{"id":381,"date":"2026-01-07T20:25:47","date_gmt":"2026-01-07T12:25:47","guid":{"rendered":"https:\/\/flycloud.io\/?p=381"},"modified":"2026-10-07T20:26:41","modified_gmt":"2026-10-07T12:26:41","slug":"%e4%bd%bf%e7%94%a8%e9%98%bf%e9%87%8c%e4%ba%91%e7%99%be%e7%82%bc%e6%9e%84%e5%bb%ba%e4%bc%81%e4%b8%9a%e7%a7%81%e6%9c%89-rag-%e7%9f%a5%e8%af%86%e5%ba%93","status":"publish","type":"post","link":"https:\/\/flycloud.io\/en\/2026\/01\/07\/%e4%bd%bf%e7%94%a8%e9%98%bf%e9%87%8c%e4%ba%91%e7%99%be%e7%82%bc%e6%9e%84%e5%bb%ba%e4%bc%81%e4%b8%9a%e7%a7%81%e6%9c%89-rag-%e7%9f%a5%e8%af%86%e5%ba%93\/","title":{"rendered":"\u4f7f\u7528\u963f\u91cc\u4e91\u767e\u70bc\u6784\u5efa\u4f01\u4e1a\u79c1\u6709 RAG \u77e5\u8bc6\u5e93"},"content":{"rendered":"

\u4e00\u3001\u524d\u8a00<\/h2>\n\n\n\n

\u901a\u7528\u5927\u6a21\u578b\u5728\u4f01\u4e1a\u4e1a\u52a1\u573a\u666f\u4e2d\u5b58\u5728\u4e09\u5927\u6838\u5fc3\u77ed\u677f\uff1a\u4f01\u4e1a\u5185\u90e8\u8425\u6536\u6570\u636e\u3001\u5185\u90e8\u89c4\u8303\u3001\u9879\u76ee\u6587\u6863\u7b49\u6838\u5fc3\u8d44\u6599\u65e0\u6cd5\u5bf9\u5916\u4f20\u8f93\uff0c\u5b58\u5728\u6570\u636e\u9690\u79c1\u6cc4\u9732\u98ce\u9669\uff1b\u901a\u7528\u6a21\u578b\u8bad\u7ec3\u6570\u636e\u8986\u76d6\u8303\u56f4\u6709\u9650\uff0c\u65e0\u6cd5\u7cbe\u51c6\u5339\u914d\u884c\u4e1a\u5782\u76f4\u4e13\u4e1a\u5185\u5bb9\uff1b\u6a21\u578b\u77e5\u8bc6\u5e93\u66f4\u65b0\u6ede\u540e\uff0c\u65e0\u6cd5\u54cd\u5e94\u4f01\u4e1a\u6700\u65b0\u653f\u7b56\u3001\u4ea7\u54c1\u8fed\u4ee3\u4fe1\u606f\u3002RAG\u68c0\u7d22\u589e\u5f3a\u751f\u6210\u6280\u672f\u662f\u89e3\u51b3\u4ee5\u4e0a\u75db\u70b9\u6807\u51c6\u5316\u65b9\u6848\uff0c\u901a\u8fc7\u5148\u68c0\u7d22\u79c1\u6709\u6587\u6863\u518d\u4f20\u5165\u6a21\u578b\u751f\u6210\u7b54\u6848\uff0c\u8ba9\u5927\u6a21\u578b\u7cbe\u51c6\u8c03\u7528\u4f01\u4e1a\u4e13\u5c5e\u6570\u636e\uff0c\u517c\u987e\u9690\u79c1\u5b89\u5168\u3001\u9886\u57df\u4e13\u4e1a\u6027\u4e0e\u5185\u5bb9\u5b9e\u65f6\u6027\u3002<\/p>\n\n\n\n

2026\u5e74\u963f\u91cc\u4e91\u767e\u70bc\u5927\u6a21\u578b\u5e73\u53f0\u5185\u7f6e\u539f\u751fRAG\u77e5\u8bc6\u5e93\u80fd\u529b\uff0c\u914d\u5957\u5411\u91cf\u5b58\u50a8\u3001\u6587\u6863\u89e3\u6790\u3001\u68c0\u7d22\u91cd\u6392\u5e8f\u3001\u957f\u671f\u8bb0\u5fc6\u63d2\u4ef6\u7b49\u4e00\u4f53\u5316\u5de5\u5177\uff0c\u517c\u5bb9OpenClaw\u3001Hermes\u7b49\u4e3b\u6d41AI\u667a\u80fd\u4f53\u6846\u67b6\u3002\u672c\u6587\u57fa\u4e8ePython+LangChain\u6280\u672f\u6808\uff0c\u5b8c\u6574\u8bb2\u89e3\u79c1\u6709RAG\u77e5\u8bc6\u5e93\u642d\u5efa\u5168\u6d41\u7a0b\uff0c\u5305\u542b\u73af\u5883\u4f9d\u8d56\u3001\u591a\u7c7b\u578b\u6587\u6863\u89e3\u6790\u3001\u5411\u91cf\u5165\u5e93\u3001\u68c0\u7d22\u95ee\u7b54\u3001\u591a\u6570\u636e\u6e90\u62d3\u5c55\u3001\u6548\u679c\u4f18\u5316\u3001\u7f51\u9875\u53ef\u89c6\u5316\u90e8\u7f72\u5168\u5957\u53ef\u8fd0\u884c\u4ee3\u7801\uff0c\u5168\u7a0b\u65e0\u5916\u90e8\u94fe\u63a5\u3001\u8868\u683c\u3001\u56fe\u7247\uff0c\u4e0d\u5305\u542b\u5176\u4ed6\u4e91\u5382\u5546\u76f8\u5173\u5185\u5bb9\uff0c\u9002\u914d\u4e2a\u4eba\u5f00\u53d1\u3001\u4e2d\u5c0f\u4f01\u4e1a\u79c1\u6709\u5316\u77e5\u8bc6\u5e93\u843d\u5730\u4f7f\u7528\u3002<\/p>\n\n\n\n

\u4e8c\u3001RAG\u6838\u5fc3\u5de5\u4f5c\u539f\u7406<\/h2>\n\n\n\n

RAG\u5168\u79f0Retrieval-Augmented Generation\uff0c\u68c0\u7d22\u589e\u5f3a\u751f\u6210\uff0c\u6574\u5957\u94fe\u8def\u5206\u4e3a\u4e94\u5927\u56fa\u5b9a\u73af\u8282\uff1a<\/p>\n\n\n\n

    \n
