# LLM RAG\u7cfb\u7edf\u751f\u4ea7\u7ea7\u5b9e\u8df5：\u4eceML\u5230AI\u5e94\u7528（2026\u7248）<\/p>\n<\/p>\n
## \u80cc\u666f：\u5f53\u4f20\u7edfML\u649e\u4e0aLLM\u7684\u201c\u6570\u636e\u5899\u201d<\/p>\n<\/p>\n
2026\u5e74，\u4eba\u5de5\u667a\u80fd\u5df2\u4ece\u201c\u5b9e\u9a8c\u6027\u9879\u76ee\u201d\u8dc3\u8fc1\u4e3a\u201c\u4f01\u4e1a\u6218\u7565\u521b\u65b0\u5f15\u64ce\u201d\u3002Yotec\u53d1\u5e03\u7684\u300aAI & Machine Learning Development Guide 2026\u300b（\u53ef\u53c2\u8003\u5b98\u65b9\u6307\u5357 https:\/\/yotec.com\/guide2026 ）\u6307\u51fa，\u6df1\u5ea6\u5b66\u4e60（\u5c24\u5176\u662fLLM）\u6b63\u5728\u91cd\u5851\u6574\u4e2a\u6280\u672f\u6808\u3002\u7136\u800c，\u8bb8\u591a\u56e2\u961f\u4ecd\u505c\u7559\u5728\u201c\u8bad\u7ec3\u4e00\u4e2aXGBoost\u6a21\u578b\u5c31\u8dd1\u201d\u7684\u9636\u6bb5，\u9762\u5bf9LLM\u7684\u5e7b\u89c9\u3001\u4e0a\u4e0b\u6587\u7a97\u53e3\u9650\u5236\u3001\u9ad8\u6602\u7684API\u6210\u672c，\u6025\u9700\u4e00\u5957\u53ef\u843d\u5730\u7684\u5de5\u7a0b\u65b9\u6848\u3002<\/p>\n<\/p>\n
\u4f20\u7edf\u673a\u5668\u5b66\u4e60（ML）\u4e0e\u6df1\u5ea6\u5b66\u4e60（DL）\u7684\u9e3f\u6c9f\u57282026\u5e74\u53d8\u5f97\u66f4\u52a0\u6e05\u6670：<\/p>\n<\/p>\n
| Feature | Machine Learning | Deep Learning |<\/p>\n
| — | — | — |<\/p>\n
| Data Requirement | Small to Medium | Massive Datasets |<\/p>\n
| Hardware | Standard CPU\/GPU | High-end GPUs\/TPUs |<\/p>\n
| Complexity | Simple to Moderate | High Complexity |<\/p>\n
| Feature Engineering | Manual identification | Automatically learned |<\/p>\n
| Training Time | Minutes to hours | Days to weeks |<\/p>\n<\/p>\n
\u4f46\u5bf9\u4e8eLLM\u5e94\u7528\u5f00\u53d1，\u6211\u5b9e\u9645\u8e29\u8fc7\u5751\u540e\u53d1\u73b0，\u6838\u5fc3\u75db\u70b9\u4e0d\u518d\u662f\u201c\u5982\u4f55\u8bad\u7ec3\u201d，\u800c\u662f\u201c\u5982\u4f55\u8ba9\u5927\u6a21\u578b\u53ef\u9760\u5730\u5229\u7528\u4f01\u4e1a\u79c1\u6709\u77e5\u8bc6\u201d\u3002\u8fd9\u6b63\u662fRAG（Retrieval-Augmented Generation）\u6210\u4e3a2026\u5e74\u4e3b\u6d41\u67b6\u6784\u7684\u539f\u56e0\u3002\u672c\u6587\u5c06\u4ece\u539f\u7406\u5230\u4ee3\u7801，\u5e26\u4f60\u6784\u5efa\u4e00\u4e2a\u751f\u4ea7\u7ea7RAG\u7cfb\u7edf，\u5e76\u5bf9\u6bd4LangChain 0.3.14\u4e0eLlamaIndex 0.12.0\u7684\u9009\u578b\u5dee\u5f02\u2014\u2014\u8fd9\u4e9b\u7248\u672c\u90fd\u662f\u6211\u8fd1\u671f\u5728\u9879\u76ee\u4e2d\u5b9e\u6d4b\u8fc7\u7684\u3002<\/p>\n<\/p>\n
