{"id":1483,"date":"2024-05-16T03:35:47","date_gmt":"2024-05-16T03:35:47","guid":{"rendered":"https:\/\/www.nicekj.com\/?p=1483"},"modified":"2024-05-16T03:36:19","modified_gmt":"2024-05-16T03:36:19","slug":"goujianninde-ai-bokezhushouyi","status":"publish","type":"post","link":"https:\/\/www.nicekj.com\/goujianninde-ai-bokezhushouyi.html","title":{"rendered":"\u6784\u5efa\u60a8\u7684 AI \u64ad\u5ba2\u52a9\u624b[\u8bd1]"},"content":{"rendered":"

\u628a\u542c\u64ad\u5ba2\u7684\u6570\u5c0f\u65f6\u8f6c\u6362\u6210\u5373\u65f6\u7684\u6df1\u523b\u89c1\u89e3<\/h3>\n

\u4f5c\u4e3a\u4e00\u4e2a\u5bf9\u5065\u5eb7\u6216\u5546\u4e1a\u4e3b\u9898\u64ad\u5ba2\u60c5\u6709\u72ec\u949f\u7684\u7231\u597d\u8005\uff0c\u6211\u5e38\u5e38\u611f\u5230\u6709\u592a\u591a\u7684\u8282\u76ee\u8981\u542c\uff0c\u800c\u65f6\u95f4\u5374\u8fdc\u8fdc\u4e0d\u591f\u3002\u50cf Huberman Lab \u8fd9\u6837\u7684\u64ad\u5ba2\uff0c\u4e00\u671f\u8282\u76ee\u901a\u5e38\u80fd\u957f\u8fbe\u4e09\u56db\u4e2a\u5c0f\u65f6\u3002\u95ee\u9898\u4e0d\u4ec5\u662f\u627e\u51fa\u65f6\u95f4\u6765\u542c\uff0c\u66f4\u591a\u65f6\u5019\uff0c\u6211\u5bf9\u8c08\u8bdd\u4e2d\u7684\u67d0\u4e9b\u7279\u5b9a\u90e8\u5206\u611f\u5174\u8da3\u3002\u6b63\u56e0\u5982\u6b64\uff0c\u6211\u5229\u7528 OpenAI \u6a21\u578b\u3001LangChain \u548c Streamlit \u5f00\u53d1\u4e86\u4e00\u4e2a\u7b80\u6d01\u7684\u5e94\u7528\u7a0b\u5e8f\u3002\u5b83\u80fd\u8ba9\u6211\u7c98\u8d34\u4e00\u4e2a YouTube \u64ad\u5ba2\u94fe\u63a5\uff08\u6bd4\u5982 Huberman Lab \u6216 The Diary of a CEO\uff09\uff0c\u8fc5\u901f\u4e3a\u6211\u63d0\u4f9b\u8282\u76ee\u6458\u8981\u3002\u4e4b\u540e\uff0c\u6211\u8fd8\u53ef\u4ee5\u9488\u5bf9\u6027\u5730\u63d0\u95ee\uff0c\u5f97\u5230\u7acb\u523b\u7684\u7b54\u6848\u3002\u8fd9\u6837\u4e00\u6765\uff0c\u6211\u5c31\u80fd\u76f4\u63a5\u83b7\u53d6\u6211\u6240\u9700\u7684\u4fe1\u606f\uff0c\u65e0\u9700\u89c2\u770b\u6574\u671f\u8282\u76ee\u3002<\/p>\n

\n

\"image.png\"<\/figure>\n<\/p>\n

\u8ba4\u8bc6 RAG<\/h3>\n

\u5728\u6df1\u5165\u7814\u7a76\u4ee3\u7801\u4e4b\u524d\uff0c\u8ba9\u6211\u4eec\u5148\u7b80\u5355\u4e86\u89e3\u4e00\u4e0b RAG\u3002\u5b83\u662f\u4e00\u79cd\u7ed3\u5408\u4e86\u50cf GPT-4 \u8fd9\u7c7b\u6a21\u578b\u7684\u6587\u672c\u751f\u6210\u80fd\u529b\u548c\u4fe1\u606f\u68c0\u7d22\u529f\u80fd\u7684\u65b9\u6cd5\uff0c\u80fd\u591f\u63d0\u4f9b\u7cbe\u786e\u3001\u4e0e\u4e0a\u4e0b\u6587\u76f8\u5173\u7684\u4fe1\u606f\u3002\u5927\u8bed\u8a00\u6a21\u578b\u6709\u65f6\u5019\u53ef\u80fd\u65e0\u6cd5\u63a5\u89e6\u5230\u6211\u4eec\u9700\u8981\u7684\u7279\u5b9a\u6570\u636e\u3002RAG \u5141\u8bb8\u6211\u4eec\u8f93\u5165\u76f8\u5173\u6570\u636e\uff08\u6bd4\u5982\u64ad\u5ba2\u7684\u6587\u5b57\u8bb0\u5f55\uff09\uff0c\u8fd9\u6837\u6a21\u578b\u5c31\u80fd\u591f\u7ed9\u51fa\u5207\u9898\u7684\u7b54\u6848\u3002<\/p>\n

\u5176\u8fc7\u7a0b\u5f88\u7b80\u5355\uff1a\u6211\u4eec\u8f7d\u5165\u6570\u636e\uff08\u6bd4\u5982\u64ad\u5ba2\u7684\u6587\u5b57\u8bb0\u5f55\uff09\uff0c\u5c06\u5176\u5206\u5272\u6210\u5c0f\u7247\u6bb5\uff0c\u8fdb\u884c\u5d4c\u5165\u5904\u7406\uff0c\u7136\u540e\u5b58\u50a8\u5728\u5411\u91cf\u6570\u636e\u5e93\u4e2d\u3002<\/p>\n

\n

\"image.png\"<\/figure>\n<\/p>\n

\u5f53\u6211\u4eec\u63d0\u51fa\u67e5\u8be2\u65f6\uff0c\u6a21\u578b\u4f1a\u5c06\u6211\u4eec\u7684\u95ee\u9898\u4e0e\u8fd9\u4e9b\u6570\u636e\u7247\u6bb5\u8fdb\u884c\u5339\u914d\u3002\u6311\u9009\u51fa\u6700\u76f8\u5173\u7684\u6587\u672c\uff0c\u5c06\u5176\u8fde\u540c\u95ee\u9898\u4e00\u8d77\u8f93\u5165\u5927\u8bed\u8a00\u6a21\u578b\uff0c\u4ece\u800c\u751f\u6210\u51c6\u786e\u7684\u7b54\u6848\u3002\u7b80\u800c\u8a00\u4e4b\uff0cRAG \u901a\u8fc7\u68c0\u7d22\u76f8\u5173\u4fe1\u606f\u5e76\u5229\u7528\u8fd9\u4e9b\u4fe1\u606f\u6765\u56de\u7b54\u6211\u4eec\u7684\u95ee\u9898\u3002<\/p>\n

\n

\"image.png\"<\/figure>\n\n\u603b\u7684\u6765\u8bf4\uff0c\u8fd9\u5c31\u662f RAG \u7684\u5de5\u4f5c\u539f\u7406\u3002\u5b83\u627e\u5230\u4e0e\u4f60\u7684\u67e5\u8be2\u76f8\u5173\u7684\u4fe1\u606f\u7247\u6bb5\uff0c\u7136\u540e\u7ed3\u5408\u8fd9\u4e9b\u4fe1\u606f\u751f\u6210\u7b54\u6848\u3002<\/p>\n

\u5e94\u7528\u7a0b\u5e8f\u7684\u5de5\u4f5c\u539f\u7406<\/h3>\n

\u8fd9\u6b3e\u5e94\u7528\u7a0b\u5e8f\u975e\u5e38\u5bb9\u6613\u4e0a\u624b\uff1a\u53ea\u9700\u5c06 YouTube \u64ad\u5ba2\u7684\u94fe\u63a5\u7c98\u8d34\u5230 Streamlit \u754c\u9762\uff0c\u8f93\u5165\u4f60\u7684 OpenAI \u5bc6\u94a5\uff0c\u8f6c\u77ac\u95f4\uff0c\u4f60\u5c31\u80fd\u5f97\u5230\u4e00\u4e2a\u64ad\u5ba2\u6458\u8981\u3002\u63a5\u4e0b\u6765\uff0c\u4f60\u53ef\u4ee5\u63d0\u51fa\u5177\u4f53\u7684\u95ee\u9898\uff0c\u6bd4\u5982\u201c\u63a8\u8350\u4e86\u54ea\u4e9b\u4e66\uff1f\u201d\u6216\u8005\u201c\u63d0\u5230\u4e86\u54ea\u4e9b\u6700\u4f73\u7761\u7720\u5efa\u8bae\uff1f\u201d\u3002<\/p>\n

