沈書禾（River Shen）的個人網站。東吳大學資訊管理學系，做 RAG、資訊檢索、後端與系統設計。這裡有我的研究、專案、文章，還有很主觀的 F1。https://rivershen.com/https://rivershen.com/posts/hello-world/https://rivershen.com/posts/hello-world/終於有一個放東西的地方了。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/學技術時留下來的筆記。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/從「為什麼需要 RAG」到「怎麼評估和改進」，一章一章整理我學 RAG 的筆記。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/01-%E6%AA%A2%E7%B4%A2%E5%A2%9E%E5%BC%B7%E7%94%9F%E6%88%90%E7%B0%A1%E4%BB%8B/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/01-%E6%AA%A2%E7%B4%A2%E5%A2%9E%E5%BC%B7%E7%94%9F%E6%88%90%E7%B0%A1%E4%BB%8B/先搞懂 RAG 要解決什麼問題、它怎麼運作，以及它跟微調差在哪。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/01-%E6%AA%A2%E7%B4%A2%E5%A2%9E%E5%BC%B7%E7%94%9F%E6%88%90%E7%B0%A1%E4%BB%8B/1-1-%E6%A8%99%E6%BA%96%E5%A4%A7%E5%9E%8B%E8%AA%9E%E8%A8%80%E6%A8%A1%E5%9E%8B%E7%9A%84%E5%B1%80%E9%99%90%E6%80%A7/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/01-%E6%AA%A2%E7%B4%A2%E5%A2%9E%E5%BC%B7%E7%94%9F%E6%88%90%E7%B0%A1%E4%BB%8B/1-1-%E6%A8%99%E6%BA%96%E5%A4%A7%E5%9E%8B%E8%AA%9E%E8%A8%80%E6%A8%A1%E5%9E%8B%E7%9A%84%E5%B1%80%E9%99%90%E6%80%A7/LLM 能生成流暢文字，但仍有知識截止、幻覺、專業知識不足與來源難以追溯等問題。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/01-%E6%AA%A2%E7%B4%A2%E5%A2%9E%E5%BC%B7%E7%94%9F%E6%88%90%E7%B0%A1%E4%BB%8B/1-2-%E4%BB%80%E9%BA%BC%E6%98%AF%E6%AA%A2%E7%B4%A2%E5%A2%9E%E5%BC%B7%E7%94%9F%E6%88%90/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/01-%E6%AA%A2%E7%B4%A2%E5%A2%9E%E5%BC%B7%E7%94%9F%E6%88%90%E7%B0%A1%E4%BB%8B/1-2-%E4%BB%80%E9%BA%BC%E6%98%AF%E6%AA%A2%E7%B4%A2%E5%A2%9E%E5%BC%B7%E7%94%9F%E6%88%90/RAG 的定義，以及一次問答從頭到尾經過哪三個步驟。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/01-%E6%AA%A2%E7%B4%A2%E5%A2%9E%E5%BC%B7%E7%94%9F%E6%88%90%E7%B0%A1%E4%BB%8B/1-3-rag%E7%B3%BB%E7%B5%B1%E7%9A%84%E6%A0%B8%E5%BF%83%E6%9E%B6%E6%A7%8B/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/01-%E6%AA%A2%E7%B4%A2%E5%A2%9E%E5%BC%B7%E7%94%9F%E6%88%90%E7%B0%A1%E4%BB%8B/1-3-rag%E7%B3%BB%E7%B5%B1%E7%9A%84%E6%A0%B8%E5%BF%83%E6%9E%B6%E6%A7%8B/一個 RAG 系統由檢索器和生成器組成，再加上事先準備好的知識庫。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/01-%E6%AA%A2%E7%B4%A2%E5%A2%9E%E5%BC%B7%E7%94%9F%E6%88%90%E7%B0%A1%E4%BB%8B/1-4-rag%E5%B0%8D%E6%AF%94%E5%BE%AE%E8%AA%BF/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/01-%E6%AA%A2%E7%B4%A2%E5%A2%9E%E5%BC%B7%E7%94%9F%E6%88%90%E7%B0%A1%E4%BB%8B/1-4-rag%E5%B0%8D%E6%AF%94%E5%BE%AE%E8%AA%BF/兩種都能讓模型更適合你的用途，但解決的是不同的問題。