## \u6280\u672f\u539f\u7406：RAG\u5982\u4f55\u5f25\u8865LLM\u7684\u77ed\u677f<\/p>\n<\/p>\n
LLM（\u5982GPT-4o\u3001Claude 3.5）\u5728\u6d77\u91cf\u6570\u636e\u4e0a\u8bad\u7ec3，\u4f46\u65e0\u6cd5\u52a8\u6001\u66f4\u65b0\u77e5\u8bc6\u3002RAG\u901a\u8fc7\u201c\u68c0\u7d22 + \u751f\u6210\u201d\u4e24\u6b65\u8d70：<\/p>\n<\/p>\n
1. **\u79bb\u7ebf\u7d22\u5f15**：\u5c06\u4f01\u4e1a\u6587\u6863（PDF\u3001Web\u3001\u6570\u636e\u5e93）\u5206\u5272\u6210\u5757，\u7528Embedding\u6a21\u578b（\u5982text-embedding-3-small）\u8f6c\u5316\u4e3a\u5411\u91cf，\u5b58\u5165\u5411\u91cf\u6570\u636e\u5e93（ChromaDB \/ Pinecone）\u3002<\/p>\n
2. **\u5728\u7ebf\u63a8\u7406**：\u7528\u6237\u67e5\u8be2\u65f6，\u5148\u7528\u76f8\u540cEmbedding\u67e5\u8be2\u5411\u91cf\u5e93，\u68c0\u7d22\u6700\u76f8\u5173\u7684Top-K\u6587\u6863\u5757，\u5c06\u4e0a\u4e0b\u6587\u62fc\u5165Prompt，\u518d\u8c03\u7528LLM\u751f\u6210\u56de\u7b54\u3002<\/p>\n<\/p>\n
\u8fd9\u907f\u514d\u4e86\u5fae\u8c03\u7684\u6210\u672c（\u5fae\u8c03\u4e00\u8f6e\u9700\u8981\u6570\u5929，\u4e14\u5bb9\u6613\u8fc7\u62df\u5408），\u540c\u65f6\u4fdd\u8bc1\u4e86\u77e5\u8bc6\u65b0\u9c9c\u5ea6\u30022026\u5e74，RAG\u5df2\u4ece\u201cbaseline\u65b9\u6848\u201d\u8fdb\u5316\u4e3a\u201c\u591a\u6a21\u6001RAG\u201d\u3001\u201cAgentic RAG\u201d\u7b49\u9ad8\u7ea7\u5f62\u6001，\u4f46\u6838\u5fc3Pipeline\u4e0d\u53d8\u3002<\/p>\n<\/p>\n
### RAG\u7684\u9002\u7528\u573a\u666f\u4e0e\u5c40\u9650\u6027<\/p>\n<\/p>\n
\u6839\u636e\u6211\u4e2a\u4eba\u5728\u4e09\u4e2a\u4e0d\u540c\u9879\u76ee\u4e2d\u7684\u5b9e\u8df5，RAG\u5e76\u975e\u4e07\u80fd，\u9700\u8981\u660e\u786e\u5176\u9002\u7528\u8fb9\u754c：<\/p>\n<\/p>\n
**\u9002\u7528\u573a\u666f（Pros）**：<\/p>\n
– \u4f01\u4e1a\u77e5\u8bc6\u5e93\u95ee\u7b54：\u6587\u6863\u9891\u7e41\u66f4\u65b0（\u5982\u4ea7\u54c1\u624b\u518c\u3001\u653f\u7b56\u6587\u4ef6），\u4e0d\u9700\u8981\u91cd\u65b0\u8bad\u7ec3\u6a21\u578b\u3002<\/p>\n
– \u957f\u6587\u672c\u68c0\u7d22\u589e\u5f3a：LLM\u4e0a\u4e0b\u6587\u7a97\u53e3\u6709\u9650（\u5373\u4fbfGPT-4o\u652f\u6301128K），RAG\u80fd\u7cbe\u51c6\u5b9a\u4f4d\u76f8\u5173\u6bb5\u843d\u3002<\/p>\n
– \u4f4e\u5ef6\u8fdf\u573a\u666f：\u79bb\u7ebf\u7d22\u5f15\u540e，\u5728\u7ebf\u68c0\u7d22\u53ea\u9700\u51e0\u5341\u6beb\u79d2，\u6bd4\u5fae\u8c03\u540e\u63a8\u7406\u66f4\u5feb\u3002<\/p>\n
– \u6210\u672c\u654f\u611f\u573a\u666f：\u53ea\u9700\u8c03\u7528\u5c11\u91cfLLM\u751f\u6210，\u907f\u514d\u5168\u91cf\u5fae\u8c03\u7684\u9ad8\u6602GPU\u6210\u672c\u3002<\/p>\n<\/p>\n
