上线一周就翻车？RAG生产部署避坑指南
这是RAG系列的最后一篇。前面6篇我们从原理讲到高级架构，从切分讲到评估，你的RAG应该已经能在本地跑出不错的效果了。但"本地跑得好"跟"线上稳定运行"之间隔着十万八千里。
我见过太多RAG项目，演示的时候很完美，上线一周就翻车：文档更新了但向量库没更新、检索延迟突然从200ms飙升到5秒、Embedding API限流导致整个服务不可用、新增一个客户的数据后检索结果混了……
今天这篇我把自己在生产环境踩过的坑整理成一份部署指南，照着做你的RAG能少走半年弯路。
架构设计：RAG服务的标准分层
生产级RAG不是"一个Python脚本跑个Flask"，而是分层架构：
<span leaf="">┌──────────────────────────────────┐</span><br><span leaf="">│ 接入层 │ API网关、认证、限流</span><br><span leaf="">├──────────────────────────────────┤</span><br><span leaf="">│ 编排层 │ 查询路由、对话管理</span><br><span leaf="">├──────────────────────────────────┤</span><br><span leaf="">│ 检索层 │ 混合检索、Rerank</span><br><span leaf="">├──────────────────────────────────┤</span><br><span leaf="">│ 生成层 │ LLM调用、流式输出</span><br><span leaf="">├──────────────────────────────────┤</span><br><span leaf="">│ 存储层 │ 向量库、文档库、缓存</span><br><span leaf="">├──────────────────────────────────┤</span><br><span leaf="">│ 数据管道 │ 文档入库、增量更新</span><br><span leaf="">└──────────────────────────────────┘</span><br>
每一层都可以独立扩展和替换。别把所有逻辑写在一个文件里——后面改一个地方牵一发动全身。
Java类比：这就像你搭一个Spring Boot微服务架构——API网关（接入层）→业务Service（编排层）→数据访问层（检索层）→外部服务调用（生成层）→数据库（存储层）。分层解耦是工程基本功，RAG也一样。
增量索引：文档更新不用重建
生产环境的文档是动态变化的——每天都有新文档入库、旧文档更新、错误文档删除。你不能每次变更都重建整个向量库。
增量更新流程
<span><span leaf="">import</span></span><span leaf=""> hashlib</span><br><span><span leaf="">from</span></span><span leaf=""> datetime </span><span><span leaf="">import</span></span><span leaf=""> datetime</span><br><br><span><span><span leaf="">class</span></span><span leaf=""> </span><span><span leaf="">DocumentManager</span></span><span leaf="">:</span></span><br><span leaf=""> </span><span><span leaf="">"""文档增量管理器"""</span></span><br><span leaf=""> </span><br><span leaf=""> </span><span><span><span leaf="">def</span></span><span leaf=""> </span><span><span leaf="">__init__</span></span><span><span leaf="">(self, vectorstore, doc_registry)</span></span><span leaf="">:</span></span><br><span leaf=""> self.vectorstore = vectorstore</span><br><span leaf=""> self.registry = doc_registry </span><span><span leaf=""># 文档注册表（数据库）</span></span><br><span leaf=""> </span><br><span leaf=""> </span><span><span><span leaf="">def</span></span><span leaf=""> </span><span><span leaf="">_get_doc_hash</span></span><span><span leaf="">(self, content: str)</span></span><span leaf=""> -> str:</span></span><br><span leaf=""> </span><span><span leaf="">"""计算文档内容哈希"""</span></span><br><span leaf=""> </span><span><span leaf="">return</span></span><span leaf=""> hashlib.md5(content.encode()).hexdigest()</span><br><span leaf=""> </span><br><span leaf=""> </span><span><span><span leaf="">def</span></span><span leaf=""> </span><span><span leaf="">add_document</span></span><span><span leaf="">(self, doc_path: str)</span></span><span leaf="">:</span></span><br><span leaf=""> </span><span><span leaf="">"""新增文档"""</span></span><br><span leaf=""> content = self._load_document(doc_path)</span><br><span leaf=""> doc_hash = self._get_doc_hash(content)</span><br><span leaf=""> </span><br><span