  1. \u6587\u6863\u8f7d\u5165\uff1a\u89e3\u6790PDF\u3001TXT\u3001DOCX\u3001\u7f51\u9875\u3001\u77e5\u8bc6\u5e93\u7b49\u591a\u683c\u5f0f\u539f\u59cb\u6587\u4ef6\uff1b<\/li>\n\n\n\n
  2. \u6587\u672c\u5206\u5757\uff1a\u8d85\u957f\u6587\u6863\u5207\u5272\u4e3a\u56fa\u5b9a\u957f\u5ea6\u6587\u672c\u7247\u6bb5\uff0c\u907f\u514d\u4e0a\u4e0b\u6587\u8d85\u9650\uff1b<\/li>\n\n\n\n
  3. \u5d4c\u5165\u5411\u91cf\u5316\uff1a\u5c06\u6587\u672c\u8f6c\u6362\u4e3a\u6d6e\u70b9\u5411\u91cf\uff0c\u5b58\u5165\u5411\u91cf\u6570\u636e\u5e93\uff1b<\/li>\n\n\n\n
  4. \u68c0\u7d22\u53ec\u56de\uff1a\u7528\u6237\u63d0\u95ee\u65f6\uff0c\u5c06\u95ee\u9898\u540c\u6b65\u5411\u91cf\u5316\uff0c\u5339\u914d\u5411\u91cf\u5e93\u76f8\u4f3c\u5ea6\u6700\u9ad8\u7684\u6587\u6863\u7247\u6bb5\uff1b<\/li>\n\n\n\n
  5. \u589e\u5f3a\u751f\u6210\uff1a\u628a\u68c0\u7d22\u5230\u7684\u76f8\u5173\u6587\u6863\u4f5c\u4e3a\u4e0a\u4e0b\u6587\u9644\u52a0\u81f3\u7528\u6237\u95ee\u9898\uff0c\u4f20\u5165\u5927\u6a21\u578b\u751f\u6210\u8d34\u5408\u4f01\u4e1a\u4e1a\u52a1\u7684\u7cbe\u51c6\u56de\u7b54\u3002
    \u65e0RAG\u52a0\u6301\u7684\u901a\u7528\u6a21\u578b\u4ec5\u4f9d\u9760\u8bad\u7ec3\u6570\u636e\u4f5c\u7b54\uff0c\u65e0\u6cd5\u83b7\u53d6\u4f01\u4e1a\u5185\u90e8\u79c1\u6709\u5185\u5bb9\uff1b\u63a5\u5165\u79c1\u6709\u5411\u91cf\u77e5\u8bc6\u5e93\u540e\uff0c\u6a21\u578b\u6240\u6709\u56de\u7b54\u5747\u4f9d\u6258\u81ea\u6709\u6587\u6863\uff0c\u5f7b\u5e95\u89e3\u51b3\u4fe1\u606f\u6ede\u540e\u3001\u9886\u57df\u77e5\u8bc6\u7f3a\u5931\u3001\u6570\u636e\u5916\u6cc4\u95ee\u9898\u3002
    <\/li>\n\n\n\n
  6. langchain\uff1a\u5c01\u88c5RAG\u5b8c\u6574\u94fe\u8def\uff0c\u7edf\u4e00\u6587\u6863\u52a0\u8f7d\u3001\u68c0\u7d22\u3001\u95ee\u7b54\u63a5\u53e3\uff1b<\/li>\n\n\n\n
  7. pypdf\/python-docx\uff1a\u5206\u522b\u89e3\u6790PDF\u3001Word\u6587\u6863\uff1b<\/li>\n\n\n\n
  8. chromadb\uff1a\u672c\u5730\u8f7b\u91cf\u5316\u5411\u91cf\u6570\u636e\u5e93\uff0c\u65e0\u9700\u989d\u5916\u90e8\u7f72\u670d\u52a1\uff1b<\/li>\n\n\n\n
  9. streamlit\uff1a\u5feb\u901f\u642d\u5efa\u7f51\u9875\u95ee\u7b54\u524d\u7aef\uff1b<\/li>\n\n\n\n
  10. cohere\uff1a\u63d0\u4f9b\u68c0\u7d22\u91cd\u6392\u5e8f\u80fd\u529b\uff0c\u63d0\u5347\u53ec\u56de\u5185\u5bb9\u76f8\u5173\u6027\uff1b<\/li>\n\n\n\n
  11. python-dotenv\uff1a\u7edf\u4e00\u7ba1\u7406\u6a21\u578b\u5bc6\u94a5\u3001\u73af\u5883\u53d8\u91cf\uff0c\u907f\u514d\u786c\u7f16\u7801\u51ed\u8bc1\u3002<\/li>\n<\/ol>\n\n\n\n

    \u56db\u3001\u79c1\u6709\u77e5\u8bc6\u5e93\u5b8c\u6574\u6838\u5fc3\u4ee3\u7801\u5b9e\u73b0<\/h2>\n\n\n\n

    4.1 \u9879\u76ee\u57fa\u7840\u7c7b\uff1a\u6587\u6863\u52a0\u8f7d\u3001\u5206\u5757\u3001\u5411\u91cf\u5e93\u3001\u95ee\u7b54\u94fe\u8def\u5c01\u88c5<\/h3>\n\n\n\n

    \u65b0\u5efakb_main.py\u6587\u4ef6\uff0c\u5b8c\u6574\u53ef\u8fd0\u884c\u4ee3\u7801\uff1a<\/p>\n\n\n\n

    \"\"\"\n\u4f01\u4e1a\u79c1\u6709RAG\u77e5\u8bc6\u5e93\u7cfb\u7edf\n\u652f\u6301PDF\/TXT\/DOCX\u591a\u683c\u5f0f\u6587\u6863\uff0c\u517c\u5bb9\u767e\u70bc\u901a\u4e49\u7cfb\u5217\u6a21\u578b\n\u5185\u7f6e\u5411\u91cf\u6301\u4e45\u5316\u3001\u6df7\u5408\u68c0\u7d22\u3001\u95ee\u7b54\u8fd4\u56de\u6765\u6e90\u6587\u6863\u80fd\u529b\n\"\"\"\nimport os\nfrom dotenv import load_dotenv\nfrom langchain.document_loaders import PyPDFLoader, TextLoader, UnstructuredWordDocumentLoader\nfrom langchain.text_splitter import RecursiveCharacterTextSplitter\nfrom langchain.embeddings.openai import OpenAIEmbeddings\nfrom langchain.vectorstores import Chroma\nfrom langchain.chat_models import ChatOpenAI\nfrom langchain.chains import RetrievalQA\nfrom langchain.retrievers import EnsembleRetriever, BM25Retriever\nfrom cohere import Client as CohereClient\n\n# \u52a0\u8f7d\u672c\u5730\u5bc6\u94a5\u73af\u5883\u6587\u4ef6\nload_dotenv()\n\nclass PrivateRagKnowledgeBase:\n    def __init__(self, persist_dir=\".