01:58.58121+00:00",{"id":98,"slug":99,"title":100,"created_at":101},"dc58e153-e3a8-4c06-9b96-1aa64eabbf5f","cloudflare-100x-faster-ai-agent-sandbox-zh","Cloudflare 的 AI 沙箱跑超快","2026-03-28T03:09:44.142236+00:00",{"id":103,"slug":104,"title":105,"created_at":106},"1c8afc56-253f-47a2-979f-1065ff072f2a","openai-backs-isara-agent-swarm-bet-zh","OpenAI 挺 Isara 的 agent swarm …","2026-03-28T03:15:27.513155+00:00",{"id":108,"slug":109,"title":110,"created_at":111},"7379b422-576e-45df-ad5a-d57a0d9dd467","openai-plan-automated-ai-researcher-zh","OpenAI 想做自動化 AI 研究員","2026-03-28T03:17:42.090548+00:00",{"id":113,"slug":114,"title":115,"created_at":116},"48c9889e-86df-450b-a356-e4a4b7c83c5b","harness-engineering-ai-agent-reliability-2026-zh","駕馭工程：從「馬具」到「作業系統」，AI Agent 可靠性的終極密碼","2026-03-31T06:42:53.556721+00:00",{"id":118,"slug":119,"title":120,"created_at":121},"96d8e8c8-1edd-475d-9145-b1e7a1b02b65","mcp-explained-from-prompts-to-production-zh","MCP 怎麼把提示詞變工作流","2026-04-01T09:24:39.321274+00:00",{"id":123,"slug":124,"title":125,"created_at":126},"f2ca7720-b471-4ce5-9336-2a9ac2a876fd","amazon-bedrock-agents-multi-agent-workflows-zh","Amazon Bedrock Agents 進入多代理工作流","2026-04-01T09:30:29.945429+00:00"]