leaf=""> </span><span><span leaf=""># 检查是否已存在</span></span><br><span leaf=""> existing = self.registry.get(doc_path)</span><br><span leaf=""> </span><span><span leaf="">if</span></span><span leaf=""> existing </span><span><span leaf="">and</span></span><span leaf=""> existing[</span><span><span leaf="">"hash"</span></span><span leaf="">] == doc_hash:</span><br><span leaf=""> print(</span><span><span leaf="">f"文档</span><span><span leaf="">{doc_path}</span></span><span leaf="">未变化，跳过"</span></span><span leaf="">)</span><br><span leaf=""> </span><span><span leaf="">return</span></span><br><span leaf=""> </span><br><span leaf=""> </span><span><span leaf=""># 切分 + 向量化 + 入库</span></span><br><span leaf=""> chunks = self.splitter.split_documents(content)</span><br><span leaf=""> ids = self.vectorstore.add_documents(chunks)</span><br><span leaf=""> </span><br><span leaf=""> </span><span><span leaf=""># 更新注册表</span></span><br><span leaf=""> self.registry.set(doc_path, {</span><br><span leaf=""> </span><span><span leaf="">"hash"</span></span><span leaf="">: doc_hash,</span><br><span leaf=""> </span><span><span leaf="">"chunk_ids"</span></span><span leaf="">: ids,</span><br><span leaf=""> </span><span><span leaf="">"updated_at"</span></span><span leaf="">: datetime.now().isoformat(),</span><br><span leaf=""> })</span><br><span leaf=""> </span><br><span leaf=""> </span><span><span><span leaf="">def</span></span><span leaf=""> </span><span><span leaf="">update_document</span></span><span><span leaf="">(self, doc_path: str)</span></span><span leaf="">:</span></span><br><span leaf=""> </span><span><span leaf="">"""更新文档：删旧的→加新的"""</span></span><br><span leaf=""> existing = self.registry.get(doc_path)</span><br><span leaf=""> </span><span><span leaf="">if</span></span><span leaf=""> existing:</span><br><span leaf=""> </span><span><span leaf=""># 删除旧chunks</span></span><br><span leaf=""> </span><span><span leaf="">for</span></span><span leaf=""> chunk_id </span><span><span leaf="">in</span></span><span leaf=""> existing[</span><span><span leaf="">"chunk_ids"</span></span><span leaf="">]:</span><br><span leaf=""> self.vectorstore.delete([chunk_id])</span><br><span leaf=""> </span><br><span leaf=""> </span><span><span leaf=""># 重新入库</span></span><br><span leaf=""> self.add_document(doc_path)</span><br><span leaf=""> </span><br><span leaf=""> </span><span><span><span leaf="">def</span></span><span leaf=""> </span><span><span leaf="">delete_document</span></span><span><span leaf="">(self, doc_path: str)</span></span><span leaf="">:</span></span><br><span leaf=""> </span><span><span leaf="">"""删除文档"""</span></span><br><span leaf=""> existing = self.registry.get(doc_path)</span><br><span leaf=""> </span><span><span leaf="">if</span></span><span leaf=""> existing:</span><br><span leaf=""> </span><span><span leaf="">for</span></span><span leaf=""> chunk_id </span><span><span leaf="">in</span></span><span leaf=""> existing[</span><span><span leaf="">"chunk_ids"</span></span><span leaf="">]:</span><br><span leaf=""> self.vectorstore.delete([chunk_id])</span><br><span leaf=""> self.registry.delete(doc_path)</span><br>