\/company_vector_db\"):\n        # \u5411\u91cf\u6301\u4e45\u5316\u5b58\u50a8\u8def\u5f84\n        self.persist_directory = persist_dir\n        # \u521d\u59cb\u5316\u5d4c\u5165\u6a21\u578b\n        self.embeddings = OpenAIEmbeddings()\n        self.vectordb = None\n        self.qa_chain = None\n        self.bm25_retriever = None\n        self.cohere_client = CohereClient(api_key=os.getenv(\"COHERE_KEY\"))\n        # \u63a5\u5165\u963f\u91cc\u4e91\u767e\u70bc\u901a\u4e49\u5927\u6a21\u578b\n        self.llm = ChatOpenAI(\n            base_url=os.getenv(\"BAILIAN_BASE_URL\"),\n            api_key=os.getenv(\"BAILIAN_TOKEN\"),\n            model=os.getenv(\"BAILIAN_MODEL\"),\n            temperature=0.25,\n            max_tokens=1200\n        )\n\n    def load_all_docs(self, folder_path: str):\n        \"\"\"\u904d\u5386\u76ee\u5f55\u52a0\u8f7d\u5168\u90e8\u6587\u6863\uff0c\u652f\u6301pdf\/txt\/doc\/docx\"\"\"\n        doc_list = []\n        for root, _, files in os.walk(folder_path):\n            for file_name in files:\n                full_path = os.path.join(root, file_name)\n                try:\n                    if file_name.endswith(\".pdf\"):\n                        loader = PyPDFLoader(full_path)\n                    elif file_name.endswith(\".txt\"):\n                        loader = TextLoader(full_path, encoding=\"utf-8\")\n                    elif file_name.endswith((\".docx\", \".doc\")):\n                        loader = UnstructuredWordDocumentLoader(full_path)\n                    else:\n                        continue\n                    doc_list.extend(loader.load())\n                    print(f\"\u6210\u529f\u8f7d\u5165\u6587\u4ef6\uff1a{file_name}\")\n                except Exception as err:\n                    print(f\"\u6587\u4ef6{file_name}\u8f7d\u5165\u5931\u8d25\uff0c\u9519\u8bef\u4fe1\u606f\uff1a{str(err)}\")\n        return doc_list\n\n    def split_text_chunk(self, docs, chunk_size=600, chunk_overlap=60, file_type=\"general\"):\n        \"\"\"\u5206\u5757\u7b56\u7565\u533a\u5206\u901a\u7528\/\u6280\u672f\/\u653f\u7b56\u6587\u6863\"\"\"\n        if file == \"tech\":\n            split_rule = RecursiveCharacterTextSplitter(\n                chunk_size=300,\n                chunk_overlap=30,\n                separators=[\"\\n\\n\", \"\\n\", \"```\", \" \"]\n            )\n        elif file == \"policy\":\n            split_rule = RecursiveCharacterTextSplitter(\n                chunk_size=1000,\n                chunk_overlap=100,\n                separators=[\"\\n\\n\\n\", \"\\n\\n\", \"\u3002\", \"\uff1f\"]\n            )\n        else:\n            split_rule = RecursiveCharacterTextSplitter(\n                chunk_size=chunk_size,\n                chunk_overlap=chunk_overlap\n            )\n        return split_rule.split_documents(docs)\n\n    def build_vector_storage(self, chunk_data):\n        \"\"\"\u6784\u5efa\u672c\u5730\u5411\u91cf\u5e93\u5e76\u6301\u4e45\u5316\u4fdd\u5b58\"\"\"\n        print(\"\u6b63\u5728\u751f\u6210\u6587\u672c\u5411\u91cf\uff0c\u5b58\u5165\u5411\u91cf\u6570\u636e\u5e93...