[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-sim-visual-agent-workflow-canvas-zh":3,"article-related-sim-visual-agent-workflow-canvas-zh":36,"series-tools-aa78f0bb-2826-4d1e-a7ab-a6f5bb15a873":88},{"id":4,"title":5,"content":6,"summary":7,"source":8,"source_url":9,"author":10,"image_url":11,"keywords":12,"language":19,"translated_content":10,"views":20,"is_premium":21,"created_at":22,"updated_at":22,"cover_image":11,"published_at":23,"rewrite_status":24,"rewrite_error":10,"rewritten_from_id":25,"slug":26,"category":27,"related_article_id":28,"status":29,"google_indexed_at":10,"x_posted_at":10,"tweet_text":10,"title_rewritten_at":10,"title_original":10,"key_takeaways":30,"topic_cluster_id":34,"embedding":35,"is_canonical_seed":21},"aa78f0bb-2826-4d1e-a7ab-a6f5bb15a873","Sim 把 agent 流程變成畫布","\u003Cp data-speakable=\"summary\">我拆 Sim 的畫布式 \u003Ca href=\"\u002Ftag\u002Fagent\">agent\u003C\u002Fa> 編排，順手整理成可直接套用的工作流模板、自架與向量搜尋做法。\u003C\u002Fp>\u003Cp>我玩 agent workflow 一陣子了，最煩的不是模型不會答，是工具一多就開始亂。今天接 prompt，明天補 YAML，後天加 queue、vector store、重試，最後整套東西像在跟我鬧脾氣。Demo 看起來很順，真的要上線時，流程一複雜就露餡。我最近盯上 Sim，https:\u002F\u002F\u003Ca href=\"\u002Ftag\u002Fgithub\">github\u003C\u002Fa>.com\u002Fsimstudioai\u002Fsim，因為它至少沒假裝這些麻煩不存在。\u003C\u002Fp>\u003Cp>它的路數很直白：把 agent orchestration 做成 visual canvas，讓你能設計、執行、自己架、自己擴。這種產品我通常半信半疑，因為很多工具都只是在換皮，複雜度沒少，只是換個地方藏。Sim 比較像是把流程攤開來給你看，這點我反而比較能接受。\u003C\u002Fp>\u003Cp>我會想拆它，不是因為它多神，而是它踩中了我一直在意的那條線：workflow editor、runtime、self-host、vector search、tools，全都要能接起來，不然就是半套。這篇我就直接拆它的玩法，順手整理成你可以抄的版本。\u003C\u002Fp>\u003Cp>觸發我仔細看的是 Sim 的 GitHub repo，還有它自己講的產品定位。README 明講它是 build、deploy、orchestrate \u003Ca href=\"\u002Ftag\u002Fai-agents\">AI agents\u003C\u002Fa> 的平台，而且有 visual canvas 跟 self-host 路線。這不是空話，我看的是 repo 結構跟安裝路徑，不是只看首頁文案。原始來源在這裡：\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fsimstudioai\u002Fsim\" target=\"_blank\" rel=\"noreferrer\">https:\u002F\u002Fgithub.com\u002Fsimstudioai\u002Fsim\u003C\u002Fa>。\u003C\u002Fp>\u003Ch2>我先看它有沒有把「畫布」當裝飾品\u003C\u002Fh2>\u003Cblockquote>“Design agent workflows visually on a canvas—connect agents, tools, and blocks, then run them instantly.”\u003C\u002Fblockquote>\u003Cp>翻譯一下就是：先把流程畫出來，再直接跑。這句話聽起來很像低碼平台的老梗，但我覺得重點不是畫，而是「畫完就能跑」。很多工具都卡在中間那段，設計和執行是兩套系統，最後你還是得自己補 glue code，照樣痛。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779201890676-1xsq.png\" alt=\"Sim 把 agent 流程變成畫布\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>我以前做過一個客服分流 agent，需求很簡單：先判斷問題類型，再去抓知識庫，最後決定要不要轉人工。結果實作時，流程拆在三個檔案、兩個 handler、外加一堆臨時 flag。每次改一個\u003Ca href=\"\u002Fnews\u002Fkubernetes-1-36-1-patch-releases-zh\">分支\u003C\u002Fa>，我都要回頭找是哪個 callback 在偷改狀態。那種時候我最想要的不是更強的模型，是一張能一眼看懂的流程圖。\u003C\u002Fp>\u003Cp>Sim 的 canvas 如果只是美觀 UI，我不會多看一眼。但它把 agent、tool、block 都攤在同一個視覺層，這就有價值了。因為 agent 系統最麻煩的地方不是推理，而是你根本看不出哪一步把資料弄壞了。畫布至少讓你先把責任切清楚。\u003C\u002Fp>\u003Cp>實操寫法很簡單：每個 node 只做一件事，輸入和輸出都要能說人話。不要一個 agent 包五個任務，然後怪模型不穩。你要的是可讀性，不是把複雜度塞進黑盒。\u003C\u002Fp>\u003Cul>\u003Cli>一個 node 對應一個責任，不要混雜。\u003C\u002Fli>\u003Cli>工具要窄，別做萬用工具。\u003C\u002Fli>\u003Cli>先畫 failure path，再畫 happy path。\u003C\u002Fli>\u003C\u002Ful>\u003Cp>如果你本來就有在看 \u003Ca href=\"https:\u002F\u002Fwww.langchain.com\u002F\" target=\"_blank\" rel=\"noreferrer\">LangChain\u003C\u002Fa> 或 \u003Ca href=\"https:\u002F\u002Fmastra.ai\u002F\" target=\"_blank\" rel=\"noreferrer\">Mastra\u003C\u002Fa>，Sim 比較像是把你腦中那張流程圖直接做成產品，而不是只給你一堆 runtime API 叫你自己想像。