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/01-%E6%AA%A2%E7%B4%A2%E5%A2%9E%E5%BC%B7%E7%94%9F%E6%88%90%E7%B0%A1%E4%BB%8B/1-5-%E4%BD%BF%E7%94%A8rag%E7%9A%84%E5%84%AA%E5%8B%A2/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/01-%E6%AA%A2%E7%B4%A2%E5%A2%9E%E5%BC%B7%E7%94%9F%E6%88%90%E7%B0%A1%E4%BB%8B/1-5-%E4%BD%BF%E7%94%A8rag%E7%9A%84%E5%84%AA%E5%8B%A2/RAG 好在哪裡，以及它不是萬靈丹的地方。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/02-%E6%AA%A2%E7%B4%A2%E7%B5%84%E4%BB%B6/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/02-%E6%AA%A2%E7%B4%A2%E7%B5%84%E4%BB%B6/文字怎麼變成向量、怎麼找出意思最接近的內容，以及向量要存在哪裡。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/02-%E6%AA%A2%E7%B4%A2%E7%B5%84%E4%BB%B6/2-1-%E6%AA%A2%E7%B4%A2%E5%99%A8%E5%9C%A8rag%E4%B8%AD%E7%9A%84%E4%BD%9C%E7%94%A8/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/02-%E6%AA%A2%E7%B4%A2%E7%B5%84%E4%BB%B6/2-1-%E6%AA%A2%E7%B4%A2%E5%99%A8%E5%9C%A8rag%E4%B8%AD%E7%9A%84%E4%BD%9C%E7%94%A8/檢索器負責把「最有用的幾段資料」交給模型，它找得準不準，決定了整個 RAG 的上限。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/02-%E6%AA%A2%E7%B4%A2%E7%B5%84%E4%BB%B6/2-2-%E5%90%91%E9%87%8F%E5%B5%8C%E5%85%A5%E7%B0%A1%E4%BB%8B/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/02-%E6%AA%A2%E7%B4%A2%E7%B5%84%E4%BB%B6/2-2-%E5%90%91%E9%87%8F%E5%B5%8C%E5%85%A5%E7%B0%A1%E4%BB%8B/嵌入（embedding）把一段文字變成一串數字，讓電腦可以比較意思像不像。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/02-%E6%AA%A2%E7%B4%A2%E7%B5%84%E4%BB%B6/2-3-%E5%B8%B8%E7%94%A8%E5%B5%8C%E5%85%A5%E6%A8%A1%E5%9E%8B/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/02-%E6%AA%A2%E7%B4%A2%E7%B5%84%E4%BB%B6/2-3-%E5%B8%B8%E7%94%A8%E5%B5%8C%E5%85%A5%E6%A8%A1%E5%9E%8B/嵌入模型有雲端 API 和開源模型兩大類，選的時候要看語言、成本和部署方式。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/02-%E6%AA%A2%E7%B4%A2%E7%B5%84%E4%BB%B6/2-4-%E7%9B%B8%E4%BC%BC%E6%80%A7%E6%90%9C%E7%B4%A2/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/02-%E6%AA%A2%E7%B4%A2%E7%B5%84%E4%BB%B6/2-4-%E7%9B%B8%E4%BC%BC%E6%80%A7%E6%90%9C%E7%B4%A2/兩個向量有多像、怎麼量，以及資料很多時怎麼快速找到最接近的那幾個。