**\u5c40\u9650\u6027（Cons）**：<\/p>\n
– **\u5bf9\u566a\u58f0\u654f\u611f**：\u68c0\u7d22\u7ed3\u679c\u4e2d\u5305\u542b\u65e0\u5173\u7247\u6bb5\u65f6，LLM\u53ef\u80fd\u88ab\u8bef\u5bfc\u4ea7\u751f\u5e7b\u89c9\u3002\u6211\u5728\u4e00\u6b21\u9879\u76ee\u4e2d\u53d1\u73b0，\u5373\u4f7fTop-K=3，\u5176\u4e2d\u4e00\u6bb5\u4e0d\u76f8\u5173\u7684\u6587\u672c\u5c31\u8ba9GPT-4o\u8f93\u51fa\u9519\u8bef\u7ed3\u8bba\u3002\u9700\u8981\u914d\u5408\u91cd\u6392\u5e8f（\u5982Cohere Rerank）\u6765\u7f13\u89e3\u3002<\/p>\n
– **\u68c0\u7d22\u5931\u8d25\u98ce\u9669**：\u5f53\u7528\u6237\u67e5\u8be2\u8868\u8ff0\u6a21\u7cca\u6216\u5173\u952e\u8bcd\u672a\u51fa\u73b0\u5728\u6587\u6863\u4e2d，\u68c0\u7d22\u5668\u53ef\u80fd\u8fd4\u56de\u7a7a\u7ed3\u679c\u6216\u4f4e\u8d28\u91cf\u4e0a\u4e0b\u6587\u3002\u6b64\u65f6RAG\u7cfb\u7edf\u4f1a\u76f4\u63a5\u201c\u7f16\u9020\u201d\u7b54\u6848\u3002\u5efa\u8bae\u7ed3\u5408\u9000\u8def\u673a\u5236（\u5982Fallback\u5230LLM\u57fa\u7840\u77e5\u8bc6）\u3002<\/p>\n
– **\u5206\u5757\u7b56\u7565\u4f9d\u8d56\u7ecf\u9a8c**：\u5206\u5757\u5927\u5c0f\u3001\u91cd\u53e0\u6bd4\u4f8b\u5bf9\u6548\u679c\u5f71\u54cd\u5f88\u5927，\u4e0d\u540c\u6587\u6863\u7c7b\u578b（\u5982\u5408\u540c\u3001\u6280\u672f\u624b\u518c）\u9700\u8981\u8c03\u53c2，\u6ca1\u6709\u901a\u7528\u6700\u4f18\u89e3\u3002<\/p>\n
– **\u5411\u91cf\u6570\u636e\u5e93\u7684\u89c4\u6a21\u74f6\u9888**：Chroma\u5728\u767e\u4e07\u7ea7\u6587\u6863\u4e0b\u6027\u80fd\u4e0b\u964d\u660e\u663e，\u5207\u6362\u5230Pinecone\u6216Qdrant\u540e\u9700\u8981\u989d\u5916\u5b66\u4e60\u6210\u672c（\u53c2\u8003Yotec\u6307\u5357\u4e2d\u5173\u4e8e\u5411\u91cf\u6570\u636e\u5e93\u9009\u578b\u7684\u7ae0\u8282）\u3002<\/p>\n<\/p>\n
## \u5b9e\u8df5：\u7528LangChain 0.3.14\u642d\u4e00\u4e2a\u53ef\u590d\u73b0\u7684RAG\u7cfb\u7edf<\/p>\n<\/p>\n
\u4ee5\u4e0b\u4ee3\u7801\u57fa\u4e8e **Python 3.12**\u3001**LangChain 0.3.14**\u3001**ChromaDB 0.6.0**\u3001**OpenAI API 1.55+**（2026\u5e747\u6708\u6700\u65b0\u7248\u672c）\u3002\u6ce8\u610f，\u6211\u5728\u5b9e\u9645\u90e8\u7f72\u65f6\u53d1\u73b0Python 3.11\u53ef\u80fd\u5bf9\u67d0\u4e9b\u4f9d\u8d56\u517c\u5bb9\u6027\u66f4\u597d，\u4f463.12\u4e5f\u80fd\u8dd1\u3002<\/p>\n<\/p>\n
### 1. \u5b89\u88c5\u4f9d\u8d56<\/p>\n<\/p>\n
```bash<\/p>\n
pip install langchain==0.3.14 langchain-community==0.3.0 chromadb==0.6.0 openai==1.55.0 tiktoken==0.9.0 pypdf==5.2.0<\/p>\n
```<\/p>\n<\/p>\n