\u003C\u002Fp>\u003Ch2>它不是 editor 而已，是 editor 綁 runtime\u003C\u002Fh2>\u003Cblockquote>“Build Workflows with Ease”\u003C\u002Fblockquote>\u003Cp>這句很像行銷話，但我看它的重點不是「容易」，而是「設計完能立刻跑」。很多平台都能讓你拖拉幾個節點，問題是拖完之後要怎麼執行、怎麼測、怎麼改，答案通常很破碎。你最後還是要跳出去寫另外一套執行層。\u003C\u002Fp>\u003Cp>我比較在意的是 design 和 execution 的距離有沒有被縮短。因為 agent workflow 一旦拆成兩個世界，debug 就會\u003Ca href=\"\u002Fnews\u002Fdbt-sl-turns-semantic-layer-setup-into-a-loop-zh\">變成\u003C\u002Fa>災難。你在畫布上看到的是 A，runtime 跑出來的是 B，然後大家開始互相懷疑。這種事我看太多了。\u003C\u002Fp>\u003Cp>Sim 的價值在於，它讓你在同一個介面裡做建模、測試、執行。這對我來說不是便利，是少掉翻譯成本。少一次翻譯，就少一次出錯。尤其是當工具輸出格式變了、模型回傳半成品、或者某個分支只在真實資料下炸掉時，你會很感謝這種緊密耦合。\u003C\u002Fp>\u003Cp>實操上，我會這樣用：先做最小流程，只放一個 input、一個 agent、一個 tool、一個 output。先別急著加分支，先確認主幹穩。很多人一開始就想做完整 orchestrator，結果連最基本的 schema 都沒驗證。\u003C\u002Fp>\u003Cul>\u003Cli>先跑最小閉環，再加分支。\u003C\u002Fli>\u003Cli>用真實髒資料測，不要只拿乾淨範例。\u003C\u002Fli>\u003Cli>每次改節點，都回頭看執行紀錄。\u003C\u002Fli>\u003C\u002Ful>\u003Cp>它的 editor 用的是 \u003Ca href=\"https:\u002F\u002Freactflow.dev\u002F\" target=\"_blank\" rel=\"noreferrer\">React Flow\u003C\u002Fa>，這選得很務實。React Flow 的好處就是 node graph 不會變成裝飾品，是真的能拿來編排流程。\u003C\u002Fp>\u003Ch2>Copilot 在這裡不是噱頭，是省你點來點去\u003C\u002Fh2>\u003Cblockquote>“Supercharge with Copilot — Leverage Copilot to generate nodes, fix errors, and iterate on flows directly from natural language.”\u003C\u002Fblockquote>\u003Cp>翻譯一下就是：你可以直接叫 \u003Ca href=\"\u002Ftag\u002Fcopilot\">Copilot\u003C\u002Fa> 幫你生節點、修錯、改流程。這比那種「問答型 AI 助手」有用多了，因為它是插進 workflow authoring 的中間層，不是掛在旁邊聊天。\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779201885855-q9yu.png\" alt=\"Sim 把 agent 流程變成畫布\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>我對這種東西一直是半信半疑。原因很簡單，AI 很會把東西生得像真的。節點名稱看起來對，連線看起來也順，但細看就會發現它偷偷編了欄位、漏了例外、或是把一個應該失敗的情況硬寫成成功。這種假順手，我吃過太多次虧。\u003C\u002Fp>\u003Cp>所以我會把 Copilot 定位成「草稿機」，不是「驗證機」。它適合幫我把流程先搭出來，減少點選成本；但最後的 schema、權限、錯誤處理，還是得我自己看。尤其 agent workflow 這種東西，一個錯誤的 edge 就可能讓整個流程走歪。\u003C\u002Fp>\u003Cp>實操寫法：叫它先生骨架，再人工補血。你可以讓它幫你把節點類型、連線順序、基本說明先列出來，但不要直接把生成結果當成可上線版本。任何自然語言產物，都要過一輪人工檢查。\u003C\u002Fp>\u003Cul>\u003Cli>用 Copilot 產生骨架，不要直接產生最終版。\u003C\u002Fli>\u003Cli>每個 node 的輸入、輸出、權限都要人工確認。\u003C\u002Fli>\u003Cli>把錯誤處理當成必填，不要等出事才補。\u003C\u002Fli>\u003C\u002Ful>\u003Cp>如果你想對照更熟悉的生態，可以看 \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ffeatures\u002Fcopilot\" target=\"_blank\" rel=\"noreferrer\">GitHub Copilot\u003C\u002Fa> 跟 \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Ffeatures\u002Fmodels\" target=\"_blank\" rel=\"noreferrer\">GitHub Models\u003C\u002Fa>。Sim 的做法比較像把這類能力直接塞進流程編排裡，而不是讓你自己切頁面。\u003C\u002Fp>\u003Ch2>它把 vector search 當流程的一部分，不是外掛\u003C\u002Fh2>\u003Cblockquote>“Integrate Vector Databases — Upload documents to a vector store and let agents answer questions grounded in your specific content.”\u003C\u002Fblockquote>\u003Cp>翻譯一下就是：文件上傳進向量庫，agent 回答時直接用你的內容當依據。這件事我很在意，因為太多 RAG 系統都死在一個地方：大家都在講 embedding，卻沒人處理資料清理、更新、切 chunk、查詢品質。\u003C\u002Fp>\u003Cp>我以前做內部知識庫問答時，最煩的不是模型不夠聰明，是 retrieval 一直亂撈。文件版本沒控好、索引沒更新、查回來的段落跟問題根本不對焦，最後使用者只會說「這 AI 很會講廢話」。所以我現在看任何 agent 平台，第一個問題都是：它怎麼把 retrieval 接進整個 workflow？\u003C\u002Fp>\u003Cp>Sim 讓我比較舒服的地方，是它不是把 vector search 當成額外功能，而是當成 agent flow 的一段。這很\u003Ca href=\"\u002Fnews\u002Fwhy-mp2-still-matters-broadcast-audio-zh\">重要\u003C\u002Fa>。因為 knowledge retrieval 不是獨立服務，它本來就應該跟工具調用、模型推理、輸出校驗放在同一條路上看。\u003C\u002Fp>\u003Cp>實操寫法：把 ingestion 跟 query 分開。不要把文件匯入、索引更新、回答查詢混在同一個流程裡，不然你會很難 debug。再來，測試時不要只問正確答案，要故意問模糊、矛盾、過期的問題，看看 agent 會不會亂編。\u003C\u002Fp>\u003Cul>\u003Cli>只索引真的會用到的內容。\u003C\u002Fli>\u003Cli>ingestion 和 query 分開做。