<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>OntiCards 部落格</title>
    <link>https://onticards.com/zh-hk/blog</link>
    <description>OntiCards 產品實踐與數據智能洞察（繁體中文）</description>
    <language>zh-hk</language>
    <lastBuildDate>Mon, 21 Sep 2026 00:00:00 GMT</lastBuildDate>
    <atom:link href="https://onticards.com/zh-hk/rss.xml" rel="self" type="application/rss+xml"/>
    <item>
      <title>Gemini 越界之後：Agent 護欄要建在數據層</title>
      <link>https://onticards.com/zh-hk/blog/agent-guardrails-data-permission</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/agent-guardrails-data-permission</guid>
      <pubDate>Mon, 21 Sep 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>industry</category>
      <description>谷歌證實 Gemini 喺安全測試中侵入三間真實公司系統，四大前沿實驗室至此全部公開過智能體越界事件。事故歸因指向同一件事：權限邊界畫錯咗。企業要讓智能體碰生產數據，護欄必須內生於數據權限層，而非只靠提示詞約束。</description><enclosure url="https://onticards.com/images/blog/agent-guardrails-data-permission/cover.svg" type="image/svg+xml"/>
    </item>
    <item>
      <title>Agent 攻入電網調度：兩星期變幾個鐘，按鈕仍喺人手</title>
      <link>https://onticards.com/zh-hk/blog/agentic-grid-planning-lessons</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/agentic-grid-planning-lessons</guid>
      <pubDate>Fri, 18 Sep 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>industry</category>
      <description>AWS 聯同 Duke Energy 推出電網規劃智能體計劃，將併網研究嘅數據準備由兩星期壓到幾個鐘。真正值得企業參考嘅唔止係提速，仲有分工設計：模擬軟件計數、智能體跑流程、工程師拍板，全程可審計。</description><enclosure url="https://onticards.com/images/blog/agentic-grid-planning-lessons/cover.svg" type="image/svg+xml"/>
    </item>
    <item>
      <title>介面層戰爭開打：模型喺貶值，數據語境先至係寶</title>
      <link>https://onticards.com/zh-hk/blog/interface-war-data-context</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/interface-war-data-context</guid>
      <pubDate>Wed, 16 Sep 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>industry</category>
      <description>Dreamforce 24 小時連發 Koa 模型、AIforce 介面層同兩單 Agent 落地個案，DeepSeek 更將任務成本推到 0.07 美元。模型喺貶值，企業數據語境先至係新戰場。</description><enclosure url="https://onticards.com/images/blog/interface-war-data-context/cover.svg" type="image/svg+xml"/>
    </item>
    <item>
      <title>智能體的下半場：由編排競賽到本體圖記憶</title>
      <link>https://onticards.com/zh-hk/blog/ontology-graph-agent-memory</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/ontology-graph-agent-memory</guid>
      <pubDate>Mon, 14 Sep 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>tech</category>
      <description>過去兩日，OpenAI 將智能體編排做成公共 API，Salesforce 讓智能體追逐跨週目標，Iyuno CLOE 就揭示咗更深一層：令理解持續複利嘅唔係更大嘅上下文窗口，而係一張隨任務不斷生長嘅持久本體圖。</description><enclosure url="https://onticards.com/images/blog/ontology-graph-agent-memory/cover.svg" type="image/svg+xml"/>
    </item>
    <item>
      <title>OpenAI 進場做數據智能體：決勝點喺語義層，唔係模型</title>
      <link>https://onticards.com/zh-hk/blog/openai-data-agent-semantic-layer</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/openai-data-agent-semantic-layer</guid>
      <pubDate>Fri, 11 Sep 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>industry</category>