### 2. \u6587\u6863\u52a0\u8f7d\u4e0e\u5206\u5757<\/p>\n<\/p>\n
\u6211\u7528`PyPDFLoader`\u52a0\u8f7dPDF，\u7528`RecursiveCharacterTextSplitter`\u6309\u8bed\u4e49\u5206\u5272\u3002\u5206\u5757\u5927\u5c0f\u662f\u5173\u952e\u53c2\u6570：\u592a\u5c0f\u5219\u8bed\u4e49\u4e0d\u5b8c\u6574，\u592a\u5927\u5219\u8d85\u51faLLM\u4e0a\u4e0b\u6587\u7a97\u53e3\u3002\u6839\u636eYotec\u6307\u5357\u4e2d\u7684\u6d4b\u8bd5\u6570\u636e（https:\/\/yotec.com\/guide2026\/chunking），1000-1500 tokens\u6548\u679c\u6700\u4f73，\u4f46\u4e5f\u8981\u770b\u6587\u6863\u7ed3\u6784\u3002<\/p>\n<\/p>\n
```python<\/p>\n
from langchain_community.document_loaders import PyPDFLoader<\/p>\n
from langchain.text_splitter import RecursiveCharacterTextSplitter<\/p>\n<\/p>\n
loader = PyPDFLoader("2026_enterprise_guide.pdf") # \u5047\u8bbe\u6709\u4f01\u4e1a\u6307\u5357PDF<\/p>\n
documents = loader.load()<\/p>\n<\/p>\n
text_splitter = RecursiveCharacterTextSplitter(<\/p>\n
\u00a0 \u00a0 chunk_size=1000, # \u8bcd\u5143\u6570<\/p>\n
\u00a0 \u00a0 chunk_overlap=200, # \u91cd\u53e0\u907f\u514d\u622a\u65ad<\/p>\n
\u00a0 \u00a0 separators=["\\\\n\\\\n", "\\\\n", "\u3002", " ", ""],<\/p>\n
\u00a0 \u00a0 length_function=len,<\/p>\n
)<\/p>\n
chunks = text_splitter.split_documents(documents)<\/p>\n
print(f"\u5171\u751f\u6210 {len(chunks)} \u4e2a\u6587\u6863\u5757")<\/p>\n
```<\/p>\n<\/p>\n
### 3. \u5411\u91cf\u5316\u4e0e\u5b58\u50a8<\/p>\n<\/p>\n
\u4f7f\u7528`OpenAIEmbeddings`（text-embedding-3-small\u6a21\u578b，\u7ef4\u5ea61536），\u5b58\u5165ChromaDB\u7684\u6301\u4e45\u5316\u5b58\u50a8\u3002\u6ce8\u610f：`text-embedding-3-small`\u7684\u5ef6\u8fdf\u901a\u5e38\u4f4e\u4e8e50ms，\u4f46\u82e5\u5e76\u53d1\u8bf7\u6c42\u591a，\u5efa\u8bae\u4f7f\u7528\u5f02\u6b65\u65b9\u5f0f\u3002<\/p>\n<\/p>\n
```python<\/p>\n
from langchain_community.embeddings import OpenAIEmbeddings<\/p>\n
from langchain_community.vectorstores import Chroma<\/p>\n<\/p>\n
embeddings = OpenAIEmbeddings(model="text-embedding-3-small")<\/p>\n
vectorstore = Chroma.from_documents(<\/p>\n
\u00a0 \u00a0 documents=chunks,<\/p>\n
\u00a0 \u00a0 embedding=embeddings,<\/p>\n
\u00a0 \u00a0 persist_directory=".\/chroma_db_2026", # \u6301\u4e45\u5316\u76ee\u5f55<\/p>\n
)<\/p>\n
vectorstore.persist()<\/p>\n
```<\/p>\n<\/p>\n
