{"id":12351,"date":"2024-07-26T15:12:43","date_gmt":"2024-07-26T07:12:43","guid":{"rendered":"https:\/\/ihower.tw\/blog\/?p=12351"},"modified":"2025-07-19T19:27:34","modified_gmt":"2025-07-19T11:27:34","slug":"aie-rag","status":"publish","type":"post","link":"https:\/\/ihower.tw\/blog\/12351-aie-rag","title":{"rendered":"\u611b\u597d AI Engineer \u96fb\u5b50\u5831 \ud83d\ude80\u00a0\u6aa2\u7d22\u589e\u5f37\u751f\u6210 RAG \u7279\u8f2f #15"},"content":{"rendered":"\n
\"\"<\/figure>\n\n\n\n

Hello! \u4f60\u597d \ud83d\udc4b<\/p>\n\n\n\n

\u6211\u662f ihower\uff0c\u9019\u96c6\u662f RAG \u9032\u968e\u7279\u8f2f\u3002\u5982\u679c\u4e0d\u77e5\u9053\u4ec0\u9ebc\u662f\u57fa\u672c\u7684 naive RAG(\u6216\u53eb\u505a top-k RAG) \uff0c\u53ef\u4ee5\u5148\u770b\u6211\u7684\u5165\u9580\u6295\u5f71\u7247<\/a>\u6709\u95dc RAG \u7684\u4ecb\u7d39\u3002\u7c21\u55ae\u8b1b\uff0cRAG \u5c31\u662f\u6839\u64da\u7528\u6236\u554f\u984c\uff0c\u6aa2\u7d22\u51fa\u6700\u76f8\u95dc\u7684\u5167\u5bb9\uff0c\u7136\u5f8c\u653e\u5230 Prompt \u88e1\u9762\u8b93\u6a21\u578b\u505a\u53c3\u8003\u56de\u7b54\u3002<\/p>\n\n\n\n\n\n\n\n

\ud83d\udd1d\u4f7f\u7528\u7e41\u9ad4\u4e2d\u6587\u505a Embedding \u6a21\u578b\u5927\u8a55\u6e2c<\/a><\/h3>\n\n\n\n

\u5728 RAG \u7cfb\u7d71\u4e2d\uff0c\u5c07\u6587\u5b57\u8f49\u8a9e\u610f\u5411\u91cf\u7684 embedding \u6a21\u578b\uff0c\u662f\u975e\u5e38\u91cd\u8981\u7684\u95dc\u9375\u6aa2\u7d22\u74b0\u7bc0\u3002
\u9019\u662f\u6211\u81ea\u5df1\u505a\u7684\u8a55\u6e2c\uff0c\u7e3d\u5171 33 \u500b\u6a21\u578b\u6bd4\u8f03\u6aa2\u7d22\u80fd\u529b\u3002\u9019\u88e1\u76f4\u63a5\u5148\u8aaa\u7d50\u8ad6:<\/p>\n\n\n\n

– \u8ffd\u6c42\u6700\u597d\u6027\u80fd\u8a66\u8a66 voyage-multilingual-2
– \u4fbf\u5b9c\u597d\u4e0a\u624b\u9078 OpenAI text-embedding-3-small<\/p>\n\n\n\n

\u5b8c\u6574\u8a55\u6e2c\u7d50\u679c\u548c\u5168\u6587\u8acb\u898b\u6211\u90e8\u843d\u683c\u6587\u7ae0\u3002<\/p>\n\n\n\n

\ud83d\udd1d\u4f7f\u7528\u7e41\u9ad4\u4e2d\u6587\u8a55\u6e2c\u5404\u5bb6 Reranker \u6a21\u578b\u7684\u91cd\u6392\u80fd\u529b<\/a><\/p>\n\n\n\n

\u63a5\u7e8c\u4e0a\u4e00\u7bc7 Embedding \u6a21\u578b\u8a55\u6e2c\uff0c\u9019\u4e5f\u662f\u6211\u81ea\u5df1\u91dd\u5c0d Reranker (\u91cd\u6392)\u6a21\u578b\uff0c\u505a\u6210\u4e8c\u968e\u6bb5\u6aa2\u7d22\u7684\u8a55\u6e2c\u3002<\/p>\n\n\n\n