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/02-%E6%AA%A2%E7%B4%A2%E7%B5%84%E4%BB%B6/2-5-%E5%90%91%E9%87%8F%E8%B3%87%E6%96%99%E5%BA%AB%E7%B0%A1%E4%BB%8B/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/02-%E6%AA%A2%E7%B4%A2%E7%B5%84%E4%BB%B6/2-5-%E5%90%91%E9%87%8F%E8%B3%87%E6%96%99%E5%BA%AB%E7%B0%A1%E4%BB%8B/專門用來存向量、並且能快速找出最接近的那幾個的資料庫。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/02-%E6%AA%A2%E7%B4%A2%E7%B5%84%E4%BB%B6/2-6-%E9%81%B8%E6%93%87%E5%90%91%E9%87%8F%E8%B3%87%E6%96%99%E5%BA%AB%E8%80%83%E9%87%8F%E5%9B%A0%E7%B4%A0/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/02-%E6%AA%A2%E7%B4%A2%E7%B5%84%E4%BB%B6/2-6-%E9%81%B8%E6%93%87%E5%90%91%E9%87%8F%E8%B3%87%E6%96%99%E5%BA%AB%E8%80%83%E9%87%8F%E5%9B%A0%E7%B4%A0/沒有最好的向量資料庫，只有適合你的規模、預算和團隊的那一個。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/02-%E6%AA%A2%E7%B4%A2%E7%B5%84%E4%BB%B6/2-7-%E5%AF%A6%E8%B8%90%E7%94%9F%E6%88%90%E6%96%87%E5%AD%97%E5%B5%8C%E5%85%A5/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/02-%E6%AA%A2%E7%B4%A2%E7%B5%84%E4%BB%B6/2-7-%E5%AF%A6%E8%B8%90%E7%94%9F%E6%88%90%E6%96%87%E5%AD%97%E5%B5%8C%E5%85%A5/用一小段 Python，把句子轉成向量，並且實際算出哪些句子意思相近。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/03-%E6%BA%96%E5%82%99%E6%AA%A2%E7%B4%A2%E6%89%80%E9%9C%80%E8%B3%87%E6%96%99/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/03-%E6%BA%96%E5%82%99%E6%AA%A2%E7%B4%A2%E6%89%80%E9%9C%80%E8%B3%87%E6%96%99/把 PDF、網頁、文字檔，處理成檢索器找得到、找得準的一小塊一小塊資料。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/03-%E6%BA%96%E5%82%99%E6%AA%A2%E7%B4%A2%E6%89%80%E9%9C%80%E8%B3%87%E6%96%99/3-1-%E5%BE%9E%E4%B8%8D%E5%90%8C%E4%BE%86%E6%BA%90%E8%BC%89%E5%85%A5%E6%96%87%E4%BB%B6/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/03-%E6%BA%96%E5%82%99%E6%AA%A2%E7%B4%A2%E6%89%80%E9%9C%80%E8%B3%87%E6%96%99/3-1-%E5%BE%9E%E4%B8%8D%E5%90%8C%E4%BE%86%E6%BA%90%E8%BC%89%E5%85%A5%E6%96%87%E4%BB%B6/知識庫的資料有各種格式，第一步是把它們都讀成乾淨的純文字，並且記得它是從哪來的。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/03-%E6%BA%96%E5%82%99%E6%AA%A2%E7%B4%A2%E6%89%80%E9%9C%80%E8%B3%87%E6%96%99/3-2-%E6%96%87%E4%BB%B6%E5%88%86%E5%A1%8A%E7%9A%84%E5%BF%85%E8%A6%81%E6%80%A7/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/03-%E6%BA%96%E5%82%99%E6%AA%A2%E7%B4%A2%E6%89%80%E9%9C%80%E8%B3%87%E6%96%99/3-2-%E6%96%87%E4%BB%B6%E5%88%86%E5%A1%8A%E7%9A%84%E5%BF%85%E8%A6%81%E6%80%A7/為什麼不能把整份文件當成一筆資料，以及塊的大小怎麼取捨。