\u4ee3\u7801\u4e3b\u8981\u5206\u4e3a\u4e09\u4e2a\u90e8\u5206\uff1a\u4ece YouTube \u83b7\u53d6\u6570\u636e\u3001\u751f\u6210\u6458\u8981\u4e0e\u56de\u7b54\u4ee5\u53ca Streamlit \u524d\u7aef\u663e\u793a\u3002\u8fd8\u6709\u4e00\u4e2a\u53ef\u9009\u7684\u7b2c\u56db\u90e8\u5206\uff0c\u7528\u4e8e\u66f4\u9ad8\u7ea7\u7684 RAG \u5e94\u7528\uff0c\u4e0d\u8fc7\u57fa\u7840\u7248\u7684\u529f\u80fd\u5df2\u7ecf\u975e\u5e38\u51fa\u8272\u4e86\u3002\u8ba9\u6211\u4eec\u6765\u6df1\u5165\u4e86\u89e3\u4ee3\u7801\u5427\uff01<\/p>\n

\u7b2c\u4e00\u90e8\u5206\uff1a\u4e0e YouTube \u7684\u4e92\u52a8<\/h2>\n

\u9996\u5148\uff0c\u6211\u4eec\u7684\u5e94\u7528\u4f1a\u8fde\u63a5\u5230 YouTube\uff0c\u63d0\u53d6\u89c6\u9891\u7684\u6807\u9898\u548c\u5b57\u5e55 \u2014\u2014 \u8fd9\u662f\u7406\u89e3\u64ad\u5ba2\u5185\u5bb9\u7684\u5173\u952e\u6240\u5728\u3002<\/p>\n

\u529f\u80fd 1\uff1a\u83b7\u53d6\u89c6\u9891\u6807\u9898<\/strong><\/p>\n

\u8fd9\u4e2a\u529f\u80fd\u5229\u7528\u89c6\u9891\u7684 URL \u6765\u83b7\u53d6 YouTube \u89c6\u9891\u7684\u6807\u9898\u3002\u5b83\u53d1\u9001\u7f51\u7edc\u8bf7\u6c42\uff0c\u89e3\u6790 HTML \u4ee5\u627e\u5230\u6807\u9898\u6807\u7b7e\uff0c\u5e76\u8fd4\u56de\u6807\u9898\u3002\u8fd9\u662f\u6211\u4eec\u4e86\u89e3\u6bcf\u4e2a\u89c6\u9891\u5185\u5bb9\u7684\u8d77\u70b9\u3002<\/p>\n