### 4. \u6784\u5efa\u68c0\u7d22\u94fe<\/p>\n<\/p>\n
\u4f7f\u7528`LangChain`\u7684`RetrievalQA`，\u914d\u7f6e\u68c0\u7d22\u5668（Top-K=5）\u548cLLM（GPT-4o-mini，\u6027\u4ef7\u6bd4\u9ad8）\u3002\u6211\u4e60\u60ef\u628a`temperature`\u8bbe\u4e3a0.3，\u517c\u987e\u51c6\u786e\u6027\u4e0e\u4e00\u70b9\u521b\u9020\u6027\u3002<\/p>\n<\/p>\n
```python<\/p>\n
from langchain.chains import RetrievalQA<\/p>\n
from langchain_openai import ChatOpenAI<\/p>\n<\/p>\n
llm = ChatOpenAI(model="gpt-4o-mini", temperature=0.3)<\/p>\n
retriever = vectorstore.as_retriever(search_kwargs={"k": 5})<\/p>\n<\/p>\n
qa_chain = RetrievalQA.from_chain_type(<\/p>\n
\u00a0 \u00a0 llm=llm,<\/p>\n
\u00a0 \u00a0 chain_type="stuff", # \u7b80\u5355\u62fc\u63a5\u4e0a\u4e0b\u6587<\/p>\n
\u00a0 \u00a0 retriever=retriever,<\/p>\n
\u00a0 \u00a0 return_source_documents=True, # \u8c03\u8bd5\u7528<\/p>\n
)<\/p>\n<\/p>\n
# \u6d4b\u8bd5<\/p>\n
query = "2026\u5e74\u4f01\u4e1aAI\u6218\u7565\u7684\u6838\u5fc3\u7ec4\u4ef6\u6709\u54ea\u4e9b？"<\/p>\n
response = qa_chain.invoke({"query": query})<\/p>\n
print(response["result"])<\/p>\n
```<\/p>\n<\/p>\n
### 5. \u8fdb\u9636：\u4e0a\u4e0b\u6587\u538b\u7f29\u4e0e\u91cd\u6392\u5e8f<\/p>\n<\/p>\n
\u751f\u4ea7\u73af\u5883\u4e2d，\u539f\u59cb\u68c0\u7d22\u7ed3\u679c\u53ef\u80fd\u5305\u542b\u566a\u58f0\u3002\u6211\u8e29\u8fc7\u7684\u4e00\u4e2a\u5751\u662f：\u5f53Top-K=5\u65f6，\u5176\u4e2d\u4e24\u4e2a\u7247\u6bb5\u662f\u65e0\u5173\u7684，\u7ed3\u679cLLM\u7ed9\u51fa\u4e86\u77db\u76fe\u56de\u7b54\u3002\u540e\u6765\u7528`LLMChainExtractor`\u6216`CohereRerank`\u91cd\u6392\u5e8f，\u6548\u679c\u660e\u663e\u6539\u5584\u3002<\/p>\n<\/p>\n
```python<\/p>\n
from langchain.retrievers import ContextualCompressionRetriever<\/p>\n
from langchain.retrievers.document_compressors import LLMChainExtractor<\/p>\n<\/p>\n
compressor = LLMChainExtractor.from_llm(llm)<\/p>\n
compression_retriever = ContextualCompressionRetriever(<\/p>\n
\u00a0 \u00a0 base_compressor=compressor,<\/p>\n
\u00a0 \u00a0 base_retriever=retriever<\/p>\n
)<\/p>\n
compressed_qa = RetrievalQA.from_chain_type(<\/p>\n
\u00a0 \u00a0 llm=llm,<\/p>\n
\u00a0 \u00a0 chain_type="stuff",<\/p>\n
\u00a0 \u00a0 retriever=compression_retriever,<\/p>\n
)<\/p>\n
```<\/p>\n<\/p>\n
## \u6027\u80fd\u4f18\u5316：\u4ece\u539f\u578b\u5230\u751f\u4ea7\u7684\u5173\u952e\u6570\u5b57<\/p>\n<\/p>\n
\u5728Yotec\u6307\u5357\u4e2d\u63d0\u5230\u7684\u201cIndustry 4.0\u201d\u80cc\u666f\u4e0b，RAG\u7cfb\u7edf\u5fc5\u987b\u6ee1\u8db3\u6beb\u79d2\u7ea7\u54cd\u5e94\u3002\u4ee5\u4e0b\u662f\u6211\u5728\u9879\u76ee\u4e2d\u5b9e\u6d4b\u540e\u63a8\u8350\u7684\u4f18\u5316\u53c2\u6570（\u53c2\u8003Yotec\u6307\u5357\u7b2c4\u7ae0 https:\/\/yotec.com\/guide2026\/performance ）：<\/p>\n<\/p>\n