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/03-%E6%BA%96%E5%82%99%E6%AA%A2%E7%B4%A2%E6%89%80%E9%9C%80%E8%B3%87%E6%96%99/3-3-%E5%9B%BA%E5%AE%9A%E5%A4%A7%E5%B0%8F%E5%88%86%E5%A1%8A%E7%AD%96%E7%95%A5/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/03-%E6%BA%96%E5%82%99%E6%AA%A2%E7%B4%A2%E6%89%80%E9%9C%80%E8%B3%87%E6%96%99/3-3-%E5%9B%BA%E5%AE%9A%E5%A4%A7%E5%B0%8F%E5%88%86%E5%A1%8A%E7%AD%96%E7%95%A5/最簡單的切法：每隔固定的字數切一刀，相鄰的塊再重疊一點。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/03-%E6%BA%96%E5%82%99%E6%AA%A2%E7%B4%A2%E6%89%80%E9%9C%80%E8%B3%87%E6%96%99/3-4-%E5%85%A7%E5%AE%B9%E6%84%9F%E7%9F%A5%E7%9A%84%E5%88%86%E5%A1%8A%E6%96%B9%E6%B3%95/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/03-%E6%BA%96%E5%82%99%E6%AA%A2%E7%B4%A2%E6%89%80%E9%9C%80%E8%B3%87%E6%96%99/3-4-%E5%85%A7%E5%AE%B9%E6%84%9F%E7%9F%A5%E7%9A%84%E5%88%86%E5%A1%8A%E6%96%B9%E6%B3%95/順著文件本來的結構來切，讓每一塊都是一個完整的意思。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/03-%E6%BA%96%E5%82%99%E6%AA%A2%E7%B4%A2%E6%89%80%E9%9C%80%E8%B3%87%E6%96%99/3-5-%E8%B3%87%E6%96%99%E5%A1%8A%E7%9A%84%E5%85%83%E8%B3%87%E6%96%99%E9%97%9C%E8%81%AF/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/03-%E6%BA%96%E5%82%99%E6%AA%A2%E7%B4%A2%E6%89%80%E9%9C%80%E8%B3%87%E6%96%99/3-5-%E8%B3%87%E6%96%99%E5%A1%8A%E7%9A%84%E5%85%83%E8%B3%87%E6%96%99%E9%97%9C%E8%81%AF/每一塊除了內容，還要記得它從哪來、屬於什麼，後面的過濾、引用、權限都靠它。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/03-%E6%BA%96%E5%82%99%E6%AA%A2%E7%B4%A2%E6%89%80%E9%9C%80%E8%B3%87%E6%96%99/3-6-%E5%B0%87%E8%99%95%E7%90%86%E5%BE%8C%E7%9A%84%E8%B3%87%E6%96%99%E5%AD%98%E5%84%B2%E5%88%B0%E5%90%91%E9%87%8F%E8%B3%87%E6%96%99%E5%BA%AB/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/03-%E6%BA%96%E5%82%99%E6%AA%A2%E7%B4%A2%E6%89%80%E9%9C%80%E8%B3%87%E6%96%99/3-6-%E5%B0%87%E8%99%95%E7%90%86%E5%BE%8C%E7%9A%84%E8%B3%87%E6%96%99%E5%AD%98%E5%84%B2%E5%88%B0%E5%90%91%E9%87%8F%E8%B3%87%E6%96%99%E5%BA%AB/把每一塊的向量、原文和 metadata 一起存進去，以及資料更新時怎麼處理。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/03-%E6%BA%96%E5%82%99%E6%AA%A2%E7%B4%A2%E6%89%80%E9%9C%80%E8%B3%87%E6%96%99/3-7-%E5%8B%95%E6%89%8B%E5%AF%A6%E8%B8%90%E6%96%87%E4%BB%B6%E5%88%86%E5%A1%8A/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/03-%E6%BA%96%E5%82%99%E6%AA%A2%E7%B4%A2%E6%89%80%E9%9C%80%E8%B3%87%E6%96%99/3-7-%E5%8B%95%E6%89%8B%E5%AF%A6%E8%B8%90%E6%96%87%E4%BB%B6%E5%88%86%E5%A1%8A/用純 Python 實作三種分塊方法，並用同一份文件比較它們切出來的結果。