定时同步任务
<span><span leaf="">from</span></span><span leaf=""> apscheduler.schedulers.background </span><span><span leaf="">import</span></span><span leaf=""> BackgroundScheduler</span><br><br><span leaf="">scheduler = BackgroundScheduler()</span><br><br><span><span leaf=""># 每小时扫描一次文档目录</span></span><br><span leaf="">scheduler.add_job(</span><br><span leaf=""> scan_and_sync_documents,</span><br><span leaf=""> </span><span><span leaf="">"interval"</span></span><span leaf="">,</span><br><span leaf=""> hours=</span><span><span leaf="">1</span></span><span leaf="">,</span><br><span leaf=""> args=[DOCUMENT_DIR, doc_manager],</span><br><span leaf="">)</span><br><br><span leaf="">scheduler.start()</span><br>
关键注意点
-
文档版本管理：给每个文档存版本号或哈希值，避免重复入库
-
原子性：删旧→加新要在一个事务里，中间挂了不能出现"旧删了新没加"的状态
-
向量ID映射：chunk在向量库里的ID要跟文档注册表关联，否则删不了
-
批量操作：大量文档更新时用批量写入，别一条条加
监控告警：出了问题马上知道
RAG系统上线后，这几个指标必须监控：
关键指标
| 指标
|
告警阈值
|
说明
检索延迟P95
|
> 500ms
|
向量库或Rerank出问题
| |
生成延迟P95
|
> 5s
|
模型限流或网络问题
| |
检索空结果率
|
> 10%
|
Embedding模型故障或知识库缺失
| |
Faithfulness得分
|
< 0.7
|
模型在编造答案
| |
Embedding API错误率
|
> 5%
|
API Key过期或限流
| |
向量库查询QPS
|
接近上限
|
需要扩容
|
监控实现
<span><span leaf="">import</span></span><span leaf=""> time</span><br><span><span leaf="">import</span></span><span leaf=""> logging</span><br><span><span leaf="">from</span></span><span leaf=""> functools </span><span><span leaf="">import</span></span><span leaf=""> wraps</span><br><br><span leaf="">logger = logging.getLogger(</span><span><span leaf="">"rag_monitor"</span></span><span leaf="">)</span><br><br><span><span><span leaf="">def</span></span><span leaf=""> </span><span><span leaf="">monitor_rag</span></span><span><span leaf="">(func)</span></span><span leaf="">:</span></span><br><span leaf=""> </span><span><span leaf="">"""RAG调用监控装饰器"""</span></span><br><span><span leaf=""> @wraps(func)</span></span><br><span leaf=""> </span><span><span><span leaf="">def</span></span><span leaf=""> </span><span><span leaf="">wrapper</span></span><span><span leaf="">(*args, **kwargs)</span></span><span leaf="">:</span></span><br><span leaf=""> start = time.time()</span><br><span leaf=""> </span><span><span leaf="">try</span></span><span leaf="">:</span><br><span leaf=""> result = func(*args, **kwargs)</span><br><span leaf=""> duration = time.time() - start</span><br><span leaf=""> </span><br><span leaf=""> </span><span><span leaf=""># 记录指标</span></span><br><span leaf=""> logger.info(json.dumps({</span><br><span leaf=""> </span><span><span leaf="">"function"</span></span><span leaf="">: func.