\")\n        self.vectordb = Chroma.from_documents(\n            documents=chunk_data,\n            embedding=self.embeddings,\n            persist_directory=self.persist_directory\n        )\n        self.vectordb.persist()\n        print(f\"\u5411\u91cf\u5e93\u6784\u5efa\u5b8c\u6210\uff0c\u5b58\u50a8\u8def\u5f84\uff1a{self.persist_directory}\")\n        # \u521d\u59cb\u5316BM25\u5173\u952e\u8bcd\u68c0\u7d22\u5668\u7528\u4e8e\u6df7\u5408\u68c0\u7d22\n        self.bm25_retriever = BM25Retriever.from_documents(chunk_data)\n        return self.vectordb\n\n    def load_exist_vector_db(self):\n        \"\"\"\u8bfb\u53d6\u5df2\u6784\u5efa\u5b8c\u6210\u7684\u5411\u91cf\u5e93\uff0c\u65e0\u9700\u91cd\u590d\u5411\u91cf\u5316\"\"\"\n        if os.path.exists(self.persist_directory):\n            self.vectordb = Chroma(\n                persist_directory=self.persist_directory,\n                embedding_function=self.embeddings\n            )\n            print(\"\u672c\u5730\u5411\u91cf\u5e93\u52a0\u8f7d\u6210\u529f\")\n            return True\n        print(\"\u672a\u68c0\u6d4b\u5230\u5411\u91cf\u5e93\uff0c\u8bf7\u5148\u6267\u884c\u6587\u6863\u6784\u5efa\u6d41\u7a0b\")\n        return False\n\n    def rerank_doc(self, query, source_docs, top_n=3):\n        \"\"\"\u91cd\u6392\u5e8f\u4f18\u5316\u68c0\u7d22\u7ed3\u679c\uff0c\u8fc7\u6ee4\u4f4e\u76f8\u5173\u7247\u6bb5\"\"\"\n        doc_texts = [doc.page_content for doc in source_docs]\n        rerank_res = self.cohere.rerank(\n            query=query,\n            documents=doc_texts,\n            top_n=top_n\n        )\n        filtered_docs = [source_docs[res.index] for res in rerank_res.results]\n        return filtered_docs\n\n    def create_ensemble_retrieval(self, top_k=5):\n        \"\"\"\u6df7\u5408\u68c0\u7d22\uff1a\u5411\u91cf\u8bed\u4e49\u68c0\u7d22+\u5173\u952e\u8bcdBM25\u68c0\u7d22\"\"\"\n        vector_ret = self.vectordb.as_retriever(search_kwargs={\n   \"k\": top_k})\n        ensemble_ret = EnsembleRetriever(\n            retrievers=[vector_ret, self.bm25_retriever],\n            weights=[0.72, 0.28]\n        )\n        return ensemble_ret\n\n    def build_qa_chain(self, top_k=5):\n        \"\"\"\u642d\u5efa\u5b8c\u6574\u95ee\u7b54\u6267\u884c\u94fe\u8def\"\"\"\n        if not self.vectordb:\n            raise Exception(\"\u8bf7\u5148\u52a0\u8f7d\u6216\u6784\u5efa\u5411\u91cf\u6570\u636e\u5e93\")\n        retriever = self.create_ensemble_retrieval(top_k)\n        self.qa_chain = RetrievalQA.from_chain_type(\n            llm=self.llm,\n            chain_type=\"stuff\",\n            retriever=retriever,\n            return_source_documents=True\n        )\n        print(\"\u77e5\u8bc6\u5e93\u95ee\u7b54\u94fe\u8def\u521d\u59cb\u5316\u5b8c\u6210\uff0c\u53ef\u53d1\u8d77\u63d0\u95ee\")\n\n    def query_knowledge(self, user_question):\n        \"\"\"\u5bf9\u5916\u7edf\u4e00\u63d0\u95ee\u63a5\u53e3\uff0c\u8fd4\u56de\u7b54\u6848\u4e0e\u53c2\u8003\u6587\u6863\u7247\u6bb5\"\"\"\n        if not self.qa_chain:\n            raise Exception(\"\u95ee\u7b54\u94fe\u8def\u672a\u521d\u59cb\u5316\uff0c\u8bf7\u6267\u884cbuild_qa_chain\")\n        raw_res = self.qa_chain.invoke({\n   \"query\": user_question})\n        filtered_sources = self.rerank_doc(user_question, raw[\"source_documents\"])\n        source_summary = [doc.page_content[:220] + \"\u2026\u2026\" for doc in filtered_sources]\n        return {\n   \n            \"answer\": raw_res[\"result\"],\n            \"reference_chunk\": source_summary\n        }\n\n# \u9879\u76ee\u4e3b\u6267\u884c\u5165\u53e3\nif __name__ == \"__main__\":\n    # \u521d\u59cb\u5316\u77e5\u8bc6\u5e93\u5b9e\u4f8b\n    kb = PrivateRagKnowledgeBase(persist_dir=\".