      <description>OpenAI 喺 ChatGPT Work 推出 Data agent，直接駁通企業數據倉。但基準數據講得好白：企業級問數嘅樽頸從來唔係模型——企業 schema 上準確率得約 21%，信心遠超管控。語義層同數據治理，先至係 AI 問數落地嘅分水嶺。</description><enclosure url="https://onticards.com/images/blog/openai-data-agent-semantic-layer/cover.svg" type="image/svg+xml"/>
    </item>
    <item>
      <title>AI 問數攻入真實企業數據庫：91.7% 背後的三個架構真相</title>
      <link>https://onticards.com/zh-hk/blog/agentic-nl2sql-real-enterprise-data</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/agentic-nl2sql-real-enterprise-data</guid>
      <pubDate>Wed, 09 Sep 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>tech</category>
      <description>最新研究喺 900 條真實企業查詢上將 NL2SQL 答對率推到 91.7%，領先次優基線 54.6 個百分點。本文拆解三個架構真相：真實企業數據庫有幾難、成本感知 Agent 點樣做、點解語義資產先係最大槓桿。</description><enclosure url="https://onticards.com/images/blog/agentic-nl2sql-real-enterprise-data/cover.svg" type="image/svg+xml"/>
    </item>
    <item>
      <title>模型疲勞時代：企業 AI 底座嘅抗迭代設計</title>
      <link>https://onticards.com/zh-hk/blog/model-fatigue-data-foundation</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/model-fatigue-data-foundation</guid>
      <pubDate>Mon, 07 Sep 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>industry</category>
      <description>一星期之內 Anthropic、Meta、Google、OpenAI 齊發新模型，CNBC 造咗個新詞「模型疲勞」。本文拆解企業評測選型嘅真實負擔，並提出四條抗迭代設計原則，等數據資產穿越模型迭代週期。</description><enclosure url="https://onticards.com/images/blog/model-fatigue-data-foundation/cover.svg" type="image/svg+xml"/>
    </item>
    <item>
      <title>研讀 Qwen3.8-Flash-Next：詳解 51B N-gram Embedding 架構設計與行業影響</title>
      <link>https://onticards.com/zh-hk/blog/qwen-flash-next-51b-ngram-embedding</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/qwen-flash-next-51b-ngram-embedding</guid>
      <pubDate>Fri, 04 Sep 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>tech</category>
      <description>從一個被廣為流傳嘅「58B」參數細節切入，我哋嘗試拆解 51B N-gram Embedding 嘅真實運作機制；同時也直面三類最常見嘅誤區——佢唔係內置 RAG，亦唔會令 RAG 退出歷史舞臺，更加唔會帶嚟通用能力嘅越級躍遷。一份更冷靜嘅工程判斷，往往比一份興奮嘅測評更有價值。</description><enclosure url="https://onticards.com/images/blog/qwen-flash-next-51b-ngram-embedding/cover.svg" type="image/svg+xml"/>
    </item>
    <item>
      <title>Agent 能力衝上 Critical，企業敢用嘅前提係可審計</title>
      <link>https://onticards.com/zh-hk/blog/agent-auditability-critical</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/agent-auditability-critical</guid>
      <pubDate>Fri, 04 Sep 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>industry</category>
      <description>OpenAI 宣布 Astra 成為首個達到 Critical 網絡安全能力層級嘅大模型，隨之而嚟嘅係「不可讀推理」嘅監控爭議。當 Agent 自主性持續升級，企業落地真正嘅門檻唔係榜單分數，而係每一步操作係咪可審計、可追溯、可控。</description><enclosure url="https://onticards.com/images/blog/agent-auditability-critical/cover.svg" type="image/svg+xml"/>
    </item>
    <item>
      <title>Data Agent 自成一個賽道：IDC 首份評估嘅三件事</title>
      <link>https://onticards.com/zh-hk/blog/enterprise-data-agent-market</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/enterprise-data-agent-market</guid>
      <pubDate>Wed, 02 Sep 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>industry</category>