__name__,</span><br><span leaf=""> </span><span><span leaf="">"duration_ms"</span></span><span leaf="">: round(duration * </span><span><span leaf="">1000</span></span><span leaf="">),</span><br><span leaf=""> </span><span><span leaf="">"status"</span></span><span leaf="">: </span><span><span leaf="">"success"</span></span><span leaf="">,</span><br><span leaf=""> </span><span><span leaf="">"num_results"</span></span><span leaf="">: len(result) </span><span><span leaf="">if</span></span><span leaf=""> isinstance(result, list) </span><span><span leaf="">else</span></span><span leaf=""> </span><span><span leaf="">1</span></span><span leaf="">,</span><br><span leaf=""> }))</span><br><span leaf=""> </span><br><span leaf=""> </span><span><span leaf=""># 延迟告警</span></span><br><span leaf=""> </span><span><span leaf="">if</span></span><span leaf=""> duration > </span><span><span leaf="">2.0</span></span><span leaf="">:</span><br><span leaf=""> logger.warning(</span><span><span leaf="">f"⚠️ </span><span><span leaf="">{func.__name__}</span></span><span leaf=""> 耗时</span><span><span leaf="">{duration:</span><span><span leaf="">.1</span></span><span leaf="">f}</span></span><span leaf="">s，超过2秒阈值"</span></span><span leaf="">)</span><br><span leaf=""> </span><br><span leaf=""> </span><span><span leaf="">return</span></span><span leaf=""> result</span><br><span leaf=""> </span><span><span leaf="">except</span></span><span leaf=""> Exception </span><span><span leaf="">as</span></span><span leaf=""> e:</span><br><span leaf=""> duration = time.time() - start</span><br><span leaf=""> logger.error(json.dumps({</span><br><span leaf=""> </span><span><span leaf="">"function"</span></span><span leaf="">: func.__name__,</span><br><span leaf=""> </span><span><span leaf="">"duration_ms"</span></span><span leaf="">: round(duration * </span><span><span leaf="">1000</span></span><span leaf="">),</span><br><span leaf=""> </span><span><span leaf="">"status"</span></span><span leaf="">: </span><span><span leaf="">"error"</span></span><span leaf="">,</span><br><span leaf=""> </span><span><span leaf="">"error"</span></span><span leaf="">: str(e),</span><br><span leaf=""> }))</span><br><span leaf=""> </span><span><span leaf="">raise</span></span><br><span leaf=""> </span><span><span leaf="">return</span></span><span leaf=""> wrapper</span><br><br><span><span leaf=""># 使用</span></span><br><span><span leaf="">@monitor_rag</span></span><br><span><span><span leaf="">def</span></span><span leaf=""> </span><span><span leaf="">search_knowledge_base</span></span><span><span leaf="">(query: str)</span></span><span leaf="">:</span></span><br><span leaf=""> </span><span><span leaf="">return</span></span><span leaf=""> hybrid_retriever.invoke(query)</span><br>
成本控制：别让Token烧光你的预算
RAG的成本主要来自三块：Embedding调用、LLM生成、向量库存储。
成本优化策略
1. Embedding缓存
相同文本不重复调Embedding API：
<span><span leaf="">from</span></span><span leaf=""> functools </span><span><span leaf="">import</span></span><span leaf=""> lru_cache</span><br><br><span><span leaf="">@lru_cache(maxsize=10000)</span></span><br><span><span><span leaf="">def</span></span><span leaf=""> </span><span><span leaf="">get_embedding</span></span><span><span leaf="">(text: str)</span></span><span leaf="">:</span></span><br><span leaf=""> </span><span><span leaf="">return</span></span><span leaf=""> embedding_model.embed_query(text)</span><br>
