GPT-6 \u867d\u7136\u62e5\u6709\u5f3a\u5927\u7684\u901a\u7528\u80fd\u529b，\u4f46\u4f01\u4e1a\u79c1\u6709\u77e5\u8bc6\u5e93\u4e2d\u7684\u4e1a\u52a1\u6570\u636e\u5f80\u5f80\u4e0d\u5728\u6a21\u578b\u8bad\u7ec3\u8303\u56f4\u5185\u3002RAG（\u68c0\u7d22\u589e\u5f3a\u751f\u6210）\u901a\u8fc7\u300c\u5148\u68c0\u7d22\u3001\u540e\u751f\u6210\u300d\u7684\u65b9\u5f0f，\u628a\u4f01\u4e1a\u6587\u6863\u3001\u6570\u636e\u5e93\u4e2d\u7684\u5185\u5bb9\u5b9e\u65f6\u6ce8\u5165\u63d0\u793a\u8bcd，\u8ba9\u6a21\u578b\u57fa\u4e8e\u771f\u5b9e\u4e1a\u52a1\u8d44\u6599\u4f5c\u7b54，\u65e2\u89e3\u51b3\u4e86\u77e5\u8bc6\u65f6\u6548\u6027\u95ee\u9898，\u4e5f\u964d\u4f4e\u4e86\u5e7b\u89c9\u98ce\u9669\u3002<\/p>\n
\u76f8\u6bd4\u76f4\u63a5\u5fae\u8c03，RAG \u5177\u5907\u4e09\u4e2a\u660e\u663e\u4f18\u52bf：\u4e00\u662f\u77e5\u8bc6\u66f4\u65b0\u6210\u672c\u4f4e，\u66ff\u6362\u7d22\u5f15\u5373\u53ef\u751f\u6548；\u4e8c\u662f\u53ef\u8ffd\u6eaf\u6027\u5f3a，\u6bcf\u4e2a\u56de\u7b54\u90fd\u80fd\u5b9a\u4f4d\u5230\u6765\u6e90\u6587\u6863；\u4e09\u662f\u90e8\u7f72\u95e8\u69db\u4f4e，\u65e0\u9700\u6602\u8d35\u7684\u8bad\u7ec3\u7b97\u529b\u3002\u56e0\u6b64，RAG \u6210\u4e3a\u4f01\u4e1a\u843d\u5730\u5927\u6a21\u578b\u5e94\u7528\u7684\u9996\u9009\u67b6\u6784\u3002<\/p>\n
\u4e00\u4e2a\u5b8c\u6574\u7684\u4f01\u4e1a\u7ea7 RAG \u7cfb\u7edf\u5305\u542b\u56db\u4e2a\u6838\u5fc3\u73af\u8282：\u6587\u6863\u52a0\u8f7d\u3001\u5411\u91cf\u5316\u5b58\u50a8\u3001\u68c0\u7d22\u53ec\u56de\u3001\u751f\u6210\u56de\u7b54\u3002\u4e0b\u9762\u7528\u4e00\u5f20\u6d41\u7a0b\u56fe\u5c55\u793a\u6574\u4f53\u94fe\u8def\u3002<\/p>\n
flowchart TD
\n A[\u4f01\u4e1a\u6587\u6863] –> B[\u6587\u6863\u52a0\u8f7d\u4e0e\u5207\u5206]
\n B –> C[\u5411\u91cf\u5316 Embedding]
\n C –> D[\u5411\u91cf\u6570\u636e\u5e93\u5b58\u50a8]
\n E[\u7528\u6237\u63d0\u95ee] –> F[\u95ee\u9898\u5411\u91cf\u5316]
\n F –> G[\u76f8\u4f3c\u5ea6\u68c0\u7d22]
\n D –> G
\n G –> H[\u62fc\u63a5\u4e0a\u4e0b\u6587\u4e0e\u95ee\u9898]
\n H –> I[GPT-6 \u751f\u6210\u56de\u7b54]
\n I –> J[\u8fd4\u56de\u7ed3\u679c\u5e76\u9644\u6765\u6e90]<\/p>\n
\u5176\u4e2d，\u6587\u6863\u5207\u5206\u8d28\u91cf\u3001\u5411\u91cf\u6a21\u578b\u9009\u62e9\u3001\u68c0\u7d22\u7b56\u7565\u548c\u63d0\u793a\u8bcd\u8bbe\u8ba1，\u662f\u51b3\u5b9a\u7cfb\u7edf\u6548\u679c\u7684\u5173\u952e\u56e0\u7d20\u3002\u4e0b\u9762\u9010\u4e00\u5c55\u5f00\u3002<\/p>\n