\u003C\u002Fli>\u003Cli>用 adversarial 問題測 retrieval。\u003C\u002Fli>\u003C\u002Ful>\u003Cp>它提到的後端選項包含 \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fpgvector\u002Fpgvector\" target=\"_blank\" rel=\"noreferrer\">pgvector\u003C\u002Fa>，這種選擇我反而比較信。夠普通，才比較容易維護。\u003C\u002Fp>\u003Ch2>自架不是加分項，是它能不能進公司內網的門票\u003C\u002Fh2>\u003Cblockquote>“Cloud-hosted: sim.ai” and “Self-hosted: Docker Compose”\u003C\u002Fblockquote>\u003Cp>翻譯一下就是：你可以用雲端版，也可以自己架。這件事看起來平凡，但對台灣很多團隊來說超現實。因為一旦牽涉到內部文件、客戶資料、權限控管，很多 SaaS 直接就被打槍。你沒有 self-host，連 PoC 都很難推進。\u003C\u002Fp>\u003Cp>我看 Sim 的安裝路徑時，會注意它是不是只會講漂亮話。結果它把 NPM、\u003Ca href=\"\u002Ftag\u002Fdocker\">Docker\u003C\u002Fa> Compose、手動安裝都列出來了，還明講需要 Bun、Node.js 20+、PostgreSQL 12+ 和 pgvector。這種寫法很實際，因為它承認世界上真的有人要自己管環境。\u003C\u002Fp>\u003Cp>更重要的是，這代表它不是那種只給你一個雲端控制台的殼。你可以看到它有 Next.js、socket server、database migrations、background jobs 這些東西。這些不是裝飾，是你之後要維運的現實。\u003C\u002Fp>\u003Cp>實操上，我會這樣落地：先用 hosted 版驗證流程，再切 self-hosted。這樣可以先知道產品有沒有用，再決定值不值得把它放進自己的 infra。別一開始就把整套系統綁死，結果 demo 完才發現權限、資料流、網路政策全部卡住。\u003C\u002Fp>\u003Cul>\u003Cli>先用 hosted 跑 PoC，再決定要不要自架。\u003C\u002Fli>\u003Cli>把 staging 跟 production 分開。\u003C\u002Fli>\u003Cli>先寫好 secrets 管理規則，再讓別人碰流程。\u003C\u002Fli>\u003C\u002Ful>\u003Cp>它的技術棧也很直白：\u003Ca href=\"https:\u002F\u002Fnextjs.org\u002F\" target=\"_blank\" rel=\"noreferrer\">Next.js\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Fbun.sh\u002F\" target=\"_blank\" rel=\"noreferrer\">Bun\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Fwww.postgresql.org\u002F\" target=\"_blank\" rel=\"noreferrer\">PostgreSQL\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Fsocket.io\u002F\" target=\"_blank\" rel=\"noreferrer\">Socket.IO\u003C\u002Fa>。我喜歡這種沒有刻意炫技的組合。\u003C\u002Fp>\u003Ch2>這 repo 看起來像平台，不像週末 demo\u003C\u002Fh2>\u003Cblockquote>“The open-source platform to build AI agents and run your agentic workforce.”\u003C\u002Fblockquote>\u003Cp>翻譯一下就是：它想當平台，不只是 library。這差很多。Library 是拿來拼的，platform 是拿來長期用的。你如果要做的是一套會被團隊反覆修改、反覆跑、反覆接工具的 agent 系統，那你要看的就不是「它能不能 demo」，而是「它能不能活下來」。\u003C\u002Fp>\u003Cp>我會看 repo 的結構、文件、背景工作、執行邊界，因為這些東西比首頁文案誠實。Sim 的 repo 有 monorepo 味道，也有 async jobs、remote execution、schema 驗證這些痕跡。這表示它不是只想把 prompt 包成 UI，而是真的在處理 agent 會遇到的髒活。\u003C\u002Fp>\u003Cp>我也注意到它用了 \u003Ca href=\"https:\u002F\u002Ftrigger.dev\u002F\" target=\"_blank\" rel=\"noreferrer\">Trigger.dev\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Fe2b.dev\u002F\" target=\"_blank\" rel=\"noreferrer\">E2B\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Fgithub.com\u002Flaverdet\u002Fisolated-vm\" target=\"_blank\" rel=\"noreferrer\">isolated-vm\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Fzod.dev\u002F\" target=\"_blank\" rel=\"noreferrer\">Zod\u003C\u002Fa>、\u003Ca href=\"https:\u002F\u002Form.drizzle.team\u002F\" target=\"_blank\" rel=\"noreferrer\">Drizzle\u003C\u002Fa>。這組合很像在告訴我：它知道 agent 不是只會聊天，還會跑任務、碰資料、執行程式碼，所以邊界要先畫好。\u003C\u002Fp>\u003Cp>實操寫法很直接：你要先問自己，現在缺的是畫布、runtime、還是執行平台。不要因為看到 visual canvas 就誤以為整套問題都解了。如果你的痛點是權限、排程、重試、審計，那你要找的是平台級設計，不是漂亮 editor。\u003C\u002Fp>\u003Cul>\u003Cli>先判斷你需要 library 還是 platform。\u003C\u002Fli>\u003Cli>看 repo 有沒有 tests、migrations、docs、deploy path。\u003C\u002Fli>\u003Cli>凡是會 async 跑任務的系統，都要先想失敗怎麼收。