<\/div>
python<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>import<\/span> requests<\/span>\nfrom<\/span> bs4 import<\/span> BeautifulSoup<\/span>\n<\/span>\ndef<\/span> get_youtube_video_title<\/span>(video_url<\/span>):<\/span>\n    response = requests.get(video_url)<\/span>\n    soup = BeautifulSoup(response.content, 'html.parser'<\/span>)<\/span>\n    title = soup.find('meta'<\/span>, property<\/span>='og:title'<\/span>)<\/span>\n    return<\/span> title['content'<\/span>] if<\/span> title else<\/span> \"Title not found\"<\/span><\/span>\n<\/code><\/pre>\n
\u529f\u80fd 2\uff1a\u63d0\u53d6\u5b57\u5e55<\/strong><\/p>\n
\u7d27\u63a5\u7740\uff0c\u6211\u4eec\u4f1a\u83b7\u53d6\u89c6\u9891\u7684\u5b57\u5e55\u3002\u6211\u4eec\u4f7f\u7528\u83b7\u53d6\u6807\u9898\u7684\u529f\u80fd\uff0c\u5e76\u501f\u52a9 LangChain \u4e2d\u7684 YoutubeLoader \u6765\u52a0\u8f7d\u5b57\u5e55\u3002\u8fd9\u4e9b\u6587\u5b57\u8bb0\u5f55\u5c06\u88ab\u8f93\u5165\u6211\u4eec\u7684 RAG \u7cfb\u7edf\uff0c\u4ee5\u751f\u6210\u76f8\u5173\u7684\u56de\u7b54\u3002<\/p>\n
<\/div>
ini<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>from langchain.document_loaders import YoutubeLoader<\/span>\nfrom langchain.schema import Document<\/span>\n<\/span>\ndef fetch_youtube_captions(video_url):<\/span>\n    title<\/span> = get_youtube_video_title(video_url)<\/span>\n    loader<\/span> = YoutubeLoader.from_youtube_url(video_url)<\/span>\n    docs<\/span> = loader.load()<\/span>\n    if docs and len(docs) > 0:<\/span>\n        intro_sentence<\/span> = \"This is the title of the video\/transcription\/conversation: \"<\/span><\/span>\n        title_content<\/span> = intro_sentence + title<\/span>\n        docs[0]<\/span> = Document(page_content<\/span>=title_content + \"nn\"<\/span> + docs[0<\/span>].page_content)<\/span>\n    return docs<\/span>\n<\/code><\/pre>\n
\u7b2c\u4e8c\u90e8\u5206\uff1a\u6570\u636e\u5904\u7406\u4e0e AI \u6574\u5408<\/h2>\n
\u6838\u5fc3\u5e94\u7528\u903b\u8f91\uff1a\u5c06\u5bf9\u8bdd\u8f6c\u6362\u4e3a\u6570\u636e<\/h3>\n
\u8fd9\u91cc\u662f\u201c\u53d8\u9b54\u672f\u201d\u7684\u5730\u65b9\u3002\u6211\u4eec\u5c06\u64ad\u5ba2\u5185\u5bb9\u5206\u89e3\u6210\u6613\u4e8e\u6d88\u5316\u7684\u5c0f\u5757\uff0c\u8f6c\u6362\u6210\u9002\u5408 AI \u5904\u7406\u7684\u683c\u5f0f\uff0c\u5e76\u8fdb\u884c\u5b58\u50a8\uff0c\u4ee5\u4fbf\u6211\u4eec\u968f\u65f6\u68c0\u7d22\u548c\u751f\u6210\u6458\u8981\u3002<\/p>\n
\u73af\u5883\u642d\u5efa\u548c\u5168\u5c40\u53d8\u91cf\u8bbe\u7f6e<\/strong><\/p>\n
\u6211\u4eec\u9996\u5148\u642d\u5efa\u7f16\u7a0b\u73af\u5883\u5e76\u5b9a\u4e49\u4e00\u4e9b\u5168\u5c40\u53d8\u91cf\u3002\u5176\u4e2d\u5305\u62ec\u521d\u59cb\u5316\u4e00\u4e2a\u6570\u636e\u5e93\u548c\u5bf9\u8bdd\u5185\u5bb9\u7684\u5b58\u50a8\u7cfb\u7edf\u3002<\/p>\n
<\/div>
python<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>from<\/span> langchain.text_splitter import<\/span> RecursiveCharacterTextSplitter<\/span>\nfrom<\/span> langchain.embeddings.openai import<\/span> OpenAIEmbeddings<\/span>\nfrom<\/span> langchain.vectorstores import<\/span> Chroma<\/span>\nfrom<\/span> sklearn.cluster import<\/span> KMeans<\/span>\nfrom<\/span> langchain.chat_models import<\/span> ChatOpenAI<\/span>\nfrom<\/span> langchain.chains.summarize import<\/span> load_summarize_chain<\/span>\nfrom<\/span> langchain.schema import<\/span> Document<\/span>\nimport<\/span> numpy as<\/span> np<\/span>\nfrom<\/span> langchain.chains import<\/span> ConversationalRetrievalChain<\/span>\nimport<\/span> logging<\/span>\nfrom<\/span> langchain.prompts import<\/span> PromptTemplate<\/span>\nfrom<\/span> langchain.memory import<\/span> ConversationBufferMemory<\/span>\n<\/span>\n<\/span>\nlogging.basicConfig(level=logging.INFO) <\/span>\n# Initialize global variables<\/span><\/span>\nglobal_chromadb = None<\/span><\/span>\nglobal_documents = None<\/span><\/span>\nglobal_short_documents = None<\/span><\/span>\n<\/code><\/pre>\n
\u6570\u636e\u7ba1\u7406\u51fd\u6570<\/strong><\/p>\n
\u8fd9\u4e9b\u51fd\u6570\u5bf9\u4e8e\u7ba1\u7406\u5e94\u7528\u6570\u636e\u81f3\u5173\u91cd\u8981\u3002reset_globals \u7528\u4e8e\u91cd\u7f6e\u5168\u5c40\u53d8\u91cf\uff0cinit_chromadb \u5219\u7528\u4e8e\u7528\u5904\u7406\u8fc7\u7684\u6570\u636e\u521d\u59cb\u5316\u6570\u636e\u5e93\u3002<\/p>\n
<\/div>
ini<\/span>
\u590d\u5236\u4ee3\u7801<\/div><\/div><\/div># Initialize the memory outside the function so it persists across different calls<\/span><\/span>\nconversation_memory<\/span> = ConversationBufferMemory(<\/span>\n    memory_key<\/span>=\"chat_history\"<\/span>,<\/span>\n    max_len<\/span>=50<\/span>,<\/span>\n    input_key<\/span>=\"question\"<\/span>,<\/span>\n    output_key<\/span>=\"answer\"<\/span>,<\/span>\n    return_messages<\/span>=True<\/span>,<\/span>\n)<\/span>\n<\/span>\n# Function to reset global variables<\/span><\/span>\ndef reset_globals():<\/span>\n    global global_chromadb, global_documents, global_short_documents<\/span>\n    global_chromadb<\/span> = None<\/span>\n    global_documents<\/span> = None<\/span>\n    global_short_documents<\/span> = None<\/span>\n    # Reset the conversation memory<\/span><\/span>\n    if conversation_memory:<\/span>\n        conversation_memory.clear()<\/span>\n<\/span>\ndef init_chromadb(openai_api_key):<\/span>\n    global global_chromadb, global_short_documents<\/span>\n    if global_chromadb is None and global_short_documents is not None:<\/span>\n        global_chromadb<\/span> = Chroma.from_documents(documents=global_short_documents, embedding=OpenAIEmbeddings(openai_api_key=openai_api_key))<\/span>\n        <\/span>\n<\/code><\/pre>\n
\u5904\u7406\u5b57\u5e55\u5e76\u751f\u6210\u6458\u8981<\/strong><\/p>\n
process_and_cluster_captions: \u8fd9\u91cc\uff0c\u6211\u4eec\u5bf9 YouTube \u5b57\u5e55\u8fdb\u884c\u52a0\u5de5\uff0c\u4e3a\u5206\u6790\u548c\u751f\u6210\u56de\u7b54\u505a\u597d\u51c6\u5907\u3002\u5177\u4f53\u6b65\u9aa4\u5305\u62ec\uff1a<\/p>\n
    \n
  1. \n
    \u521d\u6b65\u6570\u636e\u68c0\u67e5\uff1a<\/strong> \u6211\u4eec\u5148\u68c0\u67e5\u5b57\u5e55\u7684\u683c\u5f0f\uff0c\u786e\u4fdd\u5b83\u4eec\u9002\u5408\u540e\u7eed\u5904\u7406\u3002<\/p>\n<\/li>\n
  2. \n
    \u5207\u5206\u5b57\u5e55\uff1a<\/strong> \u5b57\u5e55\u88ab\u5206\u5272\u6210\u66f4\u5c0f\u7684\u6bb5\u843d\uff0c\u5206\u522b\u7528\u4e8e\u5236\u4f5c\u6458\u8981\u548c\u56de\u7b54\u95ee\u9898\u3002\u8fd9\u6837\u7684\u5207\u5206\u5bf9\u9ad8\u6548\u5b8c\u6210\u5177\u4f53\u4efb\u52a1\u975e\u5e38\u5173\u952e\u3002<\/p>\n<\/li>\n