\u672c\u6559\u7a0b\u4f7f\u7528 Python 3.10 \u53ca\u4ee5\u4e0a\u7248\u672c，\u6838\u5fc3\u4f9d\u8d56\u5305\u62ec LangChain\u3001OpenAI SDK\u3001\u5411\u91cf\u6570\u636e\u5e93\u548c\u6587\u6863\u89e3\u6790\u5e93\u3002\u5efa\u8bae\u4f7f\u7528\u865a\u62df\u73af\u5883\u9694\u79bb\u9879\u76ee\u4f9d\u8d56\u3002<\/p>\n
pip install langchain langchain-openai openai chromadb pypdf<\/p>\n
\u63a5\u4e0b\u6765\u5728\u9879\u76ee\u4e2d\u914d\u7f6e GPT-6 \u7684 API \u5bc6\u94a5\u3002\u63a8\u8350\u4f7f\u7528\u73af\u5883\u53d8\u91cf\u7ba1\u7406\u5bc6\u94a5，\u907f\u514d\u786c\u7f16\u7801\u6cc4\u9732\u3002<\/p>\n
import os
\nfrom langchain_openai import ChatOpenAI, OpenAIEmbeddings
\nos.environ["OPENAI_API_KEY"] = "your-api-key"
\n\u521d\u59cb\u5316 GPT-6 \u5bf9\u8bdd\u6a21\u578b
\nllm = ChatOpenAI(
\nmodel="gpt-6",
\ntemperature=0.2,
\nmax_tokens=2048
\n)
\n\u521d\u59cb\u5316\u5411\u91cf\u5316\u6a21\u578b
\nembeddings = OpenAIEmbeddings(model="text-embedding-3-large")<\/p>\n
\u4f01\u4e1a\u6587\u6863\u683c\u5f0f\u591a\u6837，\u5e38\u89c1\u7684\u6709 PDF\u3001Word\u3001Markdown \u7b49\u3002LangChain \u63d0\u4f9b\u4e86\u7edf\u4e00\u7684\u52a0\u8f7d\u5668\u63a5\u53e3，\u8fd9\u91cc\u4ee5 PDF \u4e3a\u4f8b\u6f14\u793a\u52a0\u8f7d\u6d41\u7a0b\u3002<\/p>\n
from langchain_community.document_loaders import PyPDFLoader
\nloader = PyPDFLoader("enterprise_manual.pdf")
\ndocuments = loader.load()
\nprint(f"\u5171\u52a0\u8f7d {len(documents)} \u9875\u6587\u6863")<\/p>\n
\u6587\u6863\u5207\u5206\u662f\u5f71\u54cd\u68c0\u7d22\u8d28\u91cf\u7684\u6838\u5fc3\u73af\u8282\u3002\u5207\u5206\u8fc7\u5927，\u4e0a\u4e0b\u6587\u566a\u58f0\u591a\u3001\u6210\u672c\u9ad8；\u5207\u5206\u8fc7\u5c0f，\u8bed\u4e49\u4e0d\u5b8c\u6574\u3001\u53ec\u56de\u4e0d\u51c6\u3002\u63a8\u8350\u6309\u7ae0\u8282\u7ed3\u6784\u4f18\u5148\u5207\u5206，\u518d\u7ed3\u5408\u56fa\u5b9a\u957f\u5ea6\u515c\u5e95\u3002<\/p>\n
from langchain.text_splitter import RecursiveCharacterTextSplitter
\ntext_splitter = RecursiveCharacterTextSplitter(
\nchunk_size=800,
\nchunk_overlap=120,
\nseparators=["\\\\n\\\\n", "\\\\n", "\u3002", "！", "？", "；", " ", ""]
\n)
\nchunks = text_splitter.split_documents(documents)
\nprint(f"\u5207\u5206\u540e\u5171 {len(chunks)} \u4e2a\u6587\u672c\u5757")<\/p>\n
\u5207\u5206\u5b8c\u6210\u540e，\u9700\u8981\u628a\u6bcf\u4e2a\u6587\u672c\u5757\u5411\u91cf\u5316\u5e76\u5199\u5165\u5411\u91cf\u6570\u636e\u5e93\u3002\u8fd9\u91cc\u4f7f\u7528 Chroma \u4f5c\u4e3a\u672c\u5730\u5411\u91cf\u5e93，\u4fbf\u4e8e\u5feb\u901f\u9a8c\u8bc1；\u751f\u4ea7\u73af\u5883\u53ef\u66ff\u6362\u4e3a Milvus \u6216 Elasticsearch\u3002<\/p>\n
