"use strict"; const {Fragment: _Fragment, jsx: _jsx, jsxs: _jsxs} = arguments[0]; const {useMDXComponents: _provideComponents} = arguments[0]; function _createMdxContent(props) { const _components = { p: "p", ..._provideComponents(), ...props.components }, {Badge, Heading, Mermaid} = _components; if (!Badge) _missingMdxReference("Badge", true); if (!Heading) _missingMdxReference("Heading", true); if (!Mermaid) _missingMdxReference("Mermaid", true); return _jsxs(_Fragment, { children: [_jsx(Heading, { level: "1", id: "graph-rag", children: "Graph RAG" }), "\n", _jsxs(_components.p, { children: [_jsx(Badge, { icon: "clock", color: "green", children: "Written: 2026.06" }), "\n本页整理业务流程动画中的 Mermaid 流程图。"] }), "\n", _jsx(Heading, { level: "2", id: "1-%E5%9C%A8%E7%BA%BF%E6%8F%90%E9%97%AE%E6%B5%81%E7%A8%8B", children: "1. 在线提问流程" }), "\n", _jsx(Mermaid, { chart: "flowchart TD\n USER[\"👤 用户浏览器
输入问题 + 选择业务分类\"]\n WS[\"🔌 WS /api/stream
在线问答唯一入口\"]\n\n USER -->|\"1. 当前页面直接建立长连接\"| WS\n\n subgraph WSAPI[\"WebSocket 入口层\"]\n WACCEPT[\"接受连接
websocket.accept()\"]\n WRATE[\"消息级限流
check_rate_limit()\"]\n WPARSE[\"解析 JSON
QueryServiceContext\"]\n WACCEPT --> WRATE --> WPARSE\n end\n\n WS --> WSAPI\n\n subgraph PIPELINE[\"RAG Pipeline - stream_query() 在线主链路\"]\n S0[\"🏗️ Stage 0
create_query_context()
场景/数据域/会话/trace_id/KB版本\"]\n S1[\"🧭 Stage 1
decide_route()
direct_answer / faq_exact / retrieval\"]\n S3[\"🧠 Stage 2
prepare_retrieval()
历史加载→意图识别→source/计划→变体\"]\n S4[\"📋 Stage 3
search_faq()
FAQ 混合检索 + 置信度判断\"]\n S5[\"📄 Stage 4
search_doc() + Reranker
文档混合检索→CrossEncoder精排\"]\n S6[\"📊 Stage 5
prepare_answer()
上下文筛选→去重→格式化来源\"]\n S7[\"🤖 Stage 6
stream_llm_answer()
LLM 流式生成→引用增强\"]\n S8[\"💾 Stage 7
finish_success()
保存历史→Trace→end事件\"]\n\n S0 --> S1\n S1 -->|\"route=retrieval\"| S3\n S1 -->|\"route=direct_answer\"| DIRECT[\"🚫 直接返回\"]\n S1 -->|\"route=faq_exact\"| FAST[\"⚡ FAQ 直出
intent=FAQ_QUERY\"]\n S3 --> S4\n S4 -->|\"分数 < 阈值\"| S5\n S4 -->|\"分数 ≥ 阈值\"| FAQOUT[\"📋 FAQ 直出\"]\n S5 --> S6 --> S7 --> S8\n end