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/04-%E7%94%9F%E6%88%90%E6%A8%A1%E7%B5%84%E8%88%87%E5%A2%9E%E5%BC%B7/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/04-%E7%94%9F%E6%88%90%E6%A8%A1%E7%B5%84%E8%88%87%E5%A2%9E%E5%BC%B7/檢索到資料之後，怎麼組成提示詞交給 LLM，並且讓答案有根據、標得出處。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/04-%E7%94%9F%E6%88%90%E6%A8%A1%E7%B5%84%E8%88%87%E5%A2%9E%E5%BC%B7/4-1-%E7%94%9F%E6%88%90%E5%99%A8llm%E5%9C%A8rag%E4%B8%AD%E7%9A%84%E4%BD%9C%E7%94%A8/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/04-%E7%94%9F%E6%88%90%E6%A8%A1%E7%B5%84%E8%88%87%E5%A2%9E%E5%BC%B7/4-1-%E7%94%9F%E6%88%90%E5%99%A8llm%E5%9C%A8rag%E4%B8%AD%E7%9A%84%E4%BD%9C%E7%94%A8/在 RAG 裡，LLM 的工作從「憑記憶回答」變成「根據資料整理出答案」。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/04-%E7%94%9F%E6%88%90%E6%A8%A1%E7%B5%84%E8%88%87%E5%A2%9E%E5%BC%B7/4-2-rag%E6%8F%90%E7%A4%BA%E8%A9%9E%E7%9A%84%E7%B5%90%E6%A7%8B%E5%8C%96/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/04-%E7%94%9F%E6%88%90%E6%A8%A1%E7%B5%84%E8%88%87%E5%A2%9E%E5%BC%B7/4-2-rag%E6%8F%90%E7%A4%BA%E8%A9%9E%E7%9A%84%E7%B5%90%E6%A7%8B%E5%8C%96/一份好的 RAG 提示詞有固定的結構：指示、參考資料、問題。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/04-%E7%94%9F%E6%88%90%E6%A8%A1%E7%B5%84%E8%88%87%E5%A2%9E%E5%BC%B7/4-3-%E4%B8%8A%E4%B8%8B%E6%96%87%E6%B3%A8%E5%85%A5%E6%96%B9%E6%B3%95/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/04-%E7%94%9F%E6%88%90%E6%A8%A1%E7%B5%84%E8%88%87%E5%A2%9E%E5%BC%B7/4-3-%E4%B8%8A%E4%B8%8B%E6%96%87%E6%B3%A8%E5%85%A5%E6%96%B9%E6%B3%95/把檢索到的資料放進提示詞，看起來只是貼上去，但放法和順序都有差。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/04-%E7%94%9F%E6%88%90%E6%A8%A1%E7%B5%84%E8%88%87%E5%A2%9E%E5%BC%B7/4-4-%E7%AE%A1%E7%90%86%E4%B8%8A%E4%B8%8B%E6%96%87%E9%95%B7%E5%BA%A6%E9%99%90%E5%88%B6/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/04-%E7%94%9F%E6%88%90%E6%A8%A1%E7%B5%84%E8%88%87%E5%A2%9E%E5%BC%B7/4-4-%E7%AE%A1%E7%90%86%E4%B8%8A%E4%B8%8B%E6%96%87%E9%95%B7%E5%BA%A6%E9%99%90%E5%88%B6/模型一次能讀的長度有限，資料放不下的時候該怎麼取捨。