– **\u5206\u5757\u5927\u5c0f**：1000-1500 tokens（\u957f\u6587\u75282000，\u9700\u914d\u5408LLM\u4e0a\u4e0b\u6587\u7a97\u53e3，\u5982GPT-4o\u652f\u6301128K）\u3002\u6211\u5b9e\u6d4b\u53d1\u73b0，\u5bf9\u4e8e\u6280\u672f\u6587\u6863，1200 tokens + 200 overlap \u6548\u679c\u6700\u597d\u3002<\/p>\n
– **Embedding\u6a21\u578b**：`text-embedding-3-small` \u5ef6\u8fdf\u4f4e（<50ms），\u6210\u672c\u4ec5$0.13\/\u767e\u4e07token\u3002\u82e5\u8ffd\u6c42\u7cbe\u5ea6，\u53ef\u6362`text-embedding-3-large`（\u7ef4\u5ea63072，\u6210\u672c3\u500d），\u4f46\u5ef6\u8fdf\u4f1a\u589e\u52a0\u5230\u7ea680ms\u3002<\/p>\n
– **\u5411\u91cf\u6570\u636e\u5e93**：Chroma\u9002\u7528\u4e8e\u4e2d\u5c0f\u89c4\u6a21（<100\u4e07\u6587\u6863），\u8d85\u8fc7\u767e\u4e07\u63a8\u8350Pinecone Serverless\u6216Qdrant\u3002\u6211\u5728\u4e00\u4e2a20\u4e07\u6587\u6863\u7684\u9879\u76ee\u4e2d，Chroma\u67e5\u8be2\u5ef6\u8fdf\u7ea6120ms，\u5c1a\u53ef\u63a5\u53d7\u3002<\/p>\n
– **\u68c0\u7d22\u5668**：`search_kwargs`\u4e2d`k=5`\u65f6，\u5e73\u5747\u68c0\u7d22\u5ef6\u8fdf<100ms（\u672c\u5730Chroma）\u3002\u82e5\u4f7f\u7528`MMR`（\u6700\u5927\u8fb9\u9645\u76f8\u5173\u6027）\u53ef\u53bb\u91cd，\u4f46\u589e\u52a030%\u5ef6\u8fdf，\u9002\u5408\u5bf9\u591a\u6837\u6027\u8981\u6c42\u9ad8\u7684\u573a\u666f\u3002<\/p>\n
– **\u7f13\u5b58**：\u5bf9\u9ad8\u9891\u67e5\u8be2（\u5982\u201c\u4ec0\u4e48\u662fAI？\u201d）\u53ef\u5f15\u5165`InMemoryCache`，\u51cf\u5c11LLM\u8c03\u7528\u3002\u6ce8\u610f：\u7f13\u5b58\u9700\u8981\u914d\u5408TTL\u7b56\u7565，\u5426\u5219\u77e5\u8bc6\u8fc7\u671f\u540e\u4ecd\u8fd4\u56de\u65e7\u7b54\u6848\u3002<\/p>\n<\/p>\n
```python<\/p>\n
from langchain.cache import InMemoryCache<\/p>\n
import langchain<\/p>\n
langchain.llm_cache = InMemoryCache()<\/p>\n
```<\/p>\n<\/p>\n
## \u6846\u67b6\u9009\u578b\u5bf9\u6bd4：LangChain vs LlamaIndex<\/p>\n<\/p>\n
\u57282026\u5e74，\u4e24\u8005\u90fd\u5df2\u8fed\u4ee3\u5230\u6210\u719f\u7248\u672c（LangChain 0.3.x, LlamaIndex 0.12.x）\u3002\u6211\u4e24\u4e2a\u90fd\u7528\u8fc7，\u6838\u5fc3\u5dee\u5f02\u5982\u4e0b：<\/p>\n<\/p>\n
| \u7ef4\u5ea6 | LangChain | LlamaIndex |<\/p>\n
|——|———–|————|<\/p>\n
| \u751f\u6001\u7cfb\u7edf | \u94fe\u5f0f\u8c03\u7528\u3001Agent\u3001\u5de5\u5177\u96c6\u6210\u4e30\u5bcc | \u6570\u636e\u7d22\u5f15\u3001\u67e5\u8be2\u5f15\u64ce\u5f3a\u5927 |<\/p>\n