2. 上下文压缩
检索到的文档可能很长，压缩后再送模型：
<span><span leaf="">from</span></span><span leaf=""> langchain.retrievers </span><span><span leaf="">import</span></span><span leaf=""> ContextualCompressionRetriever</span><br><span><span leaf="">from</span></span><span leaf=""> langchain_community.document_compressors </span><span><span leaf="">import</span></span><span leaf=""> LLMChainExtractor</span><br><br><span leaf="">compressor = LLMChainExtractor.from_llm(llm=cheap_model)</span><br><span leaf="">compression_retriever = ContextualCompressionRetriever(</span><br><span leaf=""> base_compressor=compressor,</span><br><span leaf=""> base_retriever=retriever,</span><br><span leaf="">)</span><br>
3. 模型分级
不是所有问题都需要GPT-4o：
<span><span><span leaf="">def</span></span><span leaf=""> </span><span><span leaf="">get_model_for_query</span></span><span><span leaf="">(query: str)</span></span><span leaf="">:</span></span><br><span leaf=""> </span><span><span leaf="">"""根据问题复杂度选择模型"""</span></span><br><span leaf=""> simple_keywords = [</span><span><span leaf="">"是什么"</span></span><span leaf="">, </span><span><span leaf="">"多少"</span></span><span leaf="">, </span><span><span leaf="">"什么时候"</span></span><span leaf="">, </span><span><span leaf="">"在哪里"</span></span><span leaf="">]</span><br><span leaf=""> </span><span><span leaf="">if</span></span><span leaf=""> any(kw </span><span><span leaf="">in</span></span><span leaf=""> query </span><span><span leaf="">for</span></span><span leaf=""> kw </span><span><span leaf="">in</span></span><span leaf=""> simple_keywords):</span><br><span leaf=""> </span><span><span leaf="">return</span></span><span leaf=""> </span><span><span leaf="">"gpt-4o-mini"</span></span><span leaf=""> </span><span><span leaf=""># 简单问题用便宜模型</span></span><br><span leaf=""> </span><span><span leaf="">return</span></span><span leaf=""> </span><span><span leaf="">"gpt-4o"</span></span><span leaf=""> </span><span><span leaf=""># 复杂问题用好模型</span></span><br>
4. 检索结果数量控制
Top-K不是越大越好——5条够用的别拉10条，每多一条就多几百Token：
<span><span leaf=""># 根据问题类型调整K值</span></span><br><span><span><span leaf="">def</span></span><span leaf=""> </span><span><span leaf="">get_k_for_query</span></span><span><span leaf="">(query: str)</span></span><span leaf="">:</span></span><br><span leaf=""> </span><span><span leaf="">if</span></span><span leaf=""> </span><span><span leaf="">"对比"</span></span><span leaf=""> </span><span><span leaf="">in</span></span><span leaf=""> query </span><span><span leaf="">or</span></span><span leaf=""> </span><span><span leaf="">"分析"</span></span><span leaf=""> </span><span><span leaf="">in</span></span><span leaf=""> query:</span><br><span leaf=""> </span><span><span leaf="">return</span></span><span leaf=""> </span><span><span leaf="">8</span></span><span leaf=""> </span><span><span leaf=""># 需要多条文档对比</span></span><br><span leaf=""> </span><span><span leaf="">return</span></span><span leaf=""> </span><span><span leaf="">4</span></span><span leaf=""> </span><span><span leaf=""># 简单问答4条足够</span></span><br>
