[{"data":1,"prerenderedAt":528},["ShallowReactive",2],{"blog-\u002Fblog\u002Fai-chatbot-knowledge-base-quality":3},{"id":4,"title":5,"body":6,"date":515,"description":516,"extension":517,"meta":518,"navigation":519,"path":520,"seo":521,"stem":522,"tags":523,"__hash__":527},"blog\u002Fblog\u002Fai-chatbot-knowledge-base-quality.md","你的知識庫決定 AI 客服的智商——怎麼餵資料才不會養出白痴",{"type":7,"value":8,"toc":500},"minimark",[9,14,18,21,24,32,35,39,42,45,112,118,125,129,134,137,140,222,225,230,237,241,244,247,257,260,273,278,281,302,308,312,315,318,323,326,332,356,360,363,435,442,445,448,453,456,459,464,467,470,475,478,484,488,491,494],[10,11,13],"h2",{"id":12},"你花了錢做-ai-客服結果它最常說的一句話是很抱歉我無法回答","你花了錢做 AI 客服，結果它最常說的一句話是「很抱歉，我無法回答」",[15,16,17],"p",{},"這是我聽過最多的抱怨。老闆花了十幾萬做 AI 客服系統，上線第一週，客戶問「你們的 A 方案跟 B 方案差在哪」，AI 回答「很抱歉，我目前無法提供這方面的資訊，建議您聯繫客服人員」。",[15,19,20],{},"客戶的反應：「所以這個機器人到底有什麼用？」",[15,22,23],{},"老闆的反應：「AI 果然還不行。」",[15,25,26,27,31],{},"兩個人都搞錯重點了。",[28,29,30],"strong",{},"問題不在 AI 模型笨，在你餵進去的資料根本沒有答案","。",[15,33,34],{},"這就像你請了一個記憶力超強的實習生，但只給他一份過期的產品目錄和一堆格式混亂的 Word 檔，然後期望他能回答客戶的任何問題。他當然答不出來——不是他笨，是你沒給他東西。",[10,36,38],{"id":37},"同一個模型知識庫品質差五倍","同一個模型，知識庫品質差五倍",[15,40,41],{},"我做過一個實驗：同一個 AI 模型（Claude Sonnet），接上兩套不同品質的知識庫，回答同樣 50 個客戶問題。",[15,43,44],{},"結果：",[46,47,48,64],"table",{},[49,50,51],"thead",{},[52,53,54,58,61],"tr",{},[55,56,57],"th",{},"指標",[55,59,60],{},"亂餵的知識庫",[55,62,63],{},"整理過的知識庫",[65,66,67,79,90,101],"tbody",{},[52,68,69,73,76],{},[70,71,72],"td",{},"正確回答率",[70,74,75],{},"34%",[70,77,78],{},"82%",[52,80,81,84,87],{},[70,82,83],{},"「無法回答」比例",[70,85,86],{},"41%",[70,88,89],{},"8%",[52,91,92,95,98],{},[70,93,94],{},"答案含幻覺（編造資訊）",[70,96,97],{},"25%",[70,99,100],{},"10%",[52,102,103,106,109],{},[70,104,105],{},"平均回應成本",[70,107,108],{},"每則 8 元",[70,110,111],{},"每則 2.5 元",[15,113,114,117],{},[28,115,116],{},"同一個模型，差距是 2.4 倍的正確率、3.2 倍的成本","。模型沒換，唯一的變數是知識庫怎麼整理。",[15,119,120,121,124],{},"更恐怖的是「幻覺」那一行——亂餵的版本有四分之一的回答",[28,122,123],{},"看起來很有自信，但內容是錯的","。AI 不會說「我不確定」，它會很流暢地編一個聽起來合理但完全錯誤的答案。你的客戶信了，然後你就有客訴了。",[10,126,128],{"id":127},"三個決定-ai-客服智商的維度","三個決定 AI 客服智商的維度",[130,131,133],"h3",{"id":132},"維度一格式ai-讀得懂不代表讀得準","維度一：格式——AI 讀得懂，不代表讀得準",[15,135,136],{},"大部分公司丟進知識庫的是什麼？