\/company_vector_db\")\n    # \u6d41\u7a0b1\uff1a\u9996\u6b21\u4f7f\u7528\uff0c\u8f7d\u5165\u6587\u6863\u5e76\u6784\u5efa\u5411\u91cf\u5e93\n    all_docs = kb.load_all_docs(\".\/company_docs\")\n    text_chunks = kb.split_text_chunk(all_docs, file_type=\"general\")\n    kb.build_vector_storage(text_chunks)\n    kb.build_qa_chain(top_k=5)\n    # \u6d41\u7a0b2\uff1a\u540e\u7eed\u542f\u52a8\u76f4\u63a5\u52a0\u8f7d\u5df2\u6709\u5411\u91cf\u5e93\uff0c\u6ce8\u91ca\u4e0a\u65b9\u542f\u7528\u4e0b\u65b9\n    # kb.load_exist_vector_db()\n    # kb.build_qa_chain()\n    # \u4ea4\u4e92\u95ee\u7b54\u5faa\u73af\n    print(\"===\u4f01\u4e1a\u79c1\u6709\u77e5\u8bc6\u5e93\u95ee\u7b54\u7cfb\u7edf\u542f\u52a8\uff0c\u8f93\u5165exit\u9000\u51fa===\")\n    while True:\n        input_q = input(\"\u8bf7\u8f93\u5165\u4e1a\u52a1\u95ee\u9898\uff1a\")\n        if input_q.strip().lower() == \"exit\":\n            break\n        output = kb.query_knowledge(input_q)\n        print(f\"\\n\u56de\u7b54\u5185\u5bb9\uff1a\\n{output['answer']}\")\n        print(f\"\\n\u53c2\u8003\u6587\u6863\u7247\u6bb5\uff08\u5171{len(output['reference_chunk'])}\u6761\uff09\uff1a\")\n        for idx, text in enumerate(output[\"reference_chunk\"], 1):\n            print(f\"{idx}. {text}\\n\")\n<\/code><\/pre>\n\n\n\n
    4.2 \u73af\u5883\u53d8\u91cf\u914d\u7f6e\u6587\u4ef6<\/h3>\n\n\n\n
    \u5728\u9879\u76ee\u6839\u76ee\u5f55\u65b0\u5efa.env\u6587\u4ef6\uff0c\u586b\u5165\u767e\u70bc\u6a21\u578b\u8bbf\u95ee\u51ed\u8bc1\u3001\u5d4c\u5165\u6a21\u578b\u5bc6\u94a5\uff0c\u907f\u514d\u4ee3\u7801\u786c\u7f16\u7801\u654f\u611f\u4fe1\u606f\uff1a<\/p>\n\n\n\n
    BAILIAN_BASE_URL=https:\/\/compatible-mode.bailian.aliyuncs.com\/v1\nBAILIAN_TOKEN=\u4f60\u7684\u767e\u70bcToken Plan\u5bc6\u94a5\nBAIL_MODEL=qwen3.7-plus\nCOHERE_KEY=\u91cd\u6392\u5e8f\u670d\u52a1\u5bc6\u94a5\n<\/code><\/pre>\n\n\n\n
    4.3 \u591a\u6570\u636e\u6e90\u62d3\u5c55\u4ee3\u7801<\/h3>\n\n\n\n
    \u57fa\u7840\u4ee3\u7801\u4ec5\u652f\u6301\u672c\u5730\u6587\u4ef6\uff0c\u53ef\u62d3\u5c55\u7f51\u9875\u3001\u5185\u90e8\u77e5\u8bc6\u5e93\u7b49\u5916\u90e8\u6570\u636e\u6e90\u8f7d\u5165\u903b\u8f91\uff1a<\/p>\n\n\n\n
    # \u7f51\u9875\u6587\u6863\u8f7d\u5165\u62d3\u5c55\nfrom langchain.document_loaders import UnstructuredURLLoader\ndef load_web_content(url_list):\n    loader = UnstructuredURLLoader(urls=url_list)\n    return loader.load()\n\n# Notion\/\u5185\u90e8\u77e5\u8bc6\u5e93\u8f7d\u5165\u62d3\u5c55\nfrom langchain.document_loaders import NotionLoader\ndef load_notion_data(notion_url):\n    loader = NotionLoader(notion_url)\n    return loader.load()\n<\/code><\/pre>\n\n\n\n
    \u4e94\u3001\u7f51\u9875\u53ef\u89c6\u5316\u524d\u7aef\u90e8\u7f72\u4ee3\u7801<\/h2>\n\n\n\n
    \u501f\u52a9Streamlit\u5feb\u901f\u642d\u5efa\u6d4f\u89c8\u5668\u8bbf\u95ee\u754c\u9762\uff0c\u65b0\u5efaweb_ui.py\uff1a<\/p>\n\n\n\n
    import streamlit as st\nfrom kb_main import PrivateRagKnowledgeBase\n\n# \u9875\u9762\u57fa\u7840\u914d\u7f6e\nst.set_page_config(page_title=\"\u4f01\u4e1a\u79c1\u6709\u77e5\u8bc6\u5e93\", layout=\"wide\")\nst.title(\"\ud83c\udfe2 \u4f01\u4e1a\u79c1\u6709RAG\u77e5\u8bc6\u5e93\u95ee\u7b54\u5e73\u53f0\")\n\n# \u5168\u5c40\u7f13\u5b58\u77e5\u8bc6\u5e93\u5b9e\u4f8b\nif \"knowledge_engine\" not in st.session_state:\n    st.session_state.knowledge_engine = PrivateRagKnowledgeBase()\n    load_status = st.session_state.knowledge_engine.load_exist_vector()\n    if load_status:\n        st.session_state.knowledge_engine.build_qa_chain()\n        st.success(\"\u5411\u91cf\u5e93\u52a0\u8f7d\u5b8c\u6210\uff0c\u53ef\u76f4\u63a5\u63d0\u95ee\")\n    else:\n        st.warning(\"\u672a\u68c0\u6d4b\u5411\u91cf\u5e93\uff0c\u8bf7\u5148\u4e0a\u4f20\u6587\u6863\u6784\u5efa\u77e5\u8bc6\u5e93\")\n\n# \u95ee\u7b54\u8f93\u5165\u6846\nuser_input = st.text_input(\"\u8bf7\u8f93\u5165\u4e1a\u52a1\u76f8\u5173\u95ee\u9898\uff1a\")\nif user_input:\n    with st.spinner(\"\u6b63\u5728\u68c0\u7d22\u79c1\u6709\u6587\u6863\u5e76\u751f\u6210\u56de\u7b54...