  3. \n
    \u805a\u7c7b\u5f52\u7eb3\u76f8\u5173\u5185\u5bb9\uff1a<\/strong> \u6211\u4eec\u8fd0\u7528 KMeans \u7b97\u6cd5\u5bf9\u5236\u4f5c\u6458\u8981\u7684\u6bb5\u843d\u8fdb\u884c\u805a\u7c7b\u3002\u8fd9\u6837\u505a\u53ef\u4ee5\u7b5b\u9009\u51fa\u91cd\u590d\u5185\u5bb9\uff0c\u4ec5\u4fdd\u7559\u64ad\u5ba2\u4e2d\u6700\u6709\u4ee3\u8868\u6027\u7684\u90e8\u5206\u3002\u4ece\u6bcf\u4e2a\u7c7b\u522b\u4e2d\u9009\u53d6\u4e00\u4e2a\u6bb5\u843d\uff0c\u6211\u4eec\u786e\u4fdd AI \u80fd\u63a5\u6536\u5230\u65e2\u591a\u6837\u5316\u53c8\u7cbe\u70bc\u7684\u4fe1\u606f\uff0c\u975e\u5e38\u9002\u5408\u521b\u5efa\u6709\u610f\u4e49\u7684\u6458\u8981\u3002<\/p>\n<\/li>\n
  4. \n
    \u5168\u5c40\u5b58\u50a8\uff1a<\/strong> \u5206\u5272\u597d\u7684\u5b57\u5e55\u88ab\u5168\u5c40\u5b58\u50a8\uff0c\u65b9\u4fbf\u4e4b\u540e\u5236\u4f5c\u6458\u8981\u548c\u56de\u7b54\u95ee\u9898\u65f6\u968f\u65f6\u8bbf\u95ee\u3002<\/p>\n
    def process_and_cluster_captions(captions, openai_api_key, num_clusters=12):\nglobal global_documents, global_short_documents\nlogging.info(“\u6b63\u5728\u5904\u7406\u548c\u805a\u7c7b\u5b57\u5e55”)<\/p>\n
    <\/div>
    ini<\/span>
    \u590d\u5236\u4ee3\u7801<\/div><\/div><\/div># \u8bb0\u5f55\u5b57\u5e55\u5f00\u5934\u7684500\u4e2a\u5b57\u7b26\uff0c\u68c0\u67e5\u5176\u683c\u5f0f<\/span><\/span>\nlogging.info(f\"\u63a5\u6536\u5230\u7684\u5b57\u5e55\uff08\u524d500\u4e2a\u5b57\u7b26\uff09: {captions[0]<\/span>.page_content[:500]<\/span>}\")<\/span>\ncaption_content<\/span> = captions[0<\/span>].page_content<\/span>\n<\/span>\n# \u786e\u8ba4\u5b57\u5e55\u662f\u5b57\u7b26\u4e32\u683c\u5f0f\uff0c\u4ee5\u4fbf\u5904\u7406<\/span><\/span>\nif not isinstance(caption_content, str):<\/span>\n    logging.error(\"\u5b57\u5e55\u683c\u5f0f\u4e0e\u9884\u671f\u4e0d\u7b26\")<\/span>\n    return []<\/span><\/span>\n<\/span>\n# \u4e3a\u6458\u8981\u521b\u5efa\u8f83\u957f\u7684\u6587\u672c\u5757<\/span><\/span>\nsummary_splitter<\/span> = RecursiveCharacterTextSplitter(chunk_size=1000<\/span>, chunk_overlap=0<\/span>, separators=[\"nn\"<\/span>, \"n\"<\/span>, \" \"<\/span>, \"\"<\/span>])<\/span>\nsummary_docs<\/span> = summary_splitter.create_documents([caption_content])<\/span>\n<\/span>\n# \u4e3a\u95ee\u7b54\u521b\u5efa\u8f83\u77ed\u7684\u6587\u672c\u5757<\/span><\/span>\nqa_splitter<\/span> = RecursiveCharacterTextSplitter(chunk_size=200<\/span>, chunk_overlap=0<\/span>, separators=[\"nn\"<\/span>, \"n\"<\/span>, \" \"<\/span>, \"\"<\/span>])<\/span>\nqa_docs<\/span> = qa_splitter.create_documents([caption_content])<\/span>\n<\/span>\n# \u5904\u7406\u7528\u4e8e\u6458\u8981\u7684\u6587\u672c<\/span><\/span>\nsummary_embeddings<\/span> = OpenAIEmbeddings(openai_api_key=openai_api_key).embed_documents([x.page_content for x in summary_docs])<\/span>\nkmeans<\/span> = KMeans(n_clusters=num_clusters, random_state=42<\/span>).fit(summary_embeddings)<\/span>\nclosest_indices<\/span> = [np.argmin(np.linalg.norm(summary_embeddings - center, axis=1<\/span>)) for center in kmeans.cluster_centers_]<\/span>\nrepresentative_docs<\/span> = [summary_docs[i] for i in closest_indices]<\/span>\n<\/span>\n# \u5168\u5c40\u5b58\u50a8\u6587\u6863<\/span><\/span>\nglobal_documents<\/span> = summary_docs  # \u5b58\u50a8\u7528\u4e8e\u5236\u4f5c\u6458\u8981\u7684\u6587\u6863<\/span><\/span>\nglobal_short_documents<\/span> = qa_docs  # \u5b58\u50a8\u7528\u4e8e\u95ee\u7b54\u7684\u6587\u6863<\/span><\/span>\n<\/span>\ninit_chromadb(openai_api_key)  # \u5229\u7528\u8f83\u957f\u7684\u6587\u672c\u5757\u521d\u59cb\u5316\u6570\u636e\u5e93<\/span><\/span>\nreturn representative_docs<\/span>\n<\/code><\/pre>\n<\/li>\n<\/ol>\n
    generate_summary: \u8be5\u51fd\u6570\u4f7f\u7528\u5904\u7406\u8fc7\u7684\u5b57\u5e55\uff0c\u501f\u52a9 OpenAI \u7684 AI \u6a21\u578b\u751f\u6210\u89c6\u9891\u7684\u6e05\u6670\u4e14\u7b80\u6d01\u7684\u6458\u8981\u3002\u5177\u4f53\u5982\u4e0b\uff1a<\/p>\n
      \n
    1. \n
      \u6574\u5408\u5173\u952e\u6bb5\u843d\uff1a<\/strong> \u6211\u4eec\u5c06\u805a\u7c7b\u8fc7\u7a0b\u4e2d\u6311\u9009\u51fa\u7684\u6bb5\u843d\u7ed3\u5408\u8d77\u6765\uff0c\u5f62\u6210\u8fde\u8d2f\u7684\u53d9\u8ff0\u3002\u8fd9\u786e\u4fdd\u4e86\u6458\u8981\u80fd\u591f\u8986\u76d6\u64ad\u5ba2\u7684\u4e0d\u540c\u5c42\u9762\u3002<\/p>\n<\/li>\n
    2. \n
      AI \u5bfc\u5411\u7684\u6458\u8981\u5236\u4f5c\uff1a<\/strong> \u5229\u7528\u5b9a\u5236\u7684\u63d0\u793a\u8bed\uff0c\u6211\u4eec\u6307\u5bfc AI \u5236\u4f5c\u4e00\u4e2a\u7b80\u660e\u4e14\u5305\u542b\u4e30\u5bcc\u4fe1\u606f\u7684\u6458\u8981\u3002\u8fd9\u4e00\u6b65\u9aa4\u5bf9\u4e8e\u5f15\u5bfc AI \u805a\u7126\u4e8e\u603b\u4f53\u4e3b\u9898\u548c\u91cd\u70b9\u5185\u5bb9\u3001\u907f\u514d\u65e0\u5173\u7ec6\u8282\u975e\u5e38\u5173\u952e\u3002<\/p>\n<\/li>\n
    3. \n
      \u6267\u884c\u6458\u8981\u94fe\uff1a<\/strong> AI \u5904\u7406\u8fd9\u4e9b\u7ed3\u5408\u5728\u4e00\u8d77\u7684\u6587\u672c\uff0c\u751f\u6210\u6458\u8981\uff0c\u8ba9\u7528\u6237\u8fc5\u901f\u628a\u63e1\u64ad\u5ba2\u7684\u6838\u5fc3\u5185\u5bb9\u3002<\/p>\n<\/li>\n<\/ol>\n
      \u8fd9\u79cd\u65b9\u6cd5\u786e\u4fdd\u6240\u751f\u6210\u7684\u6458\u8981\u65e2\u5168\u9762\u53c8\u7cbe\u51c6\uff0c\u6355\u6349\u4e86\u64ad\u5ba2\u7684\u7cbe\u534e\u6240\u5728\u3002<\/p>\n
      <\/div>
      ini<\/span>
      \u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>def generate_summary(representative_docs, openai_api_key, model_name):<\/span>\n    logging.info(\"\u6b63\u5728\u751f\u6210\u6458\u8981\")<\/span>\n    llm4<\/span> = ChatOpenAI(model_name=model_name, temperature=0.2<\/span>, openai_api_key=openai_api_key)<\/span>\n<\/span>\n    # \u6c47\u603b\u6587\u672c\u4ee5\u51c6\u5907\u751f\u6210\u6458\u8981<\/span><\/span>\n    summary_text<\/span> = \"n\"<\/span>.join([doc.page_content for doc in representative_docs])<\/span>\n<\/span>\n    summary_prompt_template<\/span> = PromptTemplate(<\/span>\n        template<\/span>=(<\/span>\n            \"\u6839\u636e\u4e0b\u9762\u63d0\u4f9b\u7684\u6587\u672c\uff0c\u521b\u5efa\u4e00\u4e2a\u64ad\u5ba2\u5bf9\u8bdd\u7684\u7b80\u6d01\u6458\u8981\u3002\u6587\u672c\u5305\u542b\u4ece\u5bf9\u8bdd\u4e0d\u540c\u90e8\u5206\u6311\u9009\u51fa\u7684\u3001\u5177\u6709\u4ee3\u8868\u6027\u7684\u8282\u9009\u3002\"<\/span>\n            \"\u4f60\u7684\u4efb\u52a1\u662f\u5c06\u8fd9\u4e9b\u8282\u9009\u7efc\u5408\u6210\u4e00\u4e2a\u8fde\u8d2f\u4e14\u7b80\u6d01\u7684\u6458\u8981\u3002\u4e13\u6ce8\u4e8e\u64ad\u5ba2\u4e2d\u8ba8\u8bba\u7684\u603b\u4f53\u4e3b\u9898\u548c\u4e3b\u8981\u89c2\u70b9\u3002\"<\/span>\n            \"\u6458\u8981\u5e94\u8be5\u6e05\u6670\u5b8c\u6574\u5730\u4f20\u8fbe\u5bf9\u8bdd\u7684\u5173\u952e\u8bdd\u9898\u548c\u6d1e\u89c1\uff0c\u540c\u65f6\u7701\u7565\u4efb\u4f55\u4e0d\u5fc5\u8981\u7684\u7ec6\u8282\u3002\u5b83\u5e94\u8be5\u662f\u5f15\u4eba\u5165\u80dc\u4e14\u6613\u4e8e\u9605\u8bfb\u7684\uff0c\u7406\u60f3\u60c5\u51b5\u4e0b\u4e3a\u4e00\u5230\u4e24\u6bb5\u3002\u5c3d\u53ef\u80fd\u4fdd\u6301\u7b80\u77ed\"<\/span>\n            \"nn\u9009\u5b9a\u7684\u64ad\u5ba2\u8282\u9009\uff1an{text}nn\u6458\u8981\uff1a\"<\/span>\n        ),<\/span>\n        input_variables<\/span>=[\"text\"<\/span>]<\/span>\n    )<\/span>\n    # \u52a0\u8f7d\u6458\u8981\u94fe<\/span><\/span>\n    summarize_chain<\/span> = load_summarize_chain(llm=llm4, chain_type=\"stuff\"<\/span>, prompt=summary_prompt_template)<\/span>\n<\/span>\n    # \u6267\u884c\u6458\u8981\u94fe<\/span><\/span>\n    summary<\/span> = summarize_chain.run([Document(page_content=summary_text)])<\/span>\n<\/span>\n    logging.info(\"\u6458\u8981\u751f\u6210\u5b8c\u6210\")<\/span>\n    return summary<\/span>\n<\/code><\/pre>\n
      \u89e3\u7b54\u7528\u6237\u95ee\u9898<\/strong><\/p>\n
      answer_question: \u8fd9\u662f\u6211\u4eec RAG\uff08\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff09\u7cfb\u7edf\u7684\u6838\u5fc3\u529f\u80fd\u3002\u6b64\u529f\u80fd\u5229\u7528\u5df2\u5904\u7406\u7684\u6570\u636e\u56de\u7b54\u7528\u6237\u7684\u7279\u5b9a\u95ee\u9898\uff0c\u5145\u5206\u5c55\u73b0\u4e86\u7ed3\u5408\u68c0\u7d22\u7684\u751f\u6210\u6280\u672f\u7684\u80fd\u529b\u3002<\/p>\n
        \n
      1. \n
        \u6570\u636e\u5e93\u521d\u59cb\u5316\uff1a<\/strong> \u5b83\u9996\u5148\u786e\u8ba4\u5b58\u50a8\u5df2\u5904\u7406\u6bb5\u843d\u7684\u6570\u636e\u5e93\u662f\u5426\u51c6\u5907\u5c31\u7eea\u3002\u5982\u679c\u5c1a\u672a\u521d\u59cb\u5316\uff0c\u7cfb\u7edf\u5c06\u914d\u7f6e\u6570\u636e\u5e93\uff0c\u786e\u4fdd\u80fd\u591f\u9ad8\u6548\u5730\u8fdb\u884c\u4fe1\u606f\u68c0\u7d22\u3002<\/p>\n<\/li>\n
      2. \n
        \u4fe1\u606f\u68c0\u7d22\u4e0e\u56de\u7b54\u751f\u6210\uff1a<\/strong> \u63a5\u7740\uff0c\u7cfb\u7edf\u5c06\u7528\u6237\u7684\u63d0\u95ee\u4e0e\u4ece\u6570\u636e\u5e93\u4e2d\u68c0\u7d22\u5230\u7684\u76f8\u5173\u4fe1\u606f\u7ed3\u5408\uff0c\u8fdb\u884c\u667a\u80fd\u5339\u914d\u3002\u8fd9\u79cd\u7cbe\u51c6\u68c0\u7d22\u786e\u4fdd\u4e86 AI \u63d0\u4f9b\u7684\u7b54\u6848\u65e2\u51c6\u786e\u53c8\u5207\u9898\u3002<\/p>\n<\/li>\n
      3. \n
        \u751f\u6210\u5177\u6709\u60c5\u5883\u611f\u77e5\u7684\u7b54\u6848\uff1a<\/strong> AI \u5728\u638c\u63e1\u4e86\u95ee\u9898\u548c\u6700\u5339\u914d\u7684\u4e0a\u4e0b\u6587\u540e\uff0c\u4f1a\u751f\u6210\u4e00\u4e2a\u7b80\u660e\u76f4\u63a5\u7684\u7b54\u6848\u6765\u56de\u5e94\u7528\u6237\u7684\u95ee\u9898\u3002<\/p>\n<\/li>\n<\/ol>\n
        \u901a\u8fc7\u8fd9\u4e00\u8fc7\u7a0b\uff0c\u7528\u6237\u80fd\u5f97\u5230\u9488\u5bf9\u6027\u7684\u3001\u4fe1\u606f\u4e30\u5bcc\u7684\u7b54\u590d\uff0c\u8fd9\u5927\u5927\u63d0\u5347\u4e86\u4ed6\u4eec\u4e0e\u64ad\u5ba2\u5185\u5bb9\u7684\u4e92\u52a8\u4f53\u9a8c\u3002<\/p>\n
        <\/div>
        ini<\/span>
        \u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>def answer_question(question, openai_api_key, model_name):<\/span>\n        llm4<\/span> = ChatOpenAI(model_name=model_name, temperature=0<\/span>, openai_api_key=openai_api_key)<\/span>\n        global global_chromadb, global_short_documents<\/span>\n    <\/span>\n        if global_chromadb is None and global_short_documents is not None:<\/span>\n            init_chromadb(openai_api_key, documents<\/span>=global_short_documents)<\/span>\n        <\/span>\n        logging.info(f\"Answering question: {question}\")<\/span>\n        chatTemplate<\/span> = \"\"\"<\/span><\/span>\n        You are an AI assistant tasked with answering questions based on context from a podcast conversation. Use the provided context and relevant chat messages to answer. If unsure, say so. Keep your answer to three sentences or less, focusing on the most relevant information.<\/span>\n        Chat Messages (if relevant): {chat_history}<\/span>\n        Question: {question} <\/span>\n        Context from Podcast: {context} <\/span>\n        Answer:\"\"\"<\/span>\n        QA_CHAIN_PROMPT<\/span> = PromptTemplate(input_variables=[\"context\"<\/span>, \"question\"<\/span>, \"chat_history\"<\/span>],template=chatTemplate)<\/span>\n        qa_chain<\/span> = ConversationalRetrievalChain.from_llm(<\/span>\n            llm<\/span>=llm4, <\/span>\n            chain_type<\/span>=\"stuff\"<\/span>, <\/span>\n            retriever<\/span>=global_chromadb.as_retriever(search_type=\"mmr\"<\/span>, search_kwargs={\"k\"<\/span>:12<\/span>}),<\/span>\n            memory<\/span>=conversation_memory,<\/span>\n            #return_source_documents=True,<\/span><\/span>\n            combine_docs_chain_kwargs<\/span>={'prompt'<\/span>: QA_CHAIN_PROMPT},<\/span>\n        )<\/span>\n        # Log the current chat history<\/span><\/span>\n        current_chat_history<\/span> = conversation_memory.load_memory_variables({})<\/span>\n        logging.info(f\"Current Chat History: {current_chat_history}\")<\/span>\n        response<\/span> = qa_chain({\"question\"<\/span>: question}) <\/span>\n        logging.info(f\"this is the result: {response}\")<\/span>\n        output<\/span> = response['answer'<\/span>]    <\/span>\n    <\/span>\n        return output<\/span>\n<\/code><\/pre>\n
        \u7b2c 3 \u90e8\u5206\uff1aStreamlit \u7528\u6237\u754c\u9762<\/h2>\n
        \u6fc0\u6d3b\u5e94\u7528\u7684\u7075\u9b42<\/h3>\n
        \u6700\u7ec8\uff0c\u6211\u4eec\u91c7\u7528 Streamlit \u6765\u6253\u9020\u4e00\u4e2a\u4e92\u52a8\u6027\u5f3a\u7684\u7f51\u7edc\u5e94\u7528\uff0c\u5b83\u6574\u5408\u4e86\u6240\u6709\u7684\u529f\u80fd\u3002\u7528\u6237\u53ef\u4ee5\u8f93\u5165 YouTube \u89c6\u9891\u94fe\u63a5\uff0c\u63d0\u51fa\u95ee\u9898\uff0c\u5e76\u83b7\u5f97 AI \u751f\u6210\u7684\u6458\u8981\u548c\u7b54\u6848\u3002\u4ee5\u4e0b\u4ee3\u7801\u662f\u6211\u8bbe\u8ba1\u7684\u754c\u9762\u7248\u672c\uff0c\u4f60\u53ef\u4ee5\u76f4\u63a5\u91c7\u7528\u6216\u6839\u636e\u4e2a\u4eba\u559c\u597d\u8fdb\u884c\u8c03\u6574\u3002<\/p>\n
        <\/div>
        python<\/span>