PDF、Word、PPT、網頁截圖、甚至掃描的紙本文件。",[15,138,139],{},"AI 確實能「讀」這些格式，但讀的品質天差地遠：",[46,141,142,155],{},[49,143,144],{},[52,145,146,149,152],{},[55,147,148],{},"格式",[55,150,151],{},"AI 理解度",[55,153,154],{},"常見問題",[65,156,157,170,183,196,209],{},[52,158,159,164,167],{},[70,160,161],{},[28,162,163],{},"結構化 Markdown \u002F 純文字",[70,165,166],{},"95%+",[70,168,169],{},"幾乎沒有",[52,171,172,177,180],{},[70,173,174],{},[28,175,176],{},"乾淨的網頁 HTML",[70,178,179],{},"85-90%",[70,181,182],{},"導航列、頁尾被當成內容",[52,184,185,190,193],{},[70,186,187],{},[28,188,189],{},"排版簡單的 PDF",[70,191,192],{},"70-80%",[70,194,195],{},"表格錯亂、分頁斷句",[52,197,198,203,206],{},[70,199,200],{},[28,201,202],{},"複雜排版的 PDF \u002F PPT",[70,204,205],{},"40-60%",[70,207,208],{},"欄位混淆、圖文分離",[52,210,211,216,219],{},[70,212,213],{},[28,214,215],{},"掃描圖片 \u002F 截圖",[70,217,218],{},"20-50%",[70,220,221],{},"OCR 錯字、完全丟失結構",[15,223,224],{},"一家餐廳把菜單的 PDF 丟進知識庫，結果 AI 把「招牌牛肉麵 $180」讀成「招牌牛肉 麵 $18 0」，客戶問價格，AI 回答 18 元。",[15,226,227],{},[28,228,229],{},"不是 AI 的錯。是你給它的格式讓它讀錯了。",[15,231,232,233,236],{},"最低成本的改善：把核心資料轉成",[28,234,235],{},"純文字或 Markdown","。不用全部轉，先轉最常被問到的前 20 個問題涵蓋的內容就好。這一步通常花半天到一天，但可以讓正確率從 40% 跳到 75%。",[130,238,240],{"id":239},"維度二粒度一份文件回答一個問題","維度二：粒度——一份文件回答一個問題",[15,242,243],{},"這是最反直覺的一點。",[15,245,246],{},"很多公司的做法：把整本產品手冊（200 頁）丟進知識庫，覺得「反正 AI 會自己找」。",[15,248,249,250,253,254,31],{},"問題是 AI 找答案的方式叫 ",[28,251,252],{},"RAG（Retrieval-Augmented Generation）","——它不是讀完整本書再回答你，而是",[28,255,256],{},"先搜尋最相關的幾個段落，只把這幾段送進 AI 模型",[15,258,259],{},"搜尋的單位越大，精準度越低。想像你在圖書館找一個食譜：",[261,262,263,267,270],"ul",{},[264,265,266],"li",{},"搜尋單位是「整本書」→ 找到《中華料理大全》，但你要的「番茄炒蛋」在第 387 頁，AI 可能沒看到那頁",[264,268,269],{},"搜尋單位是「一個章節」→ 找到「家常菜」章節，範圍小多了",[264,271,272],{},"搜尋單位是「一道菜的食譜」→ 直接命中",[15,274,275],{},[28,276,277],{},"知識庫的理想粒度：一個段落回答一個具體問題。",[15,279,280],{},"實務上怎麼做：",[282,283,284,290,296],"ol",{},[264,285,286,289],{},[28,287,288],{},"把大文件拆成小段落","。每段 200-500 字，圍繞一個具體主題。",[264,291,292,295],{},[28,293,294],{},"每段加上標題","。標題要像客戶會問的問題：「A 方案和 B 方案的差別」比「產品比較表」好。",[264,297,298,301],{},[28,299,300],{},"重複的背景資訊不要刪","。每段都應該能獨立被理解，不能假設讀者看過前面的內容。",[15,303,304,305,31],{},"一家保險公司把 80 頁的保單條款拆成 150 個「一問一答」段落後，AI 的正確回答率從 38% 跳到 79%。",[28,306,307],{},"工作量大概是兩個人花三天",[130,309,311],{"id":310},"維度三更新頻率過期的答案比沒有答案更危險","維度三：更新頻率——過期的答案比沒有答案更危險",[15,313,314],{},"這是最多人忽略的。",[15,316,317],{},"知識庫建好那天，內容是對的。三個月後呢？價格改了、方案調了、新產品上了、舊功能下架了。",[15,319,320],{},[28,321,322],{},"AI 不知道資料過期了。它會用過期的資訊，很有自信地回答客戶。",[15,324,325],{},"一個真實場景：某公司去年底調了價格，但知識庫裡還是舊價格。AI 告訴客戶「這個方案每月 $990」，客戶下單後發現實際收費 $1,290。客訴、退款、信任崩塌——全部因為一個沒更新的數字。",[15,327,328,331],{},[28,329,330],{},"最低限度的防護","：",[282,333,334,340,346],{},[264,335,336,339],{},[28,337,338],{},"設定更新週期","。核心資料（價格、方案、政策）至少每月檢查一次。",[264,341,342,345],{},[28,343,344],{},"加上時效標記","。在知識庫段落裡寫「此價格更新於 2026 年 4 月」，AI 被問到時至少可以提醒客戶「這是 X 月的資訊，建議確認最新價格」。",[264,347,348,351,352,355],{},[28,349,350],{},"有人負責","。不是 IT 部門的事，是",[28,353,354],{},"最了解業務變動的那個人","的事。