\"):\n        res_data = st.session_state.knowledge_engine.query_knowledge(user_input)\n    st.subheader(\"\u667a\u80fd\u56de\u7b54\")\n    st.write(res_data[\"answer\"])\n    # \u5c55\u5f00\u67e5\u770b\u53c2\u8003\u6587\u6863\n    with st.expander(\"\u67e5\u770b\u68c0\u7d22\u53c2\u8003\u6587\u6863\u7247\u6bb5\"):\n        for num, content in enumerate(res_data[\"reference_chunk\"], 1):\n            st.text(f\"{num}. {content}\")\n<\/code><\/pre>\n\n\n\n
    \u542f\u52a8\u7f51\u9875\u670d\u52a1\u6267\u884c\u547d\u4ee4\uff1a<\/p>\n\n\n\n
    streamlit run web_ui --server.port 8501\n<\/code><\/pre>\n\n\n\n
    \u542f\u52a8\u540e\u672c\u5730\u6d4f\u89c8\u5668\u8bbf\u95ee\u5bf9\u5e94\u7aef\u53e3\u5730\u5740\u5373\u53ef\u53ef\u89c6\u5316\u64cd\u4f5c\u77e5\u8bc6\u5e93\u95ee\u7b54\u3002<\/p>\n\n\n\n
    \u516d\u3001RAG\u68c0\u7d22\u6548\u679c\u5206\u5c42\u4f18\u5316\u65b9\u6848<\/h2>\n\n\n\n
    6.1 \u5dee\u5f02\u5316\u6587\u672c\u5206\u5757\u7b56\u7565<\/h3>\n\n\n\n
    \u6587\u6863\u7c7b\u578b\u76f4\u63a5\u5f71\u54cd\u68c0\u7d22\u7cbe\u51c6\u5ea6\uff0c\u4ee3\u7801\u5185\u5df2\u533a\u5206\u4e09\u7c7b\u5206\u5757\u89c4\u5219\uff1a<\/p>\n\n\n\n
      \n
    1. \u6280\u672f\u6587\u6863\/\u4ee3\u7801\u624b\u518c\uff1a\u5c0f\u5206\u5757300\u5b57\u7b26\uff0c\u51cf\u5c11\u8de8\u4ee3\u7801\u7247\u6bb5\u5207\u5272\uff1b<\/li>\n\n\n\n
    2. \u653f\u7b56\u3001\u5408\u540c\u7c7b\u957f\u6587\u672c\uff1a\u5927\u5206\u57571000\u5b57\u7b26\uff0c\u4fdd\u8bc1\u5b8c\u6574\u8bed\u4e49\u6bb5\u843d\uff1b<\/li>\n\n\n\n
    3. \u901a\u7528\u4e1a\u52a1\u6587\u6863\uff1a60\u5b57\u7b26\u91cd\u53e0\uff0c\u517c\u987e\u4e0a\u4e0b\u6587\u8fde\u7eed\u6027\u30026.2 \u6df7\u5408\u68c0\u7d22\u67b6\u6784 \u653e\u5f03\u5355\u4e00\u5411\u91cf\u68c0\u7d22\u6a21\u5f0f\uff0c\u91c7\u752872%\u8bed\u4e49\u5411\u91cf\u68c0\u7d22+28%\u5173\u952e\u8bcdBM25\u68c0\u7d22\u52a0\u6743\u878d\u5408\uff0c\u517c\u987e\u8bed\u4e49\u76f8\u4f3c\u5ea6\u4e0e\u5173\u952e\u8bcd\u7cbe\u51c6\u5339\u914d\uff0c\u89e3\u51b3\u7eaf\u5411\u91cf\u68c0\u7d22\u9057\u6f0f\u4e13\u6709\u540d\u8bcd\u7684\u95ee\u9898\u30026.3 \u68c0\u7d22\u7ed3\u679c\u91cd\u6392\u5e8f \u8c03\u7528\u91cd\u6392\u5e8f\u63a5\u53e3\u5bf9\u521d\u6b21\u53ec\u56de\u6587\u6863\u4e8c\u6b21\u6253\u5206\uff0c\u8fc7\u6ee4\u4f4e\u76f8\u5173\u6027\u7247\u6bb5\uff0c\u51cf\u5c11\u65e0\u5173\u5185\u5bb9\u6df7\u5165\u6a21\u578b\u4e0a\u4e0b\u6587\uff0c\u5927\u5e45\u964d\u4f4e\u6a21\u578b\u5e7b\u89c9\u6982\u7387\u30026. \u6a21\u578b\u53c2\u6570\u8c03\u4f18 \u767e\u70bc\u901a\u4e49\u6a21\u578btemperature\u8bbe\u7f6e0.25\uff0c\u964d\u4f4e\u968f\u673a\u521b\u4f5c\u503e\u5411\uff0c\u5f3a\u5236\u4f9d\u6258\u68c0\u7d22\u6587\u6863\u4f5c\u7b54\uff0c\u51cf\u5c11\u65e0\u4f9d\u636e\u865a\u6784\u5185\u5bb9\u3002<\/li>\n<\/ol>\n\n\n\n
      \u4e03\u3001\u5bf9\u63a5OpenClaw\/Hermes\u667a\u80fd\u4f53\u62d3\u5c55\u65b9\u6848<\/h2>\n\n\n\n
