> 歡迎訂閱我的 [AI Engineer 電子報](https://aihao.eo.page/6tcs9) 和瀏覽 [[Generative AI Engineer 知識庫]] 原版: https://www.deeplearning.ai/short-courses/langchain-chat-with-your-data/ 簡中翻譯版: https://www.youtube.com/playlist?list=PLiuLMb-dLdWJX_EWk4RtQAjjrPLWaswTV 課程心得: * 蠻有收穫的，對於 RAG 元件有更深入的內容 * 包括各種 Splitting, 各種 Retrieval 以及可傳入 chat history 的 ConversationalRetrievalChain 的運作(跟我原先直覺想的運作不一樣，竟然是個 chaining) * langchain 的範例 code 也有簡單易懂 ## Introduction ![[Pasted image 20230818220143.png]] ![[Pasted image 20230818220153.png]] ![[Pasted image 20230818220209.png]] ## Document Loading ![[Pasted image 20230921145750.png]] ![[Pasted image 20230921145821.png]] ![[Pasted image 20230921150731.png]] ![[Pasted image 20230921150751.png]] ![[Pasted image 20230921151332.png]] ![[Pasted image 20230921151341.png]] ![[Pasted image 20230921151356.png]] ## Document Splitting ![[Pasted image 20230921151432.png]] ![[Pasted image 20230921151543.png]] ![[Pasted image 20230921151641.png]] ![[Pasted image 20230921160202.png]] ![[Pasted image 20230921160212.png]] Recursive 預設會先用 \n\n 切，然後再用 \n 切..... 依此類推 ![[Pasted image 20230921160234.png]] ![[Pasted image 20230921160323.png]] ## Vector stores and Embedding ![[Pasted image 20230921160509.png]] ![[Pasted image 20230921160558.png]] ![[Pasted image 20230921160700.png]] ![[Pasted image 20230921160720.png]] ![[Pasted image 20230921160754.png]] Lecture01.pdf 載入兩次，是因為想要示範模擬髒資料 ![[Pasted image 20230921160829.png]] ![[Pasted image 20230921160843.png]] ![[Pasted image 20230921160858.png]] ### Failure modes 失敗的場景 ![[Pasted image 20230921162235.png]] 前兩個回傳是重複的，因為我們之前有重複載入 ![[Pasted image 20230921162345.png]] 我們只想要第三章的 regression 資料，但是 embedding 無法捕捉結構化資料來過濾 ## Retrieval 檢索 ![[Pasted image 20230921162555.png]] ![[Pasted image 20230921163131.png]] MMR 最大邊際相關性 我們想要有多樣性 diverse ![[Pasted image 20230921163039.png]] * 先撈比較大的 fetch_k * 從從 fetch_k 中，找較有多樣性的子集 k * 這招就可以避免有重複的資料 ![[Pasted image 20230921163355.png]] self-query 自查詢: ![[Pasted image 20230921163649.png]] 多一步 Compression LLM 來縮減內容 (縮減 chunk 內容，不是合併chunks)。此招成本會增加。 https://python.langchain.com/docs/modules/data_connection/retrievers/contextual_compression/ LLMChainExtractor ![[Pasted image 20230921170148.png]] ![[Pasted image 20230921170159.png]] ![[Pasted image 20230921170219.png]] ![[Pasted image 20230921170229.png]] ![[Pasted image 20230921170257.png]] ![[Pasted image 20230921170303.png]] ![[Pasted image 20230921170316.png]] ![[Pasted image 20230921170323.png]] ![[Pasted image 20230921165847.png]] ![[Pasted image 20230921170402.png]] ![[Pasted image 20230921170410.png]] TF-IDF 的效果就沒這麼好 最後，作者最喜歡 self-query 這招，建議建立複雜的 metadata 來做過濾 你甚至可以可用 LLM 來推導出 metadata 用來過濾，甚至是 nested metadata ## Question Answering ![[Pasted image 20230921170601.png]] ![[Pasted image 20230921170611.png]] ![[Pasted image 20230921170627.png]] 若相關資料很多時，有這三種方式可以處理 refine 效果好，但是慢 ![[Pasted image 20230921173223.png]] ![[Pasted image 20230921173234.png]] ![[Pasted image 20230921173243.png]] ![[Pasted image 20230921173253.png]] ![[Pasted image 20230921173259.png]] ![[Pasted image 20230921173351.png]] ![[Pasted image 20230921173306.png]] 沒有 conversational history，下一章講 ## Chat 建構完整的問答聊天機器人 ![[Pasted image 20230921173943.png]] ![[Pasted image 20230921174023.png]] ![[Pasted image 20230921174853.png]] ![[Pasted image 20230921174904.png]] ![[Pasted image 20230921174913.png]] ![[Pasted image 20230921174932.png]] ![[Pasted image 20230921174946.png]] 根據對話歷史紀錄，以及用戶的新問題，重寫一個新問題 (chaining prompt) 再用新問題去 RAG > 這篇 https://txt.cohere.com/rag-chatbot/ 有談到為何需要根據對話紀錄紀錄，重新產生 RAG 用的 query ![[Pasted image 20230921175008.png]] ![[Pasted image 20230921175536.png]] 注意: 這裡沒有建立 Memory，因為 chat history 在 GUI 中處理了 稍後在呼叫 qa 時，需要手動傳入 chat_history ![[Pasted image 20230921175557.png]] ![[Pasted image 20230921175611.png]] ![[Pasted image 20230921175632.png]] ![[Pasted image 20230921174048.png]]