        \u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>import<\/span> streamlit as<\/span> st<\/span>\n    from<\/span> youtuber import<\/span> fetch_youtube_captions<\/span>\n    from<\/span> agent import<\/span> process_and_cluster_captions, generate_summary, answer_question, reset_globals<\/span>\n    <\/span>\n    # Set Streamlit page configuration with custom tab title<\/span><\/span>\n    st.set_page_config(page_title=\"\ud83c\udfc4GPTpod\"<\/span>, page_icon=\"\ud83c\udfc4\"<\/span>, layout=\"wide\"<\/span>)<\/span>\n    <\/span>\n    def<\/span> user_query<\/span>(question, openai_api_key, model_name<\/span>):<\/span>\n        \"\"\"Process and display the query response.\"\"\"<\/span><\/span>\n        # Add the user's question to the conversation<\/span><\/span>\n        st.session_state.conversation.append((f\"{question}<\/span>\"<\/span>, \"user-message\"<\/span>))<\/span>\n    <\/span>\n        # Check if this query has been processed before<\/span><\/span>\n        if<\/span> question not<\/span> in<\/span> st.session_state.processed_questions:<\/span>\n            # Process the query<\/span><\/span>\n            answer = answer_question(question, openai_api_key, model_name)<\/span>\n            if<\/span> isinstance<\/span>(answer, str<\/span>):<\/span>\n                st.session_state.conversation.append((f\"{answer}<\/span>\"<\/span>, \"grimoire-message\"<\/span>))<\/span>\n            else<\/span>:<\/span>\n                st.session_state.conversation.append((\"Could not find a proper answer.\"<\/span>, \"grimoire-message\"<\/span>))<\/span>\n            <\/span>\n            st.rerun()<\/span>\n    <\/span>\n            # Mark this question as processed<\/span><\/span>\n            st.session_state.processed_questions.add(question)<\/span>\n    <\/span>\n    <\/span>\n    # Initialize session state<\/span><\/span>\n    if<\/span> 'conversation'<\/span> not<\/span> in<\/span> st.session_state:<\/span>\n        st.session_state.conversation = []<\/span>\n        st.session_state.asked_questions = set<\/span>()<\/span>\n        st.session_state.processed_questions = set<\/span>()<\/span>\n    <\/span>\n    # Sidebar for input and operations<\/span><\/span>\n    with<\/span> st.sidebar:<\/span>\n        st.title(\"GPT Podcast Surfer\ud83c\udf0a\ud83c\udfc4\ud83c\udffc\"<\/span>)<\/span>\n        st.image(\"img.png\"<\/span>) <\/span>\n    <\/span>\n        # Expandable Instructions<\/span><\/span>\n        with<\/span> st.expander(\"\ud83d\udd0d How to use:\"<\/span>, expanded=False<\/span>):<\/span>\n            st.markdown(\"\"\"<\/span><\/span>\n                - \ud83d\udd10 **Enter your OpenAI API Key.**<\/span>\n                - \ud83d\udcfa **Paste a YouTube URL.**<\/span>\n                - \ud83c\udfc3\u200d\u2642\ufe0f **Click 'Run it' to process.**<\/span>\n                - \ud83d\udd75\ufe0f\u200d\u2642\ufe0f **Ask questions in the chat.**<\/span>\n            \"\"\")<\/span>\n    <\/span>\n        # Model selection in the sidebar<\/span><\/span>\n        model_choice = st.sidebar.selectbox(\"Choose Model:\"<\/span>, <\/span>\n                                            (\"GPT-4 Turbo\"<\/span>, \"GPT-3.5 Turbo\"<\/span>), <\/span>\n                                            index=0<\/span>)  # Default to GPT-4 Turbo<\/span><\/span>\n    <\/span>\n        # Map friendly names to actual model names<\/span><\/span>\n        model_name_mapping = {<\/span>\n            \"GPT-4 Turbo\"<\/span>: \"gpt-4-1106-preview\"<\/span>,<\/span>\n            \"GPT-3.5 Turbo\"<\/span>: \"gpt-3.5-turbo\"<\/span><\/span>\n        }<\/span>\n    <\/span>\n        selected_model = model_name_mapping[model_choice]<\/span>\n        st.session_state['selected_model'<\/span>] = model_name_mapping[model_choice]<\/span>\n    <\/span>\n    <\/span>\n        # Input for OpenAI API Key<\/span><\/span>\n        openai_api_key = st.text_input(\"Enter your OpenAI API Key:\"<\/span>, type<\/span>=\"password\"<\/span>)<\/span>\n    <\/span>\n        # Save the API key in session state if it's entered<\/span><\/span>\n        if<\/span> openai_api_key:<\/span>\n            st.session_state['openai_api_key'<\/span>] = openai_api_key<\/span>\n    <\/span>\n        youtube_url = st.text_input(\"Enter YouTube URL:\"<\/span>)<\/span>\n    <\/span>\n        # Button to trigger processing<\/span><\/span>\n        if<\/span> st.button(\"\ud83d\ude80Run it\"<\/span>):<\/span>\n            if<\/span> openai_api_key:<\/span>\n                if<\/span> youtube_url and<\/span> 'processed_data'<\/span> not<\/span> in<\/span> st.session_state:<\/span>\n                    reset_globals()<\/span>\n                    with<\/span> st.spinner('\ud83d\udc69\u200d\ud83c\udf73 GPT is cooking up your podcast... hang tight for a few secs\ud83c\udf73'<\/span>):<\/span>\n                        captions = fetch_youtube_captions(youtube_url)<\/span>\n                        if<\/span> captions:<\/span>\n                            representative_docs = process_and_cluster_captions(captions, st.session_state['openai_api_key'<\/span>])<\/span>\n                            summary = generate_summary(representative_docs, st.session_state['openai_api_key'<\/span>], selected_model)<\/span>\n                            st.session_state.processed_data = (representative_docs, summary)<\/span>\n                            if<\/span> 'summary_displayed'<\/span> not<\/span> in<\/span> st.session_state:<\/span>\n                                st.session_state.conversation.append((f\"Here's a rundown of the conversation: {summary}<\/span>\"<\/span>, \"summary-message\"<\/span>))<\/span>\n                                guiding_message = \"Feel free to ask me anything else about it! :)\"<\/span><\/span>\n                                st.session_state.conversation.append((guiding_message, \"grimoire-message\"<\/span>))<\/span>\n                                st.session_state['summary_displayed'<\/span>] = True<\/span><\/span>\n                        else<\/span>:<\/span>\n                            st.error(\"Failed to fetch captions.\"<\/span>)<\/span>\n            else<\/span>:<\/span>\n                st.warning(\"Please add the OpenAI API key first.\"<\/span>)<\/span>\n    <\/span>\n    <\/span>\n    # Main app logic<\/span><\/span>\n    for<\/span> message, css_class in<\/span> st.session_state.conversation:<\/span>\n        role = \"assistant\"<\/span> if<\/span> css_class in<\/span> [\"grimoire-message\"<\/span>, \"summary-message\"<\/span>, \"suggestion-message\"<\/span>] else<\/span> \"user\"<\/span><\/span>\n        with<\/span> st.chat_message(role):<\/span>\n            st.markdown(message)<\/span>\n    <\/span>\n    <\/span>\n    # Chat input field<\/span><\/span>\n    if<\/span> prompt := st.chat_input(\"Ask me anything about the podcast...\"<\/span>):<\/span>\n        user_query(prompt, st.session_state.get('openai_api_key'<\/span>, ''<\/span>), st.session_state.get('selected_model'<\/span>, 'gpt-4-1106-preview'<\/span>))<\/span>\n<\/span>\n<\/code><\/pre>\n