      \u767e\u70bcRAG\u77e5\u8bc6\u5e93\u53ef\u4f5c\u4e3a\u63d2\u4ef6\u63a5\u5165\u4e24\u6b3e\u4e3b\u6d41AI\u667a\u80fd\u4f53\uff0c\u5b9e\u73b0\u81ea\u52a8\u6587\u6863\u68c0\u7d22\u3001\u4e1a\u52a1\u95ee\u7b54\u81ea\u52a8\u5316\uff0c\u4ee5Hermes\u914d\u7f6e\u793a\u4f8b\uff1a<\/p>\n\n\n\n
        \n
      1. \u5728Hermes .env\u914d\u7f6e\u6587\u4ef6\u65b0\u589e\u77e5\u8bc6\u5e93\u63a5\u53e3\u53c2\u6570\uff1aRAG_KB_URL=http:\/\/127.0.0.1:8501\/api\/query RAG_ENABLE=true RAG_TOP_NUM=3<\/code><\/li>\n\n\n\n
      2. Hermes\u4efb\u52a1\u6267\u884c\u524d\u81ea\u52a8\u8c03\u7528\u672c\u5730RAG\u63a5\u53e3\uff0c\u68c0\u7d22\u4e1a\u52a1\u6587\u6863\u540e\u518d\u53d1\u8d77\u6a21\u578b\u8bf7\u6c42\uff1b<\/li>\n\n\n\n
      3. OpenClaw\u53ef\u901a\u8fc7Skill\u5e02\u573a\u5b89\u88c5\u767e\u70bcRAG\u4e13\u5c5e\u63d2\u4ef6\uff0c\u586b\u5165\u5411\u91cf\u5e93\u670d\u52a1\u5730\u5740\u4e00\u952e\u542f\u7528\u79c1\u6709\u6587\u6863\u68c0\u7d22\u80fd\u529b\uff0c\u667a\u80fd\u4f53\u6267\u884c\u62a5\u8868\u3001\u65b9\u6848\u64b0\u5199\u7b49\u4efb\u52a1\u65f6\u81ea\u52a8\u8c03\u53d6\u4f01\u4e1a\u5185\u90e8\u8d44\u6599\u3002<\/li>\n<\/ol>\n\n\n\n
        \u516b\u3001\u5e38\u89c1\u6545\u969c\u6392\u67e5\u4e0e\u4f18\u5316\u5efa\u8bae<\/h2>\n\n\n\n
        8.1 \u68c0\u7d22\u4e0d\u5230\u76f8\u5173\u6587\u6863<\/h3>\n\n\n\n
        \u6392\u67e5\u65b9\u6848\uff1a\u68c0\u67e5\u6587\u6863\u5206\u5757\u5c3a\u5bf8\u662f\u5426\u8fc7\u5927\/\u8fc7\u5c0f\uff0c\u8c03\u6574chunk_overlap\u91cd\u53e0\u5b57\u7b26\uff1b\u63d0\u5347\u6df7\u5408\u68c0\u7d22top_k\u53ec\u56de\u6570\u91cf\uff1b\u8865\u5145\u6587\u6863\u5173\u952e\u8bcd\uff0c\u91cd\u65b0\u6784\u5efa\u5411\u91cf\u5e93\u3002<\/p>\n\n\n\n
        8.2 \u6a21\u578b\u56de\u7b54\u5b58\u5728\u865a\u6784\u4fe1\u606f<\/h3>\n\n\n\n
        \u4f18\u5316\u65b9\u6848\uff1a\u964d\u4f4etemperature\u53c2\u6570\uff1b\u542f\u7528Cohere\u91cd\u6392\u5e8f\u8fc7\u6ee4\u4f4e\u76f8\u5173\u7247\u6bb5\uff1bPrompt\u589e\u52a0\u7ea6\u675f\uff0c\u8981\u6c42\u4ec5\u4f9d\u636e\u68c0\u7d22\u6587\u6863\u4f5c\u7b54\uff0c\u65e0\u76f8\u5173\u5185\u5bb9\u76f4\u63a5\u544a\u77e5\u65e0\u5bf9\u5e94\u8d44\u6599\u3002<\/p>\n\n\n\n
        8. \u95ee\u7b54\u54cd\u5e94\u901f\u5ea6\u7f13\u6162<\/h3>\n\n\n\n
        \u4f18\u5316\u65b9\u6848\uff1a\u5f00\u542f\u5411\u91cf\u5e93\u672c\u5730\u7f13\u5b58\uff1b\u51cf\u5c11\u5355\u6b21\u53ec\u56de\u6587\u6863\u6570\u91cf\uff1b\u62c6\u5206\u8d85\u957f\u6587\u6863\uff0c\u907f\u514d\u5411\u91cf\u5316\u8017\u65f6\u8fc7\u957f\uff1b\u4e1a\u52a1\u9ad8\u5cf0\u5207\u6362\u8f7b\u91cf\u7248Qwen\u6a21\u578b\u3002<\/p>\n\n\n\n
        8. \u6587\u6863\u8f7d\u5165\u62a5\u9519<\/h3>\n\n\n\n
        \u6392\u67e5\uff1a\u786e\u8ba4\u6587\u4ef6\u7f16\u7801\u4e3aUTF-8\uff0c\u635f\u574fPDF\u3001Word\u6587\u4ef6\u91cd\u65b0\u5bfc\u51fa\u540e\u518d\u5165\u5e93\u3002<\/p>\n\n\n\n
        \u4e5d\u3001\u6574\u5957\u65b9\u6848\u843d\u5730\u4ef7\u503c\u603b\u7ed3<\/h2>\n\n\n\n
        \u4f9d\u6258\u767e\u70bc\u5927\u6a21\u578b\u642d\u914dLangChain\u642d\u5efa\u79c1\u6709RAG\u77e5\u8bc6\u5e93\uff0c\u5f7b\u5e95\u89e3\u51b3\u901a\u7528\u5927\u6a21\u578b\u4e1a\u52a1\u76f2\u533a\u3001\u6570\u636e\u9690\u79c1\u6cc4\u9732\u4e24\u5927\u6838\u5fc3\u75db\u70b9\uff0c\u6574\u5957\u65b9\u6848\u8f7b\u91cf\u5316\u3001\u65e0\u590d\u6742\u4e2d\u95f4\u4ef6\u4f9d\u8d56\uff0c\u4e2a\u4eba\u5f00\u53d1\u8005\u3001\u4e2d\u5c0f\u4f01\u4e1a\u5747\u53ef\u5feb\u901f\u843d\u5730\u3002\u6574\u5957\u7cfb\u7edf\u652f\u6301\u591a\u683c\u5f0f\u6587\u6863\u89e3\u6790\u3001\u6df7\u5408\u68c0\u7d22\u3001\u91cd\u6392\u5e8f\u4f18\u5316\u3001\u7f51\u9875\u53ef\u89c6\u5316\u8bbf\u95ee\uff0c\u8fd8\u53ef\u65e0\u7f1d\u5bf9\u63a5OpenClaw\u3001Hermes\u7b49AI\u667a\u80fd\u4f53\u6846\u67b6\uff0c\u5c06\u79c1\u6709\u4e1a\u52a1\u6570\u636e\u8d4b\u80fd\u81ea\u52a8\u5316\u4efb\u52a1\u6267\u884c\u3002