多租户架构：不同客户的数据隔离
如果你的RAG服务SaaS化，多个客户共用一套系统，数据隔离是刚需。
方案1：命名空间隔离
在同一个向量库中用元数据字段区分不同租户：
<span><span leaf=""># 入库时标记租户</span></span><br><span leaf="">chunks = splitter.split_documents(doc)</span><br><span><span leaf="">for</span></span><span leaf=""> chunk </span><span><span leaf="">in</span></span><span leaf=""> chunks:</span><br><span leaf=""> chunk.metadata[</span><span><span leaf="">"tenant_id"</span></span><span leaf="">] = </span><span><span leaf="">"client_001"</span></span><br><span leaf="">vectorstore.add_documents(chunks)</span><br><br><span><span leaf=""># 检索时过滤</span></span><br><span leaf="">results = vectorstore.similarity_search(</span><br><span leaf=""> query=</span><span><span leaf="">"年假政策"</span></span><span leaf="">,</span><br><span leaf=""> k=</span><span><span leaf="">5</span></span><span leaf="">,</span><br><span leaf=""> filter={</span><span><span leaf="">"tenant_id"</span></span><span leaf="">: </span><span><span leaf="">"client_001"</span></span><span leaf="">}, </span><span><span leaf=""># 只搜这个客户的数据</span></span><br><span leaf="">)</span><br>
Chroma、Milvus、Qdrant都支持元数据过滤。
方案2：独立Collection/索引
每个租户一个独立的向量库：
<span><span leaf=""># 租户A的向量库</span></span><br><span leaf="">vectorstore_a = Milvus(collection_name=</span><span><span leaf="">"client_001"</span></span><span leaf="">, ...)</span><br><br><span><span leaf=""># 租户B的向量库</span></span><br><span leaf="">vectorstore_b = Milvus(collection_name=</span><span><span leaf="">"client_002"</span></span><span leaf="">, ...)</span><br>
隔离更彻底但运维成本更高，适合对数据安全要求极高的客户。
部署方案
方案1：Docker Compose（小团队推荐）
<span><span leaf=""># docker-compose.yml</span></span><br><span><span leaf="">version:</span></span><span leaf=""> </span><span><span leaf="">'3.8'</span></span><br><span><span leaf="">services:</span></span><br><span leaf=""> </span><span><span leaf="">rag-api:</span></span><br><span leaf=""> </span><span><span leaf="">build:</span></span><span leaf=""> </span><span><span leaf="">.</span></span><br><span leaf=""> </span><span><span leaf="">ports:</span></span><br><span leaf=""> </span><span><span leaf="">-</span></span><span leaf=""> </span><span><span leaf="">"8000:8000"</span></span><br><span leaf=""> </span><span><span leaf="">environment:</span></span><br><span leaf=""> </span><span><span leaf="">-</span></span><span leaf=""> </span><span><span leaf="">OPENAI_API_KEY=${OPENAI_API_KEY}</span></span><br><span leaf=""> </span><span><span leaf="">-</span></span><span leaf=""> </span><span><span leaf="">MILVUS_HOST=milvus</span></span><br><span leaf=""> </span><span><span leaf="">depends_on:</span></span><br><span leaf=""> </span><span><span leaf="">-</span></span><span leaf=""> </span><span><span leaf="">milvus</span></span><br><span leaf=""> </span><span><span leaf="">-</span></span><span leaf=""> </span><span><span leaf="">redis</span></span><br><br><span leaf=""> </span><span><span leaf="">milvus:</span></span><br><span