      <description>九月一日，IDC 首次發布中國 Data Agent 廠商評估：18 間廠商得 4 間入到領導者象限，仲預測去到 2028 年，六成中國 500 強企業會部署企業級 Data Agent。數據智能體第一次成為獨立採購品類。呢篇拆解首份評估嘅三個訊號，同落地之前要打定嘅四張地基。</description><enclosure url="https://onticards.com/images/blog/enterprise-data-agent-market/cover.svg" type="image/svg+xml"/>
    </item>
    <item>
      <title>Agent 進入隔離治理期：從 ServiceNow 九倍增長、Replica Cyber 同 Kyndryl Incore Bank，睇 OntiCards 嘅提前卡位</title>
      <link>https://onticards.com/zh-hk/blog/agent-runtime-isolation-audit</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/agent-runtime-isolation-audit</guid>
      <pubDate>Tue, 01 Sep 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>tech</category>
      <description>企業級 Agent 上週同時衝兩條主線：ServiceNow 生產部署 9 個月增長 9 倍、Replica Cyber 幫 Agent 裝咗隔離引擎、Kyndryl × Incore Bank 用語義層加守門員 Agent，將銀行 KYC 鎖入治理範圍。Gartner 預測 2027 年底四成 Agent 項目會被取消 — 唔係 Agent 唔得，係治理仲未趕到。</description><enclosure url="https://onticards.com/images/blog/agent-runtime-isolation-audit/cover.svg" type="image/svg+xml"/>
    </item>
    <item>
      <title>Text-to-SQL 入企業，準確率只剩 25%：語義層點解係必選項</title>
      <link>https://onticards.com/zh-hk/blog/enterprise-text-to-sql-reality-check</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/enterprise-text-to-sql-reality-check</guid>
      <pubDate>Mon, 31 Aug 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>tech</category>
      <description>實驗室裏接近滿分嘅 Text-to-SQL 模型，放入真實企業數據庫後準確率暴跌至 22%-25%。arXiv 最新基準測試同 dbt Labs 生產環境實驗共同指向一個結論：冇語義層，LLM 連「搵啱表」都困難。</description><enclosure url="https://onticards.com/images/blog/enterprise-text-to-sql-reality-check/cover.svg" type="image/svg+xml"/>
    </item>
    <item>
      <title>語義層成 Agent 新底座：從 Oracle、Google 同學界本週嘅共識，睇 OntiCards 點定位</title>
      <link>https://onticards.com/zh-hk/blog/semantic-layer-data-agent</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/semantic-layer-data-agent</guid>
      <pubDate>Sun, 30 Aug 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>tech</category>
      <description>Oracle 要 Agent 揀 trusted reports 而唔係寫 SQL，Google 將 Measures + Graph 並入 BigQuery，SIGMOD 2026 Semantic-Layer-mediated NL2SQL 將 Spider 2.0 準確率由 17% 推到 94.15%——三家同一個禮拜押注同一件事，語義層正式成為 Data Agent 嘅新底座。</description><enclosure url="https://onticards.com/images/blog/semantic-layer-data-agent/cover.svg" type="image/svg+xml"/>
    </item>
    <item>
      <title>Agent 進入算賬期：數據準備度決定落地成敗</title>
      <link>https://onticards.com/zh-hk/blog/agent-data-readiness</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/agent-data-readiness</guid>
      <pubDate>Sat, 29 Aug 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>industry</category>
      <description>浦發銀行部署 2500 個智能體覆蓋 440 個場景，江蘇銀行 Token 用量 6 個月內暴漲 18 倍。當 Agent 從試點進入算賬期，數據質素與查詢準確率才是 ROI 的真正分水嶺。</description><enclosure url="https://onticards.com/images/blog/agent-data-readiness/cover.svg" type="image/svg+xml"/>
    </item>
    <item>
      <title>讓數據開口說話：OntiCards 嘅四層架構同我哋嘅實踐</title>
      <link>https://onticards.com/zh-hk/blog/onticards-four-layer-architecture</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/onticards-four-layer-architecture</guid>
      <pubDate>Fri, 28 Aug 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>product</category>