from langchain_chroma import Chroma
\nvector_store = Chroma.from_documents(
\ndocuments=chunks,
\nembedding=embeddings,
\npersist_directory=".\/chroma_db"
\n)
\nprint("\u5411\u91cf\u7d22\u5f15\u6784\u5efa\u5b8c\u6210")<\/p>\n
\u6784\u5efa\u7d22\u5f15\u65f6\u9700\u8981\u6ce8\u610f\u4e24\u70b9：\u4e00\u662f\u4e3a\u6bcf\u4e2a\u6587\u672c\u5757\u4fdd\u7559\u6765\u6e90\u5143\u6570\u636e，\u4fbf\u4e8e\u56de\u7b54\u65f6\u8ffd\u6eaf；\u4e8c\u662f\u5b9a\u671f\u91cd\u5efa\u7d22\u5f15，\u4fdd\u8bc1\u65b0\u589e\u6587\u6863\u80fd\u53ca\u65f6\u88ab\u68c0\u7d22\u5230\u3002<\/p>\n
\u57fa\u7840\u68c0\u7d22\u4f7f\u7528\u5411\u91cf\u76f8\u4f3c\u5ea6\u53ec\u56de Top-K \u6587\u672c\u5757\u3002\u4f46\u4f01\u4e1a\u573a\u666f\u4e0b，\u5355\u7eaf\u4f9d\u8d56\u5411\u91cf\u68c0\u7d22\u5f80\u5f80\u4e0d\u591f\u7cbe\u51c6，\u5efa\u8bae\u5f15\u5165\u91cd\u6392\u5e8f\u6a21\u578b\u5bf9\u53ec\u56de\u7ed3\u679c\u4e8c\u6b21\u6253\u5206\u3002<\/p>\n
from langchain.retrievers import ContextualCompressionRetriever
\nfrom langchain.retrievers.document_compressors import CrossEncoderReranker
\nfrom langchain_community.cross_encoders import HuggingFaceCrossEncoder
\nbase_retriever = vector_store.as_retriever(search_kwargs={"k": 8})
\nreranker = CrossEncoderReranker(
\nmodel=HuggingFaceCrossEncoder(model_name="BAAI\/bge-reranker-v2-m3"),
\ntop_n=4
\n)
\ncompression_retriever = ContextualCompressionRetriever(
\nbase_compressor=reranker,
\nbase_retriever=base_retriever
\n)<\/p>\n
\u7ecf\u8fc7\u91cd\u6392\u5e8f\u540e，\u7cfb\u7edf\u53ea\u4fdd\u7559\u4e0e\u95ee\u9898\u6700\u76f8\u5173\u7684 4 \u4e2a\u6587\u672c\u5757\u9001\u5165\u751f\u6210\u73af\u8282，\u65e2\u63d0\u5347\u4e86\u56de\u7b54\u51c6\u786e\u7387，\u4e5f\u63a7\u5236\u4e86 Token \u6d88\u8017\u3002<\/p>\n
\u63d0\u793a\u8bcd\u662f\u8fde\u63a5\u68c0\u7d22\u7ed3\u679c\u4e0e\u6a21\u578b\u751f\u6210\u7684\u6865\u6881\u3002\u597d\u7684\u63d0\u793a\u8bcd\u5e94\u660e\u786e\u8981\u6c42\u6a21\u578b\u53ea\u4f9d\u636e\u7ed9\u5b9a\u4e0a\u4e0b\u6587\u4f5c\u7b54，\u5e76\u5728\u4fe1\u606f\u4e0d\u8db3\u65f6\u5982\u5b9e\u8bf4\u660e\u3002<\/p>\n
from langchain_core.prompts import ChatPromptTemplate
\nprompt = ChatPromptTemplate.from_messages([
\n("system", "\u4f60\u662f\u4e00\u4e2a\u4f01\u4e1a\u77e5\u8bc6\u5e93\u95ee\u7b54\u52a9\u624b\u3002\u8bf7\u4e25\u683c\u4f9d\u636e\u63d0\u4f9b\u7684\u4e0a\u4e0b\u6587\u56de\u7b54\u7528\u6237\u95ee\u9898\u3002"