\ud83d\udc4dAdvanced Retrieval for AI with Chroma \u8ab2\u7a0b<\/a><\/h3>\n\n\n\n

(\u6211\u6574\u7406\u7684\u4e0a\u8ab2\u622a\u5716\u548c\u7b46\u8a18<\/a>)<\/p>\n\n\n\n

DeepLearning.AI \u7684\u9019\u5802 RAG \u8ab2\u7a0b\u6211\u89ba\u5f97\u662f\u6bd4\u8f03\u63a8\u85a6\u7684\u597d\u4e0a\u624b\u5c0f\u8ab2\uff0c\u6559\u4e86 Query Expansion \u548c Re-Ranker\u3002<\/p>\n\n\n\n

\ud83e\udde0Systematically Improving RAG Applications \u6f14\u8b1b<\/a><\/h3>\n\n\n\n

(\u6211\u6574\u7406\u7684\u9010\u5b57\u7a3f\u548c\u622a\u5716<\/a>)<\/p>\n\n\n\n

\u8981\u5982\u4f55\u7cfb\u7d71\u6027\u5730\u6539\u9032\u60a8\u7684 RAG \u7cfb\u7d71? \u9019\u662f Jason Liu \u7684\u4e00\u5834\u6f14\u8b1b Systematically Improving RAG Applications \uff0c\u91cd\u9ede\u5982\u4e0b:<\/p>\n\n\n\n

* \u5efa\u7acb\u6709\u6548\u7684\u53cd\u994b\u6a5f\u5236: \u7d00\u9304\u7528\u6236\u8a55\u5206\u3001\u7d00\u9304\u76f8\u95dc\u6027(cosine \u6216\u662f reranker score)\u5206\u6578
* \u4e0d\u8981\u53ea\u653e\u6309\u8b9a\u5012\u8b9a\uff0c\u56e0\u70ba\u5c31\u7b97\u7b54\u6848\u6b63\u78ba\uff0c\u7528\u6236\u4e5f\u53ef\u80fd\u56e0\u70ba\u7b54\u6848\u592a\u9577\u6216\u592a\u77ed\u800c\u4e0d\u559c\u6b61\u3002\u8981\u554f “\u662f\u5426\u6709\u89e3\u6c7a\u60a8\u7684\u554f\u984c”? \u6bd4\u8f03\u597d
* \u7528 cosine \u6216\u662f reranker score \u5206\u6578\u53ef\u4ee5\u66f4\u5177\u6210\u672c\u6548\u76ca (\u76f8\u6bd4\u7528 LLM \u8a55\u4f30)
* \u91dd\u5c0d\u7528\u6236 query \u505a\u5206\u7fa4\uff0c\u627e\u51fa\u54ea\u4e9b\u4e3b\u984c\u662f\u7528\u6236\u6700\u5e38\u554f\u7684\u3001\u54ea\u4e9b\u554f\u6cd5(\u4f8b\u5982\u554f\u6bd4\u8f03\u3001\u554f\u6458\u8981)\u662f\u7528\u6236\u6700\u5e38\u554f\u7684
* \u5982\u6b64\u4f60\u6703\u6709\u4e09\u500b\u7dad\u5ea6\u53ef\u4ee5\u5206\u6790: \u554f\u984c\u983b\u6b21\u3001\u9ad8\u4f4e\u76f8\u95dc\u6027(\u53ef\u8a08\u7b97\u5e73\u5747\u9918\u5f26\u8ddd\u96e2)\u3001\u9ad8\u4f4e\u7528\u6236\u53cd\u994b
    * \u9ad8\u983b\u3001\u4f4e\u76f8\u95dc\u3001\u4e0d\u6eff\u610f: \u4fee retriever \u90e8\u5206
    * \u9ad8\u983b\u3001\u9ad8\u76f8\u95dc\u3001\u4e0d\u6eff\u610f: \u4fee generation \u90e8\u5206 
    * \u4f4e\u983b\u3001\u4e0d\u6eff\u610f: \u53ef\u4ee5\u5728 prompt \u4e2d\u544a\u8a34\u7528\u6236\u672c AI \u4e0d\u6703\u9019\u984c
* \u8fa8\u8b58\u51fa\u9ad8\u512a\u5148\u52a0\u5f37\u6539\u9032\u7684\u4e3b\u984c\uff0c\u53ef\u7528\u5408\u6210\u6578\u64da\u591a\u505a\u6e2c\u8a66
* \u6aa2\u7d22\u4e0d\u53ea\u9760 embedding\uff0c\u9084\u53ef\u4ee5\u591a\u505a\u65e5\u671f filter\u3001\u4e3b\u984c filter\u3001\u6587\u4ef6\u985e\u578b filter \u7b49\u7b49
* \u50b3\u7d71\u5168\u6587\u641c\u5c0b BM25\uff0c\u5728\u641c\u5c0b\u8005\u5c31\u662f\u6587\u4ef6\u4f5c\u8005\u7684\u60c5\u6cc1\uff0c\u6703\u6bd4\u8a9e\u610f\u641c\u5c0b\u66f4\u512a\u8d8a\uff0c\u56e0\u70ba\u4f5c\u8005\u77e5\u9053\u95dc\u9375\u5b57\u3001\u751a\u81f3\u77e5\u9053\u6a94\u6848\u540d\u7a31\uff0c\u5efa\u8b70\u5305\u62ec BM25 \u4e00\u8d77\u505a hybrid
* \u53ea\u8981\u6709\u6578\u5343\u500b\u554f\u7b54\u8cc7\u6599\uff0c\u5c31\u53ef\u4ee5\u505a\u81ea\u5df1\u7684\u5fae\u8abf embedding \u6a21\u578b\uff0c\u6027\u8cea\u8d85\u8d8a openai \u548c cohere
* RAG \u5c31\u50cf\u662f\u8a2d\u8a08\u4e00\u500b\u63a8\u85a6\u7cfb\u7d71<\/p>\n\n\n\n