      <description>將數據由「囤喺倉入面嘅欄位」升級做「俾业务人直接提問嘅知識資產」，係我哋做 OntiCards 嘅核心命題。本文講清楚我哋係點樣拆解呢件事嘅。</description><enclosure url="https://onticards.com/images/blog/onticards-four-layer-architecture/cover.svg" type="image/svg+xml"/>
    </item>
    <item>
      <title>一次真實嘅電商數據問答驗證：由訂單、退款到用戶嘅全鏈路問數</title>
      <link>https://onticards.com/zh-hk/blog/ecommerce-data-qa-walkthrough</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/ecommerce-data-qa-walkthrough</guid>
      <pubDate>Fri, 28 Aug 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>tech</category>
      <description>我哋用內部測試環境嘅電商數據源（5 張表）跑咗一次完整嘅 OntiCards 實戰：將數據接埋嚟、生成數據卡片、用自然語言問业务問題。本文記錄全過程同真實數據。</description><enclosure url="https://onticards.com/images/blog/ecommerce-data-qa-walkthrough/cover.svg" type="image/svg+xml"/>
    </item>
    <item>
      <title>自然語言問數，將年銷百萬輛級車企的營銷作戰室變成一段對話</title>
      <link>https://onticards.com/zh-hk/blog/car-marketing-warroom-natural-language-qa</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/car-marketing-warroom-natural-language-qa</guid>
      <pubDate>Fri, 28 Aug 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>solutions</category>
      <description>由 MVP 應答率超過 90%、準確率超過 85%，到覆蓋營銷作戰中心 14 個看板、支援可交互 BI，一家年銷百萬輛級的全國性車企把經營問數做成隨時可開的對話。</description><enclosure url="https://onticards.com/images/blog/car-marketing-warroom-natural-language-qa/cover.svg" type="image/svg+xml"/>
    </item>
    <item>
      <title>2026 下半年產品路線圖：由企業權限到數據治理嘅下一程</title>
      <link>https://onticards.com/zh-hk/blog/2026-h2-product-roadmap</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/2026-h2-product-roadmap</guid>
      <pubDate>Fri, 28 Aug 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>product</category>
      <description>將下半年嘅產品計劃講清楚：邊啲喺度做、邊啲做咗、合作夥伴同企業用戶可以期待咩。</description><enclosure url="https://onticards.com/images/blog/2026-h2-product-roadmap/cover.svg" type="image/svg+xml"/>
    </item>
    <item>
      <title>區域頭部城商行將全行知識，變成員工隨口一問即答</title>
      <link>https://onticards.com/zh-hk/blog/regional-bank-knowledge-platform</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/regional-bank-knowledge-platform</guid>
      <pubDate>Thu, 06 Aug 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>solutions</category>
      <description>26 個部門調研、8 大業務專屬智能體、雙維標籤與多渠道權限路由……一家資產規模 6000 億級嘅區域頭部城商行，將散落全行嘅制度知識收攏成一個入口，員工隨口一問，數秒即答。</description><enclosure url="https://onticards.com/images/blog/regional-bank-knowledge-platform/cover.svg" type="image/svg+xml"/>
    </item>
    <item>
      <title>銀行嘅數據資產，點樣由『淨係技術先睇得明嘅庫』變成『業務同事隨口問到嘅數』</title>
      <link>https://onticards.com/zh-hk/blog/bank-data-asset-semantic-layer</link>
      <guid isPermaLink="true">https://onticards.com/zh-hk/blog/bank-data-asset-semantic-layer</guid>
      <pubDate>Thu, 23 Jul 2026 00:00:00 GMT</pubDate><dc:creator>OntiCards 團隊</dc:creator><category>solutions</category>
      <description>基於 OntiCards 數據語義層 + 智能體平台交付：將晦澀嘅表結構翻譯成業務可讀嘅 DataCard，自然語言問數自動生成 SQL 同圖表。AI 產出 2 小時採納率 85%，同業務專家 8 小時產出相當，效率提升 4 倍。</description><enclosure url="https://onticards.com/images/blog/bank-data-asset-semantic-layer/cover.svg" type="image/svg+xml"/>
    </item>
  </channel>
</rss>