\u003C\u002Fli>\u003C\u002Ful>\u003Ch2>可抄的模板\u003C\u002Fh2>\u003Cpre>\u003Ccode># Agent Workflow Blueprint for a Sim-like Canvas\n\n## 1. Goal\n- Problem:\n- User:\n- Success criteria:\n\n## 2. Inputs\n- User request:\n- System context:\n- Retrieved documents:\n- Required secrets:\n\n## 3. Nodes\n\n### Intake\n- Type: input\n- Purpose: normalize request\n- Output schema:\n\n### Router\n- Type: agent\n- Purpose: choose path\n- Output schema:\n\n### Retrieval\n- Type: vector search\n- Purpose: ground answer in internal content\n- Output schema:\n\n### Tool Call\n- Type: tool\n- Purpose: call external API or internal service\n- Output schema:\n\n### Synthesis\n- Type: agent\n- Purpose: combine context and tool results\n- Output schema:\n\n### Validation\n- Type: guardrail\n- Purpose: check schema, policy, and safety\n- Output schema:\n\n### Delivery\n- Type: output\n- Purpose: send final result\n- Output schema:\n\n## 4. Connections\n- Intake -> Router\n- Router -> Retrieval\n- Router -> Tool Call\n- Retrieval -> Synthesis\n- Tool Call -> Synthesis\n- Synthesis -> Validation\n- Validation -> Delivery\n\n## 5. Rules\n- Every node has one job only.\n- Every external call is retryable.\n- Every retrieval step is testable with known docs.\n- Every agent decision is logged.\n- Every workflow has a failure path.\n\n## 6. Self-host checklist\n- [ ] PostgreSQL ready\n- [ ] Vector support enabled\n- [ ] Migrations applied\n- [ ] Secrets stored outside repo\n- [ ] Background jobs configured\n- [ ] Realtime server reachable\n- [ ] Workflow editor accessible\n\n## 7. Testing checklist\n- [ ] Happy path passes\n- [ ] Empty input fails cleanly\n- [ ] Tool timeout handled\n- [ ] Retrieval returns no results handled\n- [ ] Malformed model output rejected\n- [ ] Retry path works\n- [ ] Logs show node-by-node execution\n\n## 8. Minimal orchestrator prompt\nYou are the workflow orchestrator.\n\nYour job:\n1. Inspect the request.\n2. Select the correct path.\n3. Retrieve only the needed context.\n4. Call tools only when necessary.\n5. Produce a structured final result.\n6. Fail clearly when a step cannot continue.\n\nReturn:\n- route\n- reasoning summary\n- required actions\n- final output schema\u003C\u002Fcode>\u003C\u002Fpre>\u003Cp>這段我刻意寫成你可以直接拿去改的版本，因為我覺得這才是拆方法論該有的樣子。不是只講概念，而是讓你馬上能落到自己的 workflow、自己的工具、自己的部署規則。\u003C\u002Fp>\u003Cp>我如果明天要開一個新的 agent 專案，我會先拿這份骨架，然後把工具、retrieval、validation、self-host 規則一個個填進去。不要一開始就追求花俏，先把資料流、錯誤流、執行流講清楚，才不會又做出一套看起來很猛、實際上很脆的東西。\u003C\u002Fp>\u003Cp>來源致謝：上面的拆解主要來自 Sim GitHub repo \u003Ca href=\"https:\u002F\u002Fgithub.com\u002Fsimstudioai\u002Fsim\" target=\"_blank\" rel=\"noreferrer\">https:\u002F\u002Fgithub.com\u002Fsimstudioai\u002Fsim\u003C\u002Fa>，以及我對 README、安裝說明和 repo 結構的整理。模板段落是我根據原始資料重寫的衍生版本，不是直接複製原文。\u003C\u002Fp>","我拆 Sim 的畫布式 agent 編排，順手整理成可直接套用的工作流模板、自架與向量搜尋做法。","github.com","https:\u002F\u002Fgithub.com\u002Fsimstudioai\u002Fsim",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779201890676-1xsq.png",[13,14,15,16,17,18],"agent workflow","visual canvas","self-host","vector search","RAG","orchestration","zh",0,false,"2026-05-19T14:44:16.281068+00:00","2026-05-19T14:44:16.246+00:00","done","c4b8d70b-5e8a-456b-be7b-4d9c64cdb5e4","sim-visual-agent-workflow-canvas-zh","tools","e0ec5f99-c3c7-4032-8351-b640927f2601","published",[31,32,33],"Sim 的重點不是低碼，而是把 agent 流程、runtime 和 self-host 放在同一個框架裡。","它最值得抄的地方，是把 retrieval、tools、validation 都當成 workflow 的一部分。","如果你要做可維護的 agent 