\n"\u5982\u679c\u4e0a\u4e0b\u6587\u4e2d\u6ca1\u6709\u76f8\u5173\u4fe1\u606f，\u8bf7\u660e\u786e\u56de\u7b54'\u77e5\u8bc6\u5e93\u4e2d\u672a\u627e\u5230\u76f8\u5173\u5185\u5bb9'，\u4e0d\u8981\u7f16\u9020\u7b54\u6848\u3002"
\n"\u56de\u7b54\u672b\u5c3e\u8bf7\u6807\u6ce8\u4fe1\u606f\u6765\u6e90\u3002"),
\n("human", "\u4e0a\u4e0b\u6587：\\\\n{context}\\\\n\\\\n\u95ee\u9898：{question}")
\n])<\/p>\n
\u6700\u540e，\u628a\u68c0\u7d22\u5668\u3001\u63d0\u793a\u8bcd\u548c GPT-6 \u6a21\u578b\u4e32\u8054\u6210\u5b8c\u6574\u7684\u95ee\u7b54\u94fe\u8def\u3002<\/p>\n
from langchain_core.runnables import RunnablePassthrough
\nfrom langchain_core.output_parsers import StrOutputParser
\ndef format_docs(docs):
\nreturn "\\\\n\\\\n".join(f"\u3010\u6765\u6e90：{doc.metadata.get('source', '\u672a\u77e5')}\u3011\\\\n{doc.page_content}" for doc in docs)
\nrag_chain = (
\n{"context": compression_retriever | format_docs, "question": RunnablePassthrough()}
\n| prompt
\n| llm
\n| StrOutputParser()
\n)
\n\u6d4b\u8bd5\u95ee\u7b54
\nresult = rag_chain.invoke("\u516c\u53f8\u5e74\u5047\u653f\u7b56\u662f\u4ec0\u4e48？")
\nprint(result)<\/p>\n
\u7cfb\u7edf\u4e0a\u7ebf\u524d，\u5efa\u8bae\u51c6\u5907\u4e00\u7ec4\u8986\u76d6\u5e38\u89c1\u4e1a\u52a1\u95ee\u9898\u7684\u6d4b\u8bd5\u96c6，\u4ece\u53ec\u56de\u51c6\u786e\u7387\u3001\u56de\u7b54\u6b63\u786e\u7387\u3001\u6765\u6e90\u53ef\u8ffd\u6eaf\u6027\u4e09\u4e2a\u7ef4\u5ea6\u8bc4\u4f30\u6548\u679c\u3002\u5e38\u89c1\u95ee\u9898\u4e0e\u8c03\u4f18\u65b9\u5411\u5982\u4e0b：<\/p>\n
| \u56de\u7b54\u4e0e\u95ee\u9898\u65e0\u5173<\/td>\n | \u68c0\u7d22\u53ec\u56de\u4e0d\u51c6\u786e<\/td>\n | \u8c03\u6574\u5207\u5206\u7c92\u5ea6\u3001\u589e\u52a0\u91cd\u6392\u5e8f\u3001\u4f18\u5316\u5411\u91cf\u6a21\u578b<\/td>\n<\/tr>\n |
| \u56de\u7b54\u5185\u5bb9\u8fc7\u65f6<\/td>\n | \u7d22\u5f15\u672a\u53ca\u65f6\u66f4\u65b0<\/td>\n | \u5efa\u7acb\u6587\u6863\u589e\u91cf\u66f4\u65b0\u673a\u5236<\/td>\n<\/tr>\n |
| \u56de\u7b54\u7f3a\u4e4f\u4f9d\u636e<\/td>\n | \u4e0a\u4e0b\u6587\u62fc\u63a5\u4e0d\u5b8c\u6574<\/td>\n | \u589e\u5927 Top-K \u6570\u91cf\u3001\u4f18\u5316\u63d0\u793a\u8bcd\u7ea6\u675f<\/td>\n<\/tr>\n |