\u9664\u4e86\u9304\u5f71\uff0c\u4f5c\u8005\u4e5f\u5beb\u6210\u6587\u7ae0\u53ef\u4ee5\u4e00\u770b Systematically Improving Your RAG<\/a><\/p>\n\n\n\n

\ud83c\udfafBeyond the Basics of RAG \u6f14\u8b1b<\/a><\/h3>\n\n\n\n

(\u6211\u6574\u7406\u7684\u9010\u5b57\u7a3f\u548c\u622a\u5716<\/a>)<\/p>\n\n\n\n

\u9019\u662f\u7531 Parlance Labs \u4e3b\u8fa6\u7684 RAG \u8ab2\u7a0b\u6f14\u8b1b\uff0c\u5167\u5bb9\u5305\u62ec Bi-encoder \u505a\u5411\u91cf\u641c\u5c0b\u3001Cross-encoder \u505a Reranker\u3001\u7d50\u5408 TF-IDF\/BM25 \u505a\u95dc\u9375\u8a5e\u641c\u7d22\u3001Metadata Filtering\u3001\u9577\u4e0a\u4e0b\u6587\u7a97\u53e3\u5c0d RAG \u7684\u5f71\u97ff\u3001\u5d4c\u5165\u6a21\u578b\u7684\u5fae\u8abf\u3001\u8655\u7406\u9577\u6587\u6a94\u7684\u7b56\u7565\u7b49\u7b49\u3002<\/p>\n\n\n\n

\ud83c\udf83Back to Basics for RAG \u6f14\u8b1b<\/a><\/h3>\n\n\n\n

(\u6211\u6574\u7406\u7684\u9010\u5b57\u7a3f\u548c\u622a\u5716<\/a>)<\/p>\n\n\n\n

\u4e5f\u662f Parlance Labs \u4e3b\u8fa6\u7684 RAG \u8ab2\u7a0b\u6f14\u8b1b\uff0c\u9019\u5834\u8457\u91cd\u5728\u8a55\u4f30\u6307\u6a19\u3002<\/p>\n\n\n\n

\ud83d\udea7Build a search engine, not a vector DB<\/a><\/h3>\n\n\n\n

\u6211\u5f88\u559c\u6b61\u9019\u500b\u6a19\u984c\uff0cRAG \u9700\u8981\u7684\u6aa2\u7d22\u6280\u8853\uff0c\u4e0d\u53ea\u662f Embedding \u5411\u91cf\u641c\u5c0b\uff0c\u800c\u662f\u5728\u6253\u9020\u4e00\u500b\u641c\u5c0b\u63a8\u85a6\u5f15\u64ce\u3002<\/p>\n\n\n\n