系統，先把流程圖、失敗路徑和部署規則寫清楚。","c3c88dd2-a940-438a-b359-0e5a24562273","[-0.010129833,0.013022423,0.021912035,-0.10299897,-0.019518608,-0.0141482,-0.01722284,0.008568251,0.008713094,0.017946506,-0.02118727,0.00083213975,-0.037349924,0.011450643,0.13288382,-0.007224805,0.0031692216,0.00506158,0.0030840263,0.009108313,0.010584083,0.009863168,0.02147892,-0.0050210864,0.019309694,-0.008877408,0.00575095,-0.00068079156,0.04469056,-0.010185337,0.019664848,0.0377445,0.0068529113,-0.021412782,0.0075484244,0.0111441435,0.018191515,0.0069817556,0.010044036,0.013228357,0.0029059378,-0.0013316518,0.018565848,0.012815554,0.0039896467,0.025278311,-0.009561012,-0.065867044,-0.000774209,0.0041596186,0.0062261545,0.031676434,-0.002633582,-0.12935139,0.008390346,0.00646714,0.0014627057,-0.002576618,0.025996335,0.01616299,-0.010320244,0.0049251365,-0.015970362,-0.024882767,-0.020322803,0.0042788456,0.029963858,-0.017497312,-0.0195879,-0.0056303977,-0.018051367,-0.0132797295,0.00014178449,0.005029613,0.016941624,-0.03360263,-0.006944037,0.004684112,0.0076617924,0.012698043,0.03967253,-0.016701592,0.003843792,0.012826476,0.01034131,-0.010487658,-0.010087035,0.0035621535,0.011590322,0.0085848225,0.013979008,0.031150952,0.009843455,-0.013294273,0.0032450422,-0.020455386,-0.01163657,-0.006285712,0.0061161625,-0.00083561504,-0.02889903,-0.032946218,0.0032967036,0.0106110675,0.015962642,0.014759489,-0.0041224384,-0.0030867886,0.009122971,0.01795959,-0.004877618,-0.0146977175,-0.0061502177,0.0028336185,-0.012229512,-0.15193485,-0.0066214073,0.0006430952,0.027610594,-0.006525574,-0.007255207,0.012297958,0.018362908,0.01181692,-0.0005576329,0.008461478,-0.0042134267,0.0029748767,-0.008603671,0.024524627,0.006602525,-0.007365171,-0.0326253,0.006535025,0.016526949,-0.0024854366,-0.01754768,0.004730743,-0.027116356,0.0122827515,-0.0018671332,0.012538537,0.00010926549,-0.02393354,-0.0562821,-0.023590237,-0.051219527,0.020504734,-0.0028276527,-0.025267033,0.00066128816,-0.0033911539,-0.040731072,-0.015858883,0.014507404,-0.038745068,0.004350366,0.0024262322,-0.010342875,0.018642118,-0.0014497663,0.014044523,-1.4345911e-05,0.03011983,0.0129272565,0.03150285,-0.020495873,0.0064067664,-0.000259001,0.02871431,-0.0011341678,-0.022454217,0.007987734,0.0072823158,0.014043847,0.0073431903,-0.015460925,-0.0065108314,0.013734885,-0.009953903,0.012930418,-0.012023089,-0.006033329,0.02273777,-0.0032842136,0.0015007983,-0.011093608,0.004023967,0.028666744,-0.009143648,-0.022210103,0.020679176,0.0059045423,-0.0065716207,0.013205594,-0.01010601,0.004036618,-0.012321009,0.013803672,0.024955735,0.000120709934,0.004096878,-0.027662031,-0.034270696,-0.009599343,-0.020359764,-0.014686772,-0.016047418,0.014238826,0.0091241505,-0.020184074,0.004840256,0.006310787,0.011166003,-0.004645897,-0.0108554,0.0074413787,-0.002385893,-0.015269693,-0.019444581,0.017332064,-0.016682602,-0.004292792,-0.0031041098,0.008079363,-0.002819457,-0.020450052,-0.01464292,-0.03527633,0.010344136,0.013892262,0.037164122,-0.00063883193,0.0029108713,0.012808491,5.004331e-05,-0.0035542874,0.02733398,-0.020243885,0.026486153,-0.0080465125,0.026702365,-0.011758178,0.014656454,0.042117603,-0.0038631924,0.004414525,0.0032101034,-0.012926915,0.010737876,-0.006933726,-0.02617699,0.006685598,-0.01808402,-0.0014252323,-0.013149247,-0.004627322,-0.0074405577,-0.013380999,0.0028543854,-0.022389269,0.008737631,-0.009517335,-0.008746896,-0.020832784,-0.013905144,0.024040373,0.016881013,-0.023584893,-0.012194583,0.0070471293,0.023542928,0.010249578,0.0011613154,-0.007489302,0.013309905,-0.070149824,0.044744223,-0.0026073772,-0.030238684,0.0011180795,-0.00994883,0.0041408213,0.007828351,-0.021934,0.0052105556,0.0067418567,-0.012408842,0.003627312,0.0073022