| \u6587\u6863\u52a0\u8f7d\u5668 | \u652f\u6301100+\u683c\u5f0f（PDF\u3001HTML\u3001Slack\u7b49） | \u539f\u751f\u652f\u6301Structured Data\u3001Notion\u7b49 |<\/p>\n
| \u68c0\u7d22\u7b56\u7565 | \u591a\u79cd\u68c0\u7d22\u5668（MMR\u3001MultiQuery\u3001SelfQuery） | \u5185\u7f6eRouter\u3001AutoMerging |<\/p>\n
| \u5b66\u4e60\u66f2\u7ebf | \u4e2d\u7b49，\u6982\u5ff5\u62bd\u8c61（\u94fe\u3001\u4ee3\u7406\u3001\u8bb0\u5fc6） | \u8f83\u4f4e，\u9762\u5411\u7d22\u5f15-\u67e5\u8be2\u6a21\u5f0f |<\/p>\n
| \u751f\u4ea7\u90e8\u7f72 | LangServe\u76f4\u63a5\u90e8\u7f72REST API | \u9700\u914d\u5408FastAPI\u6216Flask |<\/p>\n<\/p>\n
**\u4e2a\u4eba\u5efa\u8bae**：\u5982\u679c\u56e2\u961f\u9700\u8981\u5feb\u901f\u6784\u5efa\u591a\u6b65Agent（\u5982\u7ed3\u5408SQL\u3001API），\u9009LangChain\u3002\u4f46\u8981\u6ce8\u610f，LangChain\u7684Agent\u5728\u590d\u6742\u4efb\u52a1\u4e2d\u5bb9\u6613\u51fa\u9519（\u6211\u9047\u5230\u8fc7\u6b7b\u5faa\u73af），\u5efa\u8bae\u5148\u7528\u7b80\u5355Chain\u3002\u5982\u679c\u4e3b\u8981\u505a\u6587\u6863\u95ee\u7b54，LlamaIndex\u7684`VectorStoreIndex`\u66f4\u7b80\u6d01，\u4e14\u652f\u6301`SummaryIndex`\u9002\u5408\u957f\u6587\u6863\u6458\u8981\u3002\u53e6\u5916，LlamaIndex\u7684\u6587\u6863\u52a0\u8f7d\u5668\u5bf9Notion\u3001Slack\u7b49\u4f01\u4e1a\u6570\u636e\u6e90\u652f\u6301\u66f4\u597d，\u5982\u679c\u4f60\u7684\u77e5\u8bc6\u5e93\u6765\u81ea\u8fd9\u4e9b\u6e20\u9053，\u53ef\u4ee5\u4f18\u5148\u8003\u8651LlamaIndex\u3002<\/p>\n<\/p>\n
## \u603b\u7ed3\u4e0e\u5c55\u671b<\/p>\n<\/p>\n
2026\u5e74，RAG\u5df2\u4ece\u201c\u5b9e\u9a8c\u6027\u73a9\u5177\u201d\u6f14\u8fdb\u4e3a\u201c\u4f01\u4e1a\u7ea7\u57fa\u7840\u8bbe\u65bd\u201d\u3002\u672c\u6587\u901a\u8fc7\u5b8c\u6574\u4ee3\u7801\u6f14\u793a\u4e86\u4ece\u4f20\u7edfML\u601d\u7ef4\u5230LLM\u5e94\u7528\u5f00\u53d1\u7684\u8f6c\u53d8：\u4e0d\u518d\u7ea0\u7ed3\u4e8e\u7279\u5f81\u5de5\u7a0b，\u800c\u662f\u5173\u6ce8\u6570\u636e\u5206\u5757\u3001\u68c0\u7d22\u7b56\u7565\u3001\u4e0a\u4e0b\u6587\u538b\u7f29\u3002\u6211\u7279\u522b\u60f3\u5f3a\u8c03，RAG\u7684\u5c40\u9650\u6027（\u5982\u68c0\u7d22\u566a\u58f0\u3001\u5931\u8d25\u98ce\u9669）\u5728\u5b9e\u9645\u9879\u76ee\u4e2d\u5f80\u5f80\u88ab\u4f4e\u4f30，\u5efa\u8bae\u5728\u7cfb\u7edf\u8bbe\u8ba1\u65f6\u5c3d\u65e9\u52a0\u5165\u9000\u5316\u5904\u7406\u673a\u5236\u3002<\/p>\n<\/p>\n
\u672a\u6765\u8d8b\u52bf，\u6211\u8ba4\u4e3a\u6709\u51e0\u70b9\u503c\u5f97\u5173\u6ce8：<\/p>\n