        \u8fd9\u662f\u7528\u6237\u754c\u9762\u7684\u6837\u5b50\uff1a<\/p>\n
        \n
        \"image.png\"<\/figure>\n<\/p>\n
        \u7b2c 4 \u90e8\u5206\uff1a\u589e\u5f3a\u7ed3\u679c\u7684\u53ef\u9009 RAG-Fusion \u6280\u672f<\/h2>\n
        RAG-Fusion \u6280\u672f\u7b80\u4ecb<\/strong><\/p>\n
        RAG-Fusion \u662f\u6211\u4e3a\u4e86\u63d0\u5347\u7cfb\u7edf\u800c\u52a0\u5165\u7684\u4e00\u4e2a\u8bd5\u9a8c\u6027\u529f\u80fd\u3002\u5c3d\u7ba1\u8fd9\u4e0d\u662f\u5fc5\u987b\u7684\uff0c\u4f46\u5b83\u80fd\u5728\u589e\u52a0\u5c11\u91cf\u4ee3\u7801\u548c\u7a0d\u5fae\u727a\u7272\u901f\u5ea6\u7684\u57fa\u7840\u4e0a\uff0c\u63d0\u4f9b\u66f4\u4f18\u7684\u7ed3\u679c\u3002\u5176\u6838\u5fc3\u601d\u60f3\u5728\u4e8e\u63d0\u9ad8 AI \u7406\u89e3\u95ee\u9898\u548c\u63d0\u9ad8\u56de\u7b54\u51c6\u786e\u5ea6\u7684\u80fd\u529b\u3002<\/p>\n
        \u4e3a\u4ec0\u4e48\u9009\u62e9 RAG-Fusion\uff1f<\/strong><\/h3>\n
          \n
        • \n
          \u5f25\u8865\u4e0d\u8db3\uff1a<\/strong> \u5b83\u901a\u8fc7\u751f\u6210\u548c\u91cd\u65b0\u8bc4\u4f30\u7528\u6237\u95ee\u9898\u7684\u591a\u4e2a\u7248\u672c\uff0c\u5f25\u8865\u4e86\u6807\u51c6 RAG \u7684\u67d0\u4e9b\u5c40\u9650\uff0c\u786e\u4fdd\u66f4\u5e7f\u6cdb\u548c\u7cbe\u51c6\u7684\u641c\u7d22\u3002<\/p>\n<\/li>\n
        • \n
          \u66f4\u4f18\u7684\u641c\u7d22\u7ed3\u679c\uff1a<\/strong> \u5b83\u7ed3\u5408\u4e86\u4e92\u60e0\u6392\u540d\u878d\u5408\u6280\u672f\u548c\u81ea\u5b9a\u4e49\u5411\u91cf\u8bc4\u5206\u65b9\u6cd5\uff0c\u5f97\u5230\u66f4\u5168\u9762\u548c\u7cbe\u786e\u7684\u7b54\u6848\u3002<\/p>\n<\/li>\n<\/ul>\n
          RAG-Fusion \u529b\u56fe\u4e0d\u53ea\u662f\u89e3\u8bfb\u7528\u6237\u6240\u95ee\u7684\u95ee\u9898\uff0c\u8fd8\u8981\u6d1e\u6089\u7528\u6237\u771f\u6b63\u610f\u56fe\u63d0\u95ee\u7684\u5185\u5bb9\uff0c\u6df1\u5165\u6316\u6398\u90a3\u4e9b\u5e38\u88ab\u5ffd\u7565\u7684\u6df1\u5c42\u89c1\u89e3\u3002<\/p>\n
          RAG-Fusion \u7684\u5de5\u4f5c\u6d41\u7a0b<\/h3>\n
            \n
          1. \n
            \u8f6c\u6362\u67e5\u8be2\uff1a<\/strong> \u6211\u4eec\u9996\u5148\u7528\u8bed\u8a00\u6a21\u578b\u5c06\u539f\u59cb\u7528\u6237\u67e5\u8be2\u8f6c\u6362\u6210\u51e0\u4e2a\u76f8\u4f3c\u4f46\u5404\u6709\u4e0d\u540c\u7684\u95ee\u9898\u3002\u8fd9\u79cd\u591a\u89d2\u5ea6\u7684\u65b9\u6cd5\u5bf9\u4e8e\u5168\u9762\u641c\u7d22\u975e\u5e38\u5173\u952e\u3002<\/p>\n<\/li>\n
          2. \n
            \u589e\u5f3a\u7684\u5411\u91cf\u641c\u7d22\uff1a<\/strong> \u8fd9\u4e9b\u65b0\u751f\u6210\u7684\u67e5\u8be2\u5c06\u8fdb\u884c\u5411\u91cf\u641c\u7d22\uff0c\u805a\u96c6\u591a\u5143\u5316\u7684\u7ed3\u679c\u3002<\/p>\n<\/li>\n
          3. \n
            \u667a\u80fd\u91cd\u6392\uff1a<\/strong> \u901a\u8fc7\u4e92\u60e0\u6392\u540d\u878d\u5408\uff0c\u6211\u4eec\u5bf9\u6240\u6709\u8fd9\u4e9b\u7ed3\u679c\u8fdb\u884c\u91cd\u7ec4\uff0c\u628a\u6700\u76f8\u5173\u7684\u7ed3\u679c\u653e\u5728\u9996\u4f4d\u3002<\/p>\n<\/li>\n
          4. \n
            \u5b9a\u5236\u6700\u7ec8\u8f93\u51fa\uff1a<\/strong> \u7ed3\u5408\u9876\u5c16\u7ed3\u679c\u548c\u65b0\u751f\u6210\u7684\u67e5\u8be2\uff0c\u5f15\u5bfc\u8bed\u8a00\u6a21\u578b\u751f\u6210\u4e00\u4e2a\u57fa\u4e8e\u66f4\u5e7f\u6cdb\u4e0a\u4e0b\u6587\u7684\u56de\u7b54\u3002<\/p>\n<\/li>\n<\/ol>\n
            RAG-Fusion \u7684\u6838\u5fc3\u529f\u80fd<\/h3>\n
              \n
            • \n
              reciprocal_rank_fusion\uff1a<\/strong> \u8fd9\u4e2a\u529f\u80fd\u4f1a\u6839\u636e\u76f8\u5173\u6027\u5206\u6570\u91cd\u65b0\u6392\u5e8f\u641c\u7d22\u7ed3\u679c\uff0c\u786e\u4fdd\u4f18\u5148\u8003\u8651\u6700\u4f73\u7b54\u6848\u3002<\/p>\n<\/li>\n
            • \n
              generate_multiple_queries\uff1a<\/strong> \u5b83\u80fd\u751f\u6210\u521d\u59cb\u67e5\u8be2\u7684\u591a\u4e2a\u53d8\u79cd\uff0c\u4ece\u800c\u6269\u5927\u641c\u7d22\u7684\u8303\u56f4\u3002<\/p>\n<\/li>\n
            • \n
              answer_question\uff1a<\/strong> \u8fd9\u662f\u5c06\u6240\u6709\u90e8\u5206\u878d\u5408\u5728\u4e00\u8d77\u7684\u529f\u80fd\u6240\u5728\u3002\u9996\u5148\uff0c\u5b83\u751f\u6210\u591a\u4e2a\u67e5\u8be2\uff0c\u5bf9\u6bcf\u4e00\u4e2a\u67e5\u8be2\u68c0\u7d22\u76f8\u5e94\u7684\u6587\u6863\uff0c\u5e76\u8fdb\u884c\u4e92\u60e0\u6392\u540d\u878d\u5408\u3002\u7136\u540e\u5b83\u5229\u7528\u8fd9\u4e9b\u7cbe\u70bc\u540e\u7684\u7ed3\u679c\u548c\u81ea\u5b9a\u4e49\u6570\u636e\u5e93\uff0c\u5f15\u5bfc AI \u5236\u5b9a\u51fa\u66f4\u52a0\u4e30\u5bcc\u548c\u7cbe\u786e\u7684\u56de\u7b54\u3002<\/p>\n<\/li>\n<\/ul>\n\n
              <\/div>
              ini<\/span>
              \u590d\u5236\u4ee3\u7801<\/div><\/div><\/div>def reciprocal_rank_fusion(results: list[list]<\/span>, k<\/span>=60<\/span>):  <\/span>\nfused_scores<\/span> = {}  <\/span>\nfor docs in results:  <\/span>\n# the docs are returned in sorted order of relevance  <\/span><\/span>\nfor rank, doc in enumerate(docs):  <\/span>\ndoc_str<\/span> = dumps(doc)  <\/span>\nlogging.info(f\"Serialized Document: {doc_str}\")  <\/span>\nif doc_str not in fused_scores:  <\/span>\nfused_scores[doc_str]<\/span> = 0  <\/span>\nprevious_score<\/span> = fused_scores[doc_str]  <\/span>\nfused_scores[doc_str]<\/span> += 1 \/ (rank + k)  <\/span>\n  <\/span>\nreranked_results<\/span> = [  <\/span>\nloads(doc)  <\/span>\nfor doc, score in sorted(fused_scores.items(), key=lambda x: x[1<\/span>], reverse=True<\/span>)  <\/span>\n]  <\/span>\nlogging.info(f\"Reciprocal Rank Fusion applied. Reranked Results: {reranked_results[:10]<\/span>}\") # Log top 10 results  <\/span><\/span>\nreturn reranked_results  <\/span>\n  <\/span>\n  <\/span>\ndef generate_multiple_queries(question, llm):  <\/span>\nprompt<\/span> = PromptTemplate(  <\/span>\ninput_variables<\/span>=[\"question\"<\/span>],  <\/span>\ntemplate<\/span>=\"\"\"You are an AI language model assistant. Your task is to OUTPUT 4  <\/span><\/span>\ndifferent versions of the given user question to retrieve relevant documents from a vector  <\/span>\ndatabase. By generating multiple perspectives on the user question, your goal is to help  <\/span>\nthe user overcome some of the limitations of the distance-based similarity search.  <\/span>\nProvide these alternative questions separated by newlines.  <\/span>\nOriginal question: {question}\"\"\",  <\/span>\n)  <\/span>\n# Create a chain with the language model and the prompt  <\/span><\/span>\nllm_chain<\/span> = LLMChain(llm=llm, prompt=prompt, output_parser=StrOutputParser())  <\/span>\n  <\/span>\n# Run the chain  <\/span><\/span>\nresponse<\/span> = llm_chain.run({\"question\"<\/span>: question})  <\/span>\nqueries<\/span> = response.split(\"n\"<\/span>)  <\/span>\n  <\/span>\nlogging.info(f\"Generated Queries: {queries}\")  <\/span>\nreturn queries  <\/span>\n  <\/span>\n  <\/span>\ndef answer_question(question, openai_api_key, model_name):  <\/span>\nllm4<\/span> = ChatOpenAI(model_name=model_name, temperature=0.1<\/span>, openai_api_key=openai_api_key)  <\/span>\nglobal global_chromadb, global_short_documents  <\/span>\n  <\/span>\nif global_chromadb is None and global_short_documents is not None:  <\/span>\ninit_chromadb(openai_api_key, documents<\/span>=global_short_documents)  <\/span>\n  <\/span>\nlogging.info(f\"Answering question: {question}\")  <\/span>\n# Generate multiple queries  <\/span><\/span>\nqueries<\/span> = generate_multiple_queries(question, llm4)  <\/span>\n  <\/span>\n# Retrieve documents for each query  <\/span><\/span>\nresults<\/span> = []  <\/span>\nfor query in queries:  <\/span>\nretrieved_docs_with_scores<\/span> = global_chromadb.similarity_search_with_score(query, k=8<\/span>)  <\/span>\n# Log the number of documents retrieved for each query and the first 3 docs  <\/span><\/span>\nlogging.info(f\"Retrieved {len(retrieved_docs_with_scores)} documents for query '{query}': {retrieved_docs_with_scores[:3]<\/span>}\")  <\/span>\nresults.append(retrieved_docs_with_scores)  <\/span>\n  <\/span>\n# Apply reciprocal rank fusion  <\/span><\/span>\nreranked_results<\/span> = reciprocal_rank_fusion(results)  <\/span>\nlogging.info(f\"Number of reranked documents: {len(reranked_results)}\")  <\/span>\n  <\/span>\n#extract the Document object only  <\/span><\/span>\nreranked_documents<\/span> = [doc for doc, _ in reranked_results]  <\/span>\n  <\/span>\n# Create a new Chroma instance with reranked results  <\/span><\/span>\ncustom_chromadb<\/span> = Chroma.from_documents(documents=reranked_documents, embedding=OpenAIEmbeddings(openai_api_key=openai_api_key))  <\/span>\n  <\/span>\nchatTemplate<\/span> = \"\"\"  <\/span><\/span>\nYou are an AI assistant tasked with answering questions based on context from a podcast conversation.  <\/span>\nUse the provided context and relevant chat messages to answer. If unsure, say so. Keep your answer to four sentences or less, focusing on the most relevant information.  <\/span>\nChat Messages (if relevant): {chat_history}  <\/span>\nQuestion: {question}  <\/span>\nContext from Podcast: {context}  <\/span>\nAnswer:\"\"\"  <\/span>\nQA_CHAIN_PROMPT<\/span> = PromptTemplate(input_variables=[\"context\"<\/span>, \"question\"<\/span>, \"chat_history\"<\/span>],template=chatTemplate)  <\/span>\nqa_chain<\/span> = ConversationalRetrievalChain.from_llm(  <\/span>\nllm<\/span>=llm4,  <\/span>\nchain_type<\/span>=\"stuff\"<\/span>,  <\/span>\nretriever<\/span>=custom_chromadb.as_retriever(search_type=\"similarity\"<\/span>, search_kwargs={\"k\"<\/span>:10<\/span>}),  <\/span>\nmemory<\/span>=conversation_memory,  <\/span>\nreturn_source_documents<\/span>=True<\/span>,  <\/span>\ncombine_docs_chain_kwargs<\/span>={'prompt'<\/span>: QA_CHAIN_PROMPT},  <\/span>\n)  <\/span>\n# Log the current chat history  <\/span><\/span>\ncurrent_chat_history<\/span> = conversation_memory.load_memory_variables({})  <\/span>\nlogging.info(f\"Current Chat History: {current_chat_history}\")  <\/span>\nresponse<\/span> = qa_chain({\"question\"<\/span>: question})  <\/span>\nlogging.info(f\"Final response: {response}\")  <\/span>\noutput<\/span> = response['answer'<\/span>]  <\/span>\n  <\/span>\nreturn output<\/span>\n<\/code><\/pre>\n
              \u603b\u7ed3\n\u81f3\u6b64\uff0c\u6211\u4eec\u5b8c\u6210\u4e86\uff01\u4ee5\u4e0a\u662f\u4e00\u4efd\u8be6\u7ec6\u6307\u5357\uff0c\u6559\u4f60\u5982\u4f55\u6253\u9020\u4e00\u4e2a\u4e92\u52a8\u6027\u5f3a\u3001\u667a\u80fd\u5316\u7684\u64ad\u5ba2\u5e94\u7528\u3002\u6b22\u8fce\u5728\u8fd9\u91cc\u8bd5\u7528\uff0c\u5e76\u7559\u4e0b\u60a8\u5b9d\u8d35\u7684\u610f\u89c1\u3002<\/p>\n
              \u611f\u8c22\u60a8\u7684\u966a\u4f34\uff0c\u795d\u60a8\u7f16\u7a0b\u6109\u5feb\uff01:)<\/p>","protected":false},"excerpt":{"rendered":"
              \u628a\u542c\u64ad\u5ba2\u7684\u6570\u5c0f\u65f6\u8f6c\u6362\u6210\u5373\u65f6\u7684\u6df1\u523b\u89c1\u89e3 \u4f5c\u4e3a\u4e00\u4e2a\u5bf9\u5065\u5eb7\u6216\u5546\u4e1a\u4e3b\u9898\u64ad\u5ba2\u60c5\u6709\u72ec\u949f\u7684\u7231\u597d\u8005\uff0c\u6211\u5e38\u5e38\u611f\u5230\u6709\u592a\u591a\u7684\u8282\u76ee\u8981\u542c\uff0c\u800c\u65f6\u95f4\u5374\u8fdc\u8fdc\u4e0d\u591f\u3002\u50cf Huberman Lab \u8fd9\u6837\u7684\u64ad\u5ba2\uff0c\u4e00\u671f\u8282\u76ee\u901a\u5e38\u80fd\u957f\u8fbe\u4e09\u56db\u4e2a\u5c0f\u65f6<\/p>\n","protected":false},"author":1,"featured_media":9051,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"rank_math_title":"","rank_math_description":"","rank_math_focus_keyword":"","views":"","footnotes":""},"categories":[3],"tags":[127,128,129,136,126],"collection":[],"class_list":["post-1483","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-fenlei2","tag-ai","tag-128","tag-129","tag-136","tag-gpt"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/posts\/1483","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/comments?post=1483"}],"version-history":[{"count":0,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/posts\/1483\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/media\/9051"}],"wp:attachment":[{"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/media?parent=1483"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/categories?post=1483"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/tags?post=1483"},{"taxonomy":"collection","embeddable":true,"href":"https:\/\/www.nicekj.com\/nicekj2024\/wp\/v2\/collection?post=1483"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}