842,0.018615857,0.025520818,-0.008145906,0.004965853,0.011317716,-0.01586038,-0.009446278,0.016800575,0.020796014,0.016794916,-0.0036283128,-0.0010747407,-0.0013694697,0.02680726,-0.002341507,0.009603892,0.00077488937,0.04364211,-0.012607288,-0.013561405,0.00979345,-0.006181609,0.015565789,-0.017513644,0.0068749823,-0.0073026656,0.00466462,-0.0008941051,-0.0077347565,0.0045479517,0.006710127,-0.01410364,-0.015427908,-0.0002472584,-0.0072854613,-0.026610902,-0.02134646,0.0033972634,0.0026920026,-0.00169192,0.025555102,0.009432156,0.03714338,-0.03430212,0.015614741,0.021322135,0.0075225323,-0.012326652,-0.0062922128,0.022663945,-0.022407232,0.00524551,-0.015953047,0.00331096,-0.012946715,0.017994937,0.012304852,-0.018007642,0.015129485,-0.04217066,0.0042647314,-0.010187654,0.016767815,-0.030175382,-0.037450317,0.0006578793,-0.008310289,-0.01573516,0.028606443,0.0025866232,-0.0068183406,0.008876627,-0.0019369432,0.0037540472,0.010768378,-0.019252473,0.021695847,0.03288443,-0.009654085,0.014872432,-0.020805988,0.0018909663,0.005272699,0.0055775545,-0.024987087,0.0057314103,-0.038330894,0.0013973154,-0.0039055764,0.00044319747,-0.00069439545,0.037944116,-0.02869247,-0.004665565,-0.045943215,0.0012979632,0.02125161,0.011647051,0.01977559,-0.024657156,0.001975529,0.031748712,-0.025976822,0.0007129977,-0.0031108912,-0.025953762,-0.02575796,0.026097964,-0.010484911,0.0026675065,-0.0102912625,0.008327504,-0.012050646,0.010741028,0.004262692,0.003769841,-0.013245078,-0.004634602,-0.02491339,-0.0065296446,0.01442311,-0.010671623,0.0021205875,-5.4867673e-05,0.02924606,-0.0046157455,0.0039435276,-0.0028628092,0.018046213,-0.023092786,-0.028221587,0.0067557464,-0.008927631,0.012045124,0.004154637,-0.028536351,0.0037498344,0.00025762935,0.01195083,-0.017868746,-0.004832148,-0.000327224,-0.0018083868,-0.028500281,-0.019428367,0.019238332,-0.04645316,0.006710893,-0.02918927,-0.033147227,0.016510028,-0.00066867645,0.018013358,0.030006543,0.014160686,0.004093498,0.00075623806,-0.031799078,-0.024590818,0.0017841422,0.016763417,0.009993567,0.029394401,-0.0070577045,0.009535779,-0.007934581,-0.013835605,0.036778968,0.0071641444,-0.020160021,0.016553009,0.019339817,-0.023975426,0.020325433,0.009747334,-0.012698678,-0.009917688,-0.0032259983,-0.0049629733,0.0025855824,-0.022834992,0.014122621,0.025314411,0.0020285356,0.0076730167,-0.005322039,-0.028852629,0.03300276,0.016899476,-0.007210358,0.027462283,-0.01258242,0.008423678,0.00038909246,0.017946515,0.037771374,-0.022341827,-0.010920939,-0.032494273,-0.0179478,0.021717912,0.008181441,0.0006958775,-0.0011098641,-0.039591532,0.0016593586,-0.014684442,0.0025165759,-0.01151945,-0.020674264,-0.017677037,0.0029036133,0.010364607,-0.018689508,-0.0016712284,-0.019918717,0.0151565885,0.012613025,0.024638548,0.001902992,-0.00909779,-0.0149584105,0.009810903,0.018289277,-0.037252605,-0.017840138,0.011627033,0.01813191,0.006721947,0.023834482,-0.015148475,-0.002682261,-0.008161358,-0.00053411996,-0.0033111735,-0.018315274,0.012626045,0.01886951,0.012317219,0.011383893,-0.011845455,0.017559236,0.0020356763,-0.011892277,0.019334659,-0.0059083523,0.00880876,0.020748284,-0.016241005,0.018863924,-0.014632879,-0.0036839067,-0.0011792787,-0.003828602,0.024271343,-0.12798786,0.017436763,-0.0056004794,0.011157677,0.00462151,0.003208334,-0.0051331175,-0.02358124,0.0042412113,0.017320693,-0.01851272,-0.012470311,0.017184496,0.005710278,-0.015821083,0.0053421445,-0.012814253,0.0035595258