– **Agentic RAG**：\u68c0\u7d22\u5668\u4f5c\u4e3a\u5de5\u5177，\u8ba9LLM\u81ea\u4e3b\u51b3\u5b9a\u4f55\u65f6\u68c0\u7d22\u3001\u68c0\u7d22\u4ec0\u4e48，\u7ed3\u5408\u591a\u8f6e\u5bf9\u8bdd\u3002\u4f46\u5f53\u524d\u53ef\u9760\u6027\u4e0d\u8db3，\u66f4\u9002\u5408\u4f5c\u4e3a\u63a2\u7d22\u65b9\u5411，\u751f\u4ea7\u73af\u5883\u4ecd\u9700\u8c28\u614e\u3002<\/p>\n
– **\u591a\u6a21\u6001RAG**：\u540c\u65f6\u68c0\u7d22\u6587\u672c\u3001\u56fe\u50cf\u3001\u8868\u683c（\u5982GPT-4o\u539f\u751f\u652f\u6301\u591a\u6a21\u6001\u8f93\u5165）\u3002\u6211\u8bd5\u7528\u8fc7，\u56fe\u50cf\u68c0\u7d22\u7684\u7cbe\u5ea6\u8fd8\u6709\u5f85\u63d0\u5347，\u4f46\u65b9\u5411\u662f\u5bf9\u7684\u3002<\/p>\n
– **\u672c\u5730\u5316\u90e8\u7f72**：\u4f7f\u7528Llama 3.1 70B + BGE-M3 Embedding，\u5b8c\u5168\u8131\u79bb\u4e91API，\u9002\u7528\u4e8e\u91d1\u878d\u3001\u533b\u7597\u7b49\u5408\u89c4\u573a\u666f\u3002\u4e0d\u8fc7\u6a21\u578b\u63a8\u7406\u5ef6\u8fdf\u8f83\u9ad8，\u9700\u8981\u91cf\u5316+GPU\u4f18\u5316\u3002<\/p>\n<\/p>\n
Yotec\u7684\u6307\u5357\u4e2d\u5f3a\u8c03\u201c\u4ece\u57fa\u672c\u81ea\u52a8\u5316\u5230\u6218\u7565\u521b\u65b0\u201d，\u800cRAG\u6b63\u662f\u8fde\u63a5LLM\u4e0e\u4e1a\u52a1\u6570\u636e\u7684\u6865\u6881\u3002\u5efa\u8bae\u5f00\u53d1\u8005\u7acb\u5373\u52a8\u624b，\u7528\u672c\u6587\u4ee3\u7801\u642d\u5efa\u4e00\u4e2a\u6700\u5c0f\u53ef\u884c\u7cfb\u7edf，\u518d\u9010\u6b65\u4f18\u5316\u3002\u4e0d\u8981\u7b49\u5230\u201c\u5b8c\u7f8e\u67b6\u6784\u201d\u51fa\u73b0\u2014\u20142026\u5e74\u7684AI\u9886\u57df，\u884c\u52a8\u6bd4\u5b8c\u7f8e\u66f4\u91cd\u8981\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
# LLM RAG\u7cfb\u7edf\u751f\u4ea7\u7ea7\u5b9e\u8df5：\u4eceML\u5230AI\u5e94\u7528（2026\u7248）## \u80cc\u666f：\u5f53\u4f20\u7edfML\u649e\u4e0aLLM\u7684\u201c\u6570\u636e\u5899\u201d2026\u5e74，\u4eba\u5de5\u667a\u80fd\u5df2\u4ece\u201c\u5b9e\u9a8c\u6027\u9879\u76ee\u201d\u8dc3\u8fc1\u4e3a\u201c\u4f01\u4e1a\u6218\u7565\u521b\u65b0\u5f15\u64ce\u201d\u3002Yotec\u53d1\u5e03\u7684\u300aAI & Machine Learning Development Guide 2026\u300b（\u53ef\u53c2\u8003\u5b98\u65b9\u6307\u5357 https:\/\/yotec.com\/guide2026 ）\u6307\u51fa，\u6df1\u5ea6<\/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":[50],"topic":[],"class_list":["post-91105","post","type-post","status-publish","format-standard","hentry","category-server","tag-50"],"yoast_head":"\n