leaf=""> </span><span><span leaf="">image:</span></span><span leaf=""> </span><span><span leaf="">milvusdb/milvus:v2.4-latest</span></span><br><span leaf=""> </span><span><span leaf="">ports:</span></span><br><span leaf=""> </span><span><span leaf="">-</span></span><span leaf=""> </span><span><span leaf="">"19530:19530"</span></span><br><span leaf=""> </span><span><span leaf="">volumes:</span></span><br><span leaf=""> </span><span><span leaf="">-</span></span><span leaf=""> </span><span><span leaf="">milvus_data:/var/lib/milvus</span></span><br><br><span leaf=""> </span><span><span leaf="">redis:</span></span><br><span leaf=""> </span><span><span leaf="">image:</span></span><span leaf=""> </span><span><span leaf="">redis:7-alpine</span></span><br><span leaf=""> </span><span><span leaf="">ports:</span></span><br><span leaf=""> </span><span><span leaf="">-</span></span><span leaf=""> </span><span><span leaf="">"6379:6379"</span></span><br><br><span><span leaf="">volumes:</span></span><br><span leaf=""> </span><span><span leaf="">milvus_data:</span></span><br>
方案2：Kubernetes（大团队推荐）
适合需要自动扩缩容、多实例部署的场景。关键组件：
-
RAG API服务：Deployment + HPA自动扩缩容
-
Milvus集群：分布式部署，读写分离
-
Redis：缓存层
-
Prometheus + Grafana：监控
方案3：Serverless（低频场景）
用云函数+托管向量库，按调用计费：
-
API层：AWS Lambda / 阿里云函数计算
-
向量库：Pinecone / Zilliz Cloud（托管Milvus）
-
LLM：API调用
适合调用量不大（每天<1万次）、不需要7x24运行的场景。
本篇要点
| 要点
|
说明
分层架构
|
接入/编排/检索/生成/存储/数据管道六层
| |
增量索引
|
文档哈希做差量检测，别全量重建
| |
监控告警
|
检索P95<500ms，生成P95<5s，空结果率<10%
| |
成本控制
|
Embedding缓存+上下文压缩+模型分级+K值控制
| |
多租户
|
命名空间隔离(轻量)或独立Collection(彻底)
| |
部署方案
|
Docker Compose(小团队)/K8s(大团队)/Serverless(低频)
| |
上线Checklist
|
12项，逐条检查再发版
|
RAG生产上线Checklist
上线前逐条检查：
-
增量索引机制已实现，文档更新不用重建
-
检索延迟P95 < 500ms，生成延迟P95 < 5s
-
监控告警已配好：延迟、错误率、空结果率
-
Embedding API有降级方案（主模型限流时切备用）
-
上下文长度控制：SummarizationMiddleware或截断中间件
-
多租户数据隔离已验证
-
RAGAS评估流水线已搭建，基线分数已记录
-
文档注册表已实现，可追踪每个chunk的来源
-
缓存层已配好，重复查询不重复调模型
-
API Key从环境变量读取，未硬编码
-
错误处理完善：检索失败有兜底、模型超时有重试
-
日志记录完整：每次查询的query、检索结果、生成答案、耗时
前7篇回顾
7篇基础篇，从"RAG到底在干什么"到"生产部署Checklist"，RAG的核心流程走了一遍：
-
基础篇（第1篇）：RAG原理、三代架构演进、与微调和长上下文的对比
-
数据处理篇（第2篇）：6种切分策略、PDF解析、chunk_size调参
-
向量化篇（第3篇）：Embedding模型选型、向量数据库对比
-
检索优化篇（第4篇）：混合检索、Rerank重排序、上下文压缩
-
高级架构篇（第5篇）：查询改写、多跳检索、GraphRAG、Agentic RAG
-
评估篇（第6篇）：RAGAS指标、黄金测试集、A/B测试、Bad Case分析
-
部署篇（第7篇）：增量索引、监控告警、成本控制、多租户
学RAG最有效的方式是拿真实业务数据做一遍。从最基础的Naive RAG开始，遇到问题再针对性升级——先加混合检索，再加Rerank，再考虑查询改写。别一上来就GraphRAG + Agentic RAG，99%的项目用不着。
后面还有3篇前沿专题：GraphRAG知识图谱检索、多模态RAG、Agentic RAG实战。基础打牢之后再看这些，你会理解得更深。
系列如果对你有帮助，转发给同样在做RAG的朋友。你们的支持是我持续输出的动力。
你的RAG上线了吗？遇到过什么生产事故？评论区聊聊。
觉得有用就点个在看，下一篇讲GraphRAG——跨文档推理的终极方案。
- 原文作者：知识铺
- 原文链接：https://index.zshipu.com/geek/post/20260822/%E4%B8%8A%E7%BA%BF%E4%B8%80%E5%91%A8%E5%B0%B1%E7%BF%BB%E8%BD%A6RAG%E7%94%9F%E4%BA%A7%E9%83%A8%E7%BD%B2%E9%81%BF%E5%9D%91%E6%8C%87%E5%8D%97/
- 版权声明：本作品采用知识共享署名-非商业性使用-禁止演绎 4.0 国际许可协议进行许可，非商业转载请注明出处（作者，原文链接），商业转载请联系作者获得授权。
- 免责声明：本页面内容均来源于站内编辑发布，部分信息来源互联网，并不意味着本站赞同其观点或者证实其内容的真实性，如涉及版权等问题，请立即联系客服进行更改或删除，保证您的合法权益。转载请注明来源，欢迎对文章中的引用来源进行考证，欢迎指出任何有错误或不够清晰的表达。也可以邮件至 sblig@126.com