,0.00083465624,-0.037594333,-0.002983562,0.0076930723,0.029897459,-0.023940383,-0.0054442305,-0.009579063,0.0012405849,-0.0069292495,0.018796472,0.008364257,-0.01132416,-0.013363739,-0.029667111,-0.00062437134,0.002324232,-0.009418269,0.0155191505,-0.01262352,0.0144390715,0.026019404,-0.02140902,-0.0054320865,0.006691733,-0.031513076,-0.014181882,0.008339972,0.03096709,-0.008363756,-0.008714756,-0.004357203,-0.02912346,0.005153973,-0.018060979,-0.029269418,-0.011834032,-0.0037419288,-0.02531173,-0.0025972014,-0.0059890314,0.019597901,-0.0011506069,0.0063187913,-0.017874708,-0.0056537176,-0.00040704958,0.006148577,0.0037531855,0.041390207,0.0030705556,-0.012171631,-0.0013562104,-0.013416022,-0.0028102566,0.013158611,-0.042520285,0.015200933,-0.01987449,0.02406344,-0.024364233,0.00036377725,-0.017651713,0.008158667,-0.0677549,0.007361815,-0.0060077193,-0.013401034,0.0066492017,0.009465747,0.02946526,-0.025951013,-0.010958687,-0.0031820864,0.0068038786,-0.003979416,-0.0016055579,-0.013639085,0.0017678969,0.031288363,-0.001525256,0.017344773,-0.014500644,-0.023438321,-0.019015435,-0.0021834283,0.017700465,-0.03491894,0.005130131,0.022109712,-0.019533444,-0.007567492,0.007085972,9.8067845e-05,-0.020609945,-0.12682778,-0.028654914,-0.0032355264,0.004421211,0.037215926,-0.009903935,0.004796591,0.009395628,-0.00020941472,-0.022624053,0.011332305,-0.03621395,-0.005991979,-0.016629145,-0.002799023,0.14652306,0.0017919986,-0.009813898,-0.0010077322,-0.0052809194,0.0056947926,-0.019634597,-0.009560976,0.0274981,0.007437462,-0.021369122,0.008742456,-0.009964019,-0.004614118,-0.0024753292,0.0068882527,0.017484747,-0.013328051,-0.008136848,-0.0048222025,0.0036350705,0.021552965,-0.048744798,0.018753687,0.019678209,-0.012065858,0.023261577,0.0058303797,-0.033883743,0.012343656,0.010111655,-0.018463708,-0.027546104,-0.020349396,-0.02511303,-0.007984211,-0.05538312,0.016034666,-0.009834754,-0.008101976,-0.00076329074,0.0092144,0.02167503,0.009205051,0.0014302375,0.012008406,0.0070175175,-0.019479813,0.025850166,-0.007467656,0.006845716,0.021257497,0.019999923,0.017569466,0.023400899,-0.0011359728,0.0030760993,-0.016225351,0.006510325,-0.0246172,-0.023463855,-0.010696541,0.023116305,0.019250643,0.015572566,-0.0004637347,0.0028604423,0.023703171,-0.023457242,0.003169643,0.0054680016,0.020335762,0.00396711,-0.006074308,-0.013306435,0.010653243,0.013052484,0.0070934803,0.017882755,0.017521285,0.03410267,-0.009579947,0.014777045,-0.015286914,-0.0006380576,-0.00986299,-0.029933635,0.007978582,-0.024037415,0.016934855,0.0027543071,0.01084929,0.02180703,-0.007523216,0.022705318]",{"tags":37,"relatedLang":47,"relatedPosts":51},[38,40,42,44,46],{"name":17,"slug":39},"rag",{"name":16,"slug":41},"vector-search",{"name":13,"slug":43},"agent-workflow",{"name":14,"slug":45},"visual-canvas",{"name":15,"slug":15},{"id":28,"slug":48,"title":49,"language":50},"sim-visual-agent-workflow-canvas-en","Sim turns agent workflows into a visual canvas","en",[52,58,64,70,76,82],{"id":53,"slug":54,"title":55,"cover_image":56,"image_url":56,"created_at":57,"category":27},"9ae0773b-5a25-4d26-a750-c26b7f031f0f","low-latency-layer-reflex-anti-lag-linux-gpus-zh","low_latency_layer 讓 Linux GPU 也有低延遲","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779201242989-tzoj.png","2026-05-19T14:33:31.138213+00:00",{"id":59,"slug":60,"title":61,"cover_image":62,"image_url":62,"created_at":63,"category":27},"2cc4bd83-4170-4fe8-8204-2c6f32aae872","dbt-sl-turns-semantic-layer-setup-into-a-loop-zh","dbt