| \u54cd\u5e94\u5ef6\u8fdf\u8fc7\u9ad8<\/td>\n | \u68c0\u7d22\u94fe\u8def\u8fc7\u957f<\/td>\n | \u7cbe\u7b80\u53ec\u56de\u6570\u91cf\u3001\u5f15\u5165\u7f13\u5b58\u673a\u5236<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n \u6b64\u5916，\u5efa\u8bae\u5728\u7cfb\u7edf\u4e2d\u52a0\u5165\u65e5\u5fd7\u8bb0\u5f55，\u6c89\u6dc0\u7528\u6237\u771f\u5b9e\u95ee\u9898，\u6301\u7eed\u8fed\u4ee3\u68c0\u7d22\u7b56\u7565\u548c\u63d0\u793a\u8bcd\u6a21\u677f\u3002<\/p>\n \u4e5d\u3001\u603b\u7ed3<\/h3>\n\u672c\u6587\u4ece\u67b6\u6784\u8bbe\u8ba1\u3001\u73af\u5883\u642d\u5efa\u3001\u6587\u6863\u5207\u5206\u3001\u5411\u91cf\u5b58\u50a8\u3001\u68c0\u7d22\u4f18\u5316\u5230\u63d0\u793a\u8bcd\u7f16\u6392，\u5b8c\u6574\u6f14\u793a\u4e86\u57fa\u4e8e GPT-6 \u4e0e LangChain \u6784\u5efa\u4f01\u4e1a\u7ea7 RAG \u95ee\u7b54\u7cfb\u7edf\u7684\u5168\u8fc7\u7a0b\u3002\u6838\u5fc3\u8981\u70b9\u53ef\u4ee5\u6982\u62ec\u4e3a：\u5207\u5206\u8d28\u91cf\u51b3\u5b9a\u53ec\u56de\u4e0a\u9650，\u91cd\u6392\u5e8f\u51b3\u5b9a\u7cbe\u5ea6，\u63d0\u793a\u8bcd\u51b3\u5b9a\u751f\u6210\u8d28\u91cf\u3002<\/p>\n \u8bfb\u8005\u53ef\u4ee5\u5728\u6b64\u57fa\u7840\u4e0a，\u7ed3\u5408\u81ea\u8eab\u4e1a\u52a1\u6587\u6863\u66ff\u6362\u6570\u636e\u6e90，\u5e76\u9010\u6b65\u5f15\u5165\u6743\u9650\u63a7\u5236\u3001\u591a\u8f6e\u5bf9\u8bdd\u548c\u6d41\u5f0f\u8f93\u51fa\u7b49\u4f01\u4e1a\u7ea7\u80fd\u529b，\u6253\u9020\u771f\u6b63\u53ef\u843d\u5730\u7684\u667a\u80fd\u95ee\u7b54\u7cfb\u7edf\u3002<\/p>\n","protected":false},"excerpt":{"rendered":" \u4e00\u3001\u4e3a\u4ec0\u4e48\u9009\u62e9 RAG \u67b6\u6784GPT-6 \u867d\u7136\u62e5\u6709\u5f3a\u5927\u7684\u901a\u7528\u80fd\u529b，\u4f46\u4f01\u4e1a\u79c1\u6709\u77e5\u8bc6\u5e93\u4e2d\u7684\u4e1a\u52a1\u6570\u636e\u5f80\u5f80\u4e0d\u5728\u6a21\u578b\u8bad\u7ec3\u8303\u56f4\u5185\u3002RAG（\u68c0\u7d22\u589e\u5f3a\u751f\u6210）\u901a\u8fc7\u300c\u5148\u68c0\u7d22\u3001\u540e\u751f\u6210\u300d\u7684\u65b9\u5f0f，\u628a\u4f01\u4e1a\u6587\u6863\u3001\u6570\u636e\u5e93\u4e2d\u7684\u5185\u5bb9\u5b9e\u65f6\u6ce8\u5165\u63d0\u793a\u8bcd，\u8ba9\u6a21\u578b\u57fa\u4e8e\u771f\u5b9e\u4e1a\u52a1\u8d44\u6599\u4f5c\u7b54，\u65e2\u89e3\u51b3\u4e86\u77e5\u8bc6\u65f6\u6548\u6027\u95ee\u9898，\u4e5f\u964d\u4f4e\u4e86\u5e7b\u89c9\u98ce\u9669\u3002\u76f8\u6bd4\u76f4\u63a5\u5fae\u8c03，RAG \u5177\u5907\u4e09\u4e2a\u660e\u663e\u4f18\u52bf：\u4e00\u662f\u77e5\u8bc6\u66f4\u65b0\u6210\u672c\u4f4e，<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[9098,1646,997,604],"topic":[],"class_list":["post-101851","post","type-post","status-publish","format-standard","hentry","category-server","tag-it","tag-gpt","tag-langchain","tag-oracle"],"yoast_head":"\n |