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/04-%E7%94%9F%E6%88%90%E6%A8%A1%E7%B5%84%E8%88%87%E5%A2%9E%E5%BC%B7/4-5-%E7%94%9F%E6%88%90%E6%9C%80%E7%B5%82%E5%9B%9E%E8%A6%86/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/04-%E7%94%9F%E6%88%90%E6%A8%A1%E7%B5%84%E8%88%87%E5%A2%9E%E5%BC%B7/4-5-%E7%94%9F%E6%88%90%E6%9C%80%E7%B5%82%E5%9B%9E%E8%A6%86/提示詞交給模型之後，怎麼控制輸出、讓答案穩定，以及拿到答案之後的處理。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/04-%E7%94%9F%E6%88%90%E6%A8%A1%E7%B5%84%E8%88%87%E5%A2%9E%E5%BC%B7/4-6-%E7%94%9F%E6%88%90%E5%85%A7%E5%AE%B9%E4%B8%AD%E7%9A%84%E4%BE%86%E6%BA%90%E6%AD%B8%E5%B1%AC/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/04-%E7%94%9F%E6%88%90%E6%A8%A1%E7%B5%84%E8%88%87%E5%A2%9E%E5%BC%B7/4-6-%E7%94%9F%E6%88%90%E5%85%A7%E5%AE%B9%E4%B8%AD%E7%9A%84%E4%BE%86%E6%BA%90%E6%AD%B8%E5%B1%AC/讓答案說明它是根據哪段資料，使用者才能查證，也才有辦法信任。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/05-%E6%90%AD%E5%BB%BA%E5%9F%BA%E6%9C%ACrag%E6%B5%81%E7%A8%8B/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/05-%E6%90%AD%E5%BB%BA%E5%9F%BA%E6%9C%ACrag%E6%B5%81%E7%A8%8B/把前四章學的東西接起來，做出一個從文件到答案、真的能跑的 RAG 系統。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/05-%E6%90%AD%E5%BB%BA%E5%9F%BA%E6%9C%ACrag%E6%B5%81%E7%A8%8B/5-1-rag%E6%A1%86%E6%9E%B6%E6%A6%82%E8%BF%B0/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/05-%E6%90%AD%E5%BB%BA%E5%9F%BA%E6%9C%ACrag%E6%B5%81%E7%A8%8B/5-1-rag%E6%A1%86%E6%9E%B6%E6%A6%82%E8%BF%B0/框架把載入、分塊、檢索、生成這些常用零件包好，但要不要用、什麼時候用，值得想一下。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/05-%E6%90%AD%E5%BB%BA%E5%9F%BA%E6%9C%ACrag%E6%B5%81%E7%A8%8B/5-2-%E7%92%B0%E5%A2%83%E9%85%8D%E7%BD%AE/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/05-%E6%90%AD%E5%BB%BA%E5%9F%BA%E6%9C%ACrag%E6%B5%81%E7%A8%8B/5-2-%E7%92%B0%E5%A2%83%E9%85%8D%E7%BD%AE/準備 Python 環境、安裝套件，以及怎麼安全地處理 API 金鑰。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/05-%E6%90%AD%E5%BB%BA%E5%9F%BA%E6%9C%ACrag%E6%B5%81%E7%A8%8B/5-3-%E5%AF%A6%E7%8F%BE%E6%AA%A2%E7%B4%A2%E5%99%A8/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/05-%E6%90%AD%E5%BB%BA%E5%9F%BA%E6%9C%ACrag%E6%B5%81%E7%A8%8B/5-3-%E5%AF%A6%E7%8F%BE%E6%AA%A2%E7%B4%A2%E5%99%A8/建立索引，並寫一個 retrieve 函式：問題進去，最相關的幾段資料出來。