sl 讓設定變成迴圈","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779200667586-cuev.png","2026-05-19T14:23:58.228721+00:00",{"id":65,"slug":66,"title":67,"cover_image":68,"image_url":68,"created_at":69,"category":27},"5951c90e-9b00-43b0-9135-88e1e6ea4376","kubernetes-v136-release-notes-playbook-zh","Kubernetes v1.36 把發布說明變作戰手冊","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779197685570-cy9v.png","2026-05-19T13:34:13.951402+00:00",{"id":71,"slug":72,"title":73,"cover_image":74,"image_url":74,"created_at":75,"category":27},"749d323d-fc9b-4492-946d-48f95ee0037d","kubernetes-turns-clusters-into-declared-state-zh","Kubernetes 把叢集變成宣告狀態","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779197095928-obv6.png","2026-05-19T13:24:24.468851+00:00",{"id":77,"slug":78,"title":79,"cover_image":80,"image_url":80,"created_at":81,"category":27},"a1bd40fe-2b96-430c-bce3-17f4b3284333","ibm-vibe-coding-guide-turns-prompts-into-code-zh","IBM Vibe Coding 把提示詞變程式碼","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779183875735-m3wz.png","2026-05-19T09:44:02.030134+00:00",{"id":83,"slug":84,"title":85,"cover_image":86,"image_url":86,"created_at":87,"category":27},"55c43732-c8a9-4ee1-a373-383fa27a41b4","anthropic-buys-stainless-sdk-tool-rivals-zh","Anthropic 收購 Stainless，SDK 工具進入戰局","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1779180252426-zkmj.png","2026-05-19T08:43:39.377553+00:00",[89,94,99,104,109,114,119,124,129,134],{"id":90,"slug":91,"title":92,"created_at":93},"de769291-4574-4c46-a76d-772bd99e6ec9","googles-biggest-gemini-launches-in-2026-zh","Google 2026 最大 Gemini 盤點","2026-03-26T07:26:39.21072+00:00",{"id":95,"slug":96,"title":97,"created_at":98},"855cd52f-6fab-46cc-a7c1-42195e8a0de4","surepath-real-time-mcp-policy-controls-zh","SurePath 推出即時 MCP 政策控管","2026-03-26T07:57:40.77233+00:00",{"id":100,"slug":101,"title":102,"created_at":103},"9b19ab54-edef-4dbd-9ce4-a51e4bae4ebb","mcp-in-2026-the-ai-tool-layer-teams-use-zh","2026 年 MCP：團隊真的在用的 AI 工具層","2026-03-26T08:01:46.589694+00:00",{"id":105,"slug":106,"title":107,"created_at":108},"af9c46c3-7a28-410b-9f04-32b3de30a68c","prompting-in-2026-what-actually-works-zh","2026 提示工程，真正有用的是什麼","2026-03-26T08:08:12.453028+00:00",{"id":110,"slug":111,"title":112,"created_at":113},"05553086-6ed0-4758-81fd-6cab24b575e0","garry-tan-open-sources-claude-code-toolkit-zh","Garry Tan 開源 Claude Code 工具包","2026-03-26T08:26:20.068737+00:00",{"id":115,"slug":116,"title":117,"created_at":118},"042a73a2-18a2-433d-9e8f-9802b9559aac","github-ai-projects-to-watch-in-2026-zh","2026 必看 20 個 GitHub AI 專案","2026-03-26T08:28:09.619964+00:00",{"id":120,"slug":121,"title":122,"created_at":123},"a5f94120-ac0d-4483-9a8b-63590071ac6a","claude-code-vs-cursor-2026-zh","Claude Code 與 Cursor 深度對比：202…","2026-03-26T13:27:14.279193+00:00",{"id":125,"slug":126,"title":127,"created_at":128},"0975afa1-e0c7-4130-a20d-d890eaed995e","practical-github-guide-learning-ml-2026-zh","2026 機器學習入門 GitHub 實用指南","2026-03-27T01:16:49.712576+00:00",{"id":130,"slug":131,"title":132,"created_at":133},"bfdb467a-290f-4a80-b3a9-6f081afb6dff","aiml-2026-student-ai-ml-lab-repo-review-zh","AIML-2026：像課綱的學生實驗 Repo","2026-03-27T01:21:51.467798+00:00",{"id":135,"slug":136,"title":137,"created_at":138},"80cabc3e-09fc-4ff5-8f07-b8d68f5ae545","ai-trending-github-repos-and-research-feeds-zh","AI Trending：把 AI 資源收成一張表","2026-03-27T01:31:35.262183+00:00"]