通常是業務主管或客服主管。",[10,357,359],{"id":358},"快速自檢清單你的知識庫及格嗎","快速自檢清單：你的知識庫及格嗎？",[15,361,362],{},"在花錢升級模型或換廠商之前，先跑這五題：",[46,364,365,378],{},[49,366,367],{},[52,368,369,372,375],{},[55,370,371],{},"#",[55,373,374],{},"問題",[55,376,377],{},"及格標準",[65,379,380,391,402,413,424],{},[52,381,382,385,388],{},[70,383,384],{},"1",[70,386,387],{},"你的核心資料是什麼格式？",[70,389,390],{},"至少前 20 個常見問題的答案是純文字或 Markdown",[52,392,393,396,399],{},[70,394,395],{},"2",[70,397,398],{},"知識庫的最小單位是什麼？",[70,400,401],{},"段落級（200-500 字），不是整份文件",[52,403,404,407,410],{},[70,405,406],{},"3",[70,408,409],{},"每個段落有標題嗎？",[70,411,412],{},"有，而且像客戶會問的問題",[52,414,415,418,421],{},[70,416,417],{},"4",[70,419,420],{},"上次更新是什麼時候？",[70,422,423],{},"三個月內",[52,425,426,429,432],{},[70,427,428],{},"5",[70,430,431],{},"誰負責更新？",[70,433,434],{},"有具體的人名，不是「IT 部門」",[15,436,437,438,441],{},"五題裡如果有三題不及格，",[28,439,440],{},"你的 AI 客服問題八成不在模型，在知識庫","。換更貴的模型不會變好，整理資料才會。",[10,443,444],{"id":444},"我怎麼幫客戶做這件事",[15,446,447],{},"通常分三步：",[15,449,450],{},[28,451,452],{},"第一步：盤點（免費，聊 LINE 就行）",[15,454,455],{},"你把現有的資料丟給我看——可能是一堆 PDF、一個 Google Drive、或者你的官網。我幫你快速評估：哪些可以直接用、哪些要轉格式、哪些要拆段、哪些根本不該放進知識庫。",[15,457,458],{},"這一步不用花錢，通常十分鐘就有初步結論。",[15,460,461],{},[28,462,463],{},"第二步：整理（最花時間，但最值得）",[15,465,466],{},"根據盤點結果，把核心資料轉成 AI 能精準讀取的格式。這步可以你自己做（我教你方法），也可以我來做。",[15,468,469],{},"通常 20-30 個核心主題的整理，工期大概 2-5 天。",[15,471,472],{},[28,473,474],{},"第三步：接上 AI 並持續校準",[15,476,477],{},"整理好的知識庫接上 LINE Bot，跑一週的真實對話，看哪些問題還是答不好，再針對性補資料。",[15,479,480,483],{},[28,481,482],{},"大部分客戶在第二步就會感受到巨大差異","——同一個 AI 系統，換了知識庫，突然變聰明了。不是魔法，是你終於給了它正確的原料。",[10,485,487],{"id":486},"你的-ai-客服笨可能只是餓了","你的 AI 客服「笨」，可能只是餓了",[15,489,490],{},"AI 模型是引擎，知識庫是燃料。你不會因為車跑不動就換引擎——先看油箱裡裝的是不是對的油。",[15,492,493],{},"如果你已經有 AI 客服但效果不好，或者正在評估要不要做——加 LINE 跟我聊，把你目前的資料狀況大概描述一下。我可以幫你判斷問題在哪一層，以及最小成本的改善方案是什麼。",[15,495,496,497,31],{},"大部分情況下，",[28,498,499],{},"花三天整理資料，效果比花三萬換模型好",{"title":501,"searchDepth":502,"depth":502,"links":503},"",2,[504,505,506,512,513,514],{"id":12,"depth":502,"text":13},{"id":37,"depth":502,"text":38},{"id":127,"depth":502,"text":128,"children":507},[508,510,511],{"id":132,"depth":509,"text":133},3,{"id":239,"depth":509,"text":240},{"id":310,"depth":509,"text":311},{"id":358,"depth":502,"text":359},{"id":444,"depth":502,"text":444},{"id":486,"depth":502,"text":487},"2026-04-13","同一個 AI 模型，接上不同知識庫，回答品質可以差五倍。問題不在模型笨，在你餵的資料亂。從格式、粒度、更新頻率三個維度，拆解怎麼讓 AI 客服真的能用。","md",{},true,"\u002Fblog\u002Fai-chatbot-knowledge-base-quality",{"title":5,"description":516},"blog\u002Fai-chatbot-knowledge-base-quality",[524,525,526],"rag","ai","smb","_Zrcn6X5T4iwLmtFIO0veBOtBSi5mqfCItgJREjwH_o",1785903592272]