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/05-%E6%90%AD%E5%BB%BA%E5%9F%BA%E6%9C%ACrag%E6%B5%81%E7%A8%8B/5-4-%E5%AF%A6%E7%8F%BE%E7%94%9F%E6%88%90%E5%99%A8%E6%95%B4%E5%90%88/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/05-%E6%90%AD%E5%BB%BA%E5%9F%BA%E6%9C%ACrag%E6%B5%81%E7%A8%8B/5-4-%E5%AF%A6%E7%8F%BE%E7%94%9F%E6%88%90%E5%99%A8%E6%95%B4%E5%90%88/組好提示詞，接上 LLM，並處理沒有金鑰、呼叫失敗這些實際情況。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/05-%E6%90%AD%E5%BB%BA%E5%9F%BA%E6%9C%ACrag%E6%B5%81%E7%A8%8B/5-5-%E7%B5%90%E5%90%88%E6%AA%A2%E7%B4%A2%E8%88%87%E7%94%9F%E6%88%90/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/05-%E6%90%AD%E5%BB%BA%E5%9F%BA%E6%9C%ACrag%E6%B5%81%E7%A8%8B/5-5-%E7%B5%90%E5%90%88%E6%AA%A2%E7%B4%A2%E8%88%87%E7%94%9F%E6%88%90/把 retrieve、build_prompt、generate 串成一個 ask 函式，並處理「找不到資料」的情況。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/05-%E6%90%AD%E5%BB%BA%E5%9F%BA%E6%9C%ACrag%E6%B5%81%E7%A8%8B/5-6-%E5%9C%A8%E6%B5%81%E7%A8%8B%E4%B8%AD%E5%9F%B7%E8%A1%8C%E6%9F%A5%E8%A9%A2/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/05-%E6%90%AD%E5%BB%BA%E5%9F%BA%E6%9C%ACrag%E6%B5%81%E7%A8%8B/5-6-%E5%9C%A8%E6%B5%81%E7%A8%8B%E4%B8%AD%E5%9F%B7%E8%A1%8C%E6%9F%A5%E8%A9%A2/實際丟幾個問題進去，觀察每一步發生了什麼，以及門檻值調整的過程。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/05-%E6%90%AD%E5%BB%BA%E5%9F%BA%E6%9C%ACrag%E6%B5%81%E7%A8%8B/5-7-%E5%8B%95%E6%89%8B%E5%AF%A6%E8%B8%90%E7%AB%AF%E5%88%B0%E7%AB%AFrag%E7%B3%BB%E7%B5%B1/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/05-%E6%90%AD%E5%BB%BA%E5%9F%BA%E6%9C%ACrag%E6%B5%81%E7%A8%8B/5-7-%E5%8B%95%E6%89%8B%E5%AF%A6%E8%B8%90%E7%AB%AF%E5%88%B0%E7%AB%AFrag%E7%B3%BB%E7%B5%B1/一個約一百行的完整 RAG 程式，從文件到答案，可以直接複製執行。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/06-%E8%A9%95%E4%BC%B0%E8%88%87%E6%94%B9%E9%80%B2/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/06-%E8%A9%95%E4%BC%B0%E8%88%87%E6%94%B9%E9%80%B2/系統搭起來之後，怎麼知道它好不好、哪裡壞了，以及怎麼一步步改進。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/06-%E8%A9%95%E4%BC%B0%E8%88%87%E6%94%B9%E9%80%B2/6-1-%E8%A9%95%E4%BC%B0rag%E7%9A%84%E6%8C%91%E6%88%B0/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/06-%E8%A9%95%E4%BC%B0%E8%88%87%E6%94%B9%E9%80%B2/6-1-%E8%A9%95%E4%BC%B0rag%E7%9A%84%E6%8C%91%E6%88%B0/為什麼「感覺不錯」不夠，以及 RAG 的評估難在哪裡。