        \u5206\u5c42\u5206\u5757\u7b56\u7565\u3001\u5411\u91cf+\u5173\u952e\u8bcd\u6df7\u5408\u68c0\u7d22\u3001\u7ed3\u679c\u91cd\u6392\u5e8f\u4e09\u5c42\u4f18\u5316\u624b\u6bb5\uff0c\u6709\u6548\u63d0\u5347\u68c0\u7d22\u7cbe\u51c6\u5ea6\uff0c\u51cf\u5c11\u6a21\u578b\u5e7b\u89c9\u3002\u672c\u5730\u5411\u91cf\u5e93\u65e0\u9700\u4f9d\u8d56\u7b2c\u4e09\u65b9\u5b58\u50a8\uff0c\u4f01\u4e1a\u5185\u90e8\u6570\u636e\u5168\u7a0b\u4e0d\u5bf9\u5916\u4f20\u8f93\uff0c\u6ee1\u8db3\u6570\u636e\u5408\u89c4\u8981\u6c42\u3002\u540c\u65f6\u914d\u5957\u5b8c\u6574\u53ef\u590d\u7528Python\u4ee3\u7801\uff0c\u4ece\u6587\u6863\u5165\u5e93\u5230\u524d\u7aef\u95ee\u7b54\u5f62\u6210\u5b8c\u6574\u95ed\u73af\uff0c\u53ef\u6839\u636e\u4e1a\u52a1\u89c4\u6a21\u6269\u5c55\u4e3a\u7ebf\u4e0a\u670d\u52a1\u3001\u667a\u80fd\u4f53\u5185\u7f6e\u77e5\u8bc6\u5e93\u3001\u6279\u91cf\u6587\u6863\u89e3\u6790\u7b49\u591a\u7c7b\u573a\u666f\u3002<\/p>\n\n\n\n
        <\/p>","protected":false},"excerpt":{"rendered":"
        \u4e00\u3001\u524d\u8a00 \u901a\u7528\u5927\u6a21\u578b\u5728\u4f01\u4e1a\u4e1a\u52a1\u573a\u666f\u4e2d\u5b58\u5728\u4e09\u5927\u6838\u5fc3\u77ed\u677f\uff1a\u4f01\u4e1a\u5185\u90e8\u8425\u6536\u6570\u636e\u3001\u5185\u90e8\u89c4\u8303\u3001\u9879\u76ee\u6587\u6863\u7b49\u6838\u5fc3\u8d44\u6599\u65e0\u6cd5\u5bf9\u5916\u4f20\u8f93\uff0c\u5b58\u5728\u6570\u636e\u9690\u79c1\u6cc4\u9732\u98ce\u9669\uff1b\u901a\u7528\u6a21\u578b\u8bad\u7ec3\u6570\u636e\u8986\u76d6\u8303\u56f4\u6709\u9650\uff0c\u65e0\u6cd5\u7cbe\u51c6\u5339\u914d\u884c\u4e1a\u5782\u76f4\u4e13\u4e1a\u5185\u5bb9\uff1b\u6a21\u578b\u77e5\u8bc6\u5e93\u66f4\u65b0\u6ede\u540e\uff0c\u65e0\u6cd5\u54cd\u5e94\u4f01\u4e1a\u6700\u65b0\u653f\u7b56\u3001\u4ea7\u54c1\u8fed\u4ee3\u4fe1\u606f\u3002RAG\u68c0\u7d22\u589e\u5f3a\u751f\u6210\u6280\u672f\u662f\u89e3\u51b3\u4ee5\u4e0a\u75db\u70b9\u6807\u51c6\u5316\u65b9\u6848\uff0c\u901a\u8fc7\u5148\u68c0\u7d22\u79c1\u6709\u6587\u6863\u518d\u4f20\u5165\u6a21\u578b\u751f\u6210\u7b54\u6848\uff0c\u8ba9\u5927\u6a21\u578b\u7cbe\u51c6\u8c03\u7528\u4f01\u4e1a\u4e13\u5c5e\u6570\u636e\uff0c\u517c\u987e\u9690\u79c1\u5b89\u5168\u3001\u9886\u57df\u4e13\u4e1a\u6027\u4e0e\u5185\u5bb9\u5b9e…<\/p>","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5,10],"tags":[],"class_list":["post-381","post","type-post","status-publish","format-standard","hentry","category-ai","category-alibaba-cloud"],"_links":{"self":[{"href":"https:\/\/flycloud.io\/en\/wp-json\/wp\/v2\/posts\/381","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/flycloud.io\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/flycloud.io\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/flycloud.io\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/flycloud.io\/en\/wp-json\/wp\/v2\/comments?post=381"}],"version-history":[{"count":1,"href":"https:\/\/flycloud.io\/en\/wp-json\/wp\/v2\/posts\/381\/revisions"}],"predecessor-version":[{"id":382,"href":"https:\/\/flycloud.io\/en\/wp-json\/wp\/v2\/posts\/381\/revisions\/382"}],"wp:attachment":[{"href":"https:\/\/flycloud.io\/en\/wp-json\/wp\/v2\/media?parent=381"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/flycloud.io\/en\/wp-json\/wp\/v2\/categories?post=381"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/flycloud.io\/en\/wp-json\/wp\/v2\/tags?post=381"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}