\ud83d\udc4aThe Basics of AI-Powered (Vector) Search<\/a><\/h3>\n\n\n\n

\u9019\u7bc7\u6587\u7ae0\u8a73\u8ff0\u76ee\u524d\u958b\u767c RAG \u9700\u8981\u7684\u641c\u5c0b\u7406\u8ad6\u77e5\u8b58\uff0c\u5305\u62ec\u50b3\u7d71\u7684 BM25 \u641c\u5c0b\u3001Bi-encoder\u3001Cross-encoder\u3001\u4e8c\u968e\u6bb5\u6aa2\u7d22\u7b49\u7b49<\/p>\n\n\n\n

\ud83c\udf0bWhat AI Engineers Should Know about Search<\/a><\/h3>\n\n\n\n

AI \u5de5\u7a0b\u5e2b\u61c9\u8a72\u4e86\u89e3\u7684\u641c\u5c0b\u6280\u8853\u77e5\u8b58 50 \u689d\uff0c\u7a81\u7136\u9700\u8981\u60e1\u88dc\u5f88\u591a\u641c\u7d22\u6280\u8853\u77e5\u8b58\uff0c\u6b61\u8fce\u4f86\u5230\u9019\u500b\u52c7\u6562\u7684\u4e16\u754c\u3002<\/p>\n\n\n\n

—-<\/p>\n\n\n\n

\u66f4\u591a\u95dc\u65bc RAG \u7684\u6280\u8853\u8cc7\u6599\uff0c\u53ef\u4ee5\u53c3\u8003\u6211\u7684 RAG \u7b46\u8a18 ihower.tw\/notes\/rag-guideline<\/a><\/p>\n\n\n\n

\u6700\u5f8c\u4f86\u5e6b\u81ea\u5df1 threads \u6dfb\u52a0\u4e00\u4e9b\u7c89\u7d72\u6578\uff0c\u5e33\u865f\u4e00\u6a23\u662f ihower<\/a>\uff0c\u6b61\u8fce\u8ffd\u8e64!<\/p>\n\n\n\n

\u2013 ihower<\/p>\n","protected":false},"excerpt":{"rendered":"

Hello! \u4f60\u597d \ud83d\udc4b \u6211\u662f ihower\uff0c\u9019\u96c6\u662f RAG \u9032\u968e\u7279\u8f2f\u3002\u5982\u679c\u4e0d\u77e5\u9053\u4ec0\u9ebc\u662f\u57fa\u672c\u7684 nai … <\/p>\n

\u95b1\u8b80\u5168\u6587\u3008\u611b\u597d AI Engineer \u96fb\u5b50\u5831 \ud83d\ude80\u00a0\u6aa2\u7d22\u589e\u5f37\u751f\u6210 RAG \u7279\u8f2f #15\u3009<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_monsterinsights_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"jetpack_post_was_ever_published":false,"_jetpack_newsletter_access":"","_jetpack_dont_email_post_to_subs":false,"_jetpack_newsletter_tier_id":0,"_jetpack_memberships_contains_paywalled_content":false,"_jetpack_memberships_contains_paid_content":false,"footnotes":"","jetpack_publicize_message":"","jetpack_publicize_feature_enabled":true,"jetpack_social_post_already_shared":true,"jetpack_social_options":{"image_generator_settings":{"template":"highway","default_image_id":0,"font":"","enabled":false},"version":2}},"categories":[91],"tags":[],"class_list":["post-12351","post","type-post","status-publish","format-standard","hentry","category-aie","entry"],"jetpack_publicize_connections":[],"jetpack_featured_media_url":"","jetpack_shortlink":"https:\/\/wp.me\/p1q6tG-3dd","jetpack_sharing_enabled":true,"jetpack_likes_enabled":true,"_links":{"self":[{"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/posts\/12351","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/comments?post=12351"}],"version-history":[{"count":3,"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/posts\/12351\/revisions"}],"predecessor-version":[{"id":13018,"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/posts\/12351\/revisions\/13018"}],"wp:attachment":[{"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/media?parent=12351"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/categories?post=12351"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/ihower.tw\/blog\/wp-json\/wp\/v2\/tags?post=12351"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}