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/06-%E8%A9%95%E4%BC%B0%E8%88%87%E6%94%B9%E9%80%B2/6-2-%E7%B5%84%E4%BB%B6%E5%B1%A4%E9%9D%A2%E8%A9%95%E4%BC%B0%E6%AA%A2%E7%B4%A2/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/06-%E8%A9%95%E4%BC%B0%E8%88%87%E6%94%B9%E9%80%B2/6-2-%E7%B5%84%E4%BB%B6%E5%B1%A4%E9%9D%A2%E8%A9%95%E4%BC%B0%E6%AA%A2%E7%B4%A2/用 Hit Rate、MRR 等指標，衡量檢索器有沒有把對的資料找出來。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/06-%E8%A9%95%E4%BC%B0%E8%88%87%E6%94%B9%E9%80%B2/6-3-%E7%B5%84%E4%BB%B6%E7%B4%9A%E5%88%A5%E8%A9%95%E4%BC%B0%E7%94%9F%E6%88%90/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/06-%E8%A9%95%E4%BC%B0%E8%88%87%E6%94%B9%E9%80%B2/6-3-%E7%B5%84%E4%BB%B6%E7%B4%9A%E5%88%A5%E8%A9%95%E4%BC%B0%E7%94%9F%E6%88%90/檢索找對了，模型有沒有照著資料回答、回答得切不切題。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/06-%E8%A9%95%E4%BC%B0%E8%88%87%E6%94%B9%E9%80%B2/6-4-%E7%AB%AF%E5%88%B0%E7%AB%AFrag%E8%A9%95%E4%BC%B0%E6%A1%86%E6%9E%B6/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/06-%E8%A9%95%E4%BC%B0%E8%88%87%E6%94%B9%E9%80%B2/6-4-%E7%AB%AF%E5%88%B0%E7%AB%AFrag%E8%A9%95%E4%BC%B0%E6%A1%86%E6%9E%B6/把檢索、生成放在一起看，建立一個可以重複執行的評估流程。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/06-%E8%A9%95%E4%BC%B0%E8%88%87%E6%94%B9%E9%80%B2/6-5-%E5%B8%B8%E8%A6%8B%E6%95%85%E9%9A%9C%E6%A8%A1%E5%BC%8F/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/06-%E8%A9%95%E4%BC%B0%E8%88%87%E6%94%B9%E9%80%B2/6-5-%E5%B8%B8%E8%A6%8B%E6%95%85%E9%9A%9C%E6%A8%A1%E5%BC%8F/RAG 常見的出錯方式，依流程的順序整理，以及怎麼判斷問題出在哪裡。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/06-%E8%A9%95%E4%BC%B0%E8%88%87%E6%94%B9%E9%80%B2/6-6-%E6%94%B9%E9%80%B2%E7%9A%84%E5%9F%BA%E6%9C%AC%E7%AD%96%E7%95%A5/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/06-%E8%A9%95%E4%BC%B0%E8%88%87%E6%94%B9%E9%80%B2/6-6-%E6%94%B9%E9%80%B2%E7%9A%84%E5%9F%BA%E6%9C%AC%E7%AD%96%E7%95%A5/從資料、檢索、提示詞各個環節，整理可以嘗試的改進方向，以及怎麼驗證有沒有用。Mon, 05 Oct 2026 00:00:00 GMThttps://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/06-%E8%A9%95%E4%BC%B0%E8%88%87%E6%94%B9%E9%80%B2/6-7-%E5%AF%A6%E8%B8%90%E5%88%86%E6%9E%90rag%E8%BC%B8%E5%87%BA%E5%93%81%E8%B3%AA/https://rivershen.com/posts/%E6%8A%80%E8%A1%93%E7%AD%86%E8%A8%98/rag/06-%E8%A9%95%E4%BC%B0%E8%88%87%E6%94%B9%E9%80%B2/6-7-%E5%AF%A6%E8%B8%90%E5%88%86%E6%9E%90rag%E8%BC%B8%E5%87%BA%E5%93%81%E8%B3%AA/用一份小測試集，實際計算檢索的 Hit Rate 和 MRR，並分析「找不到」門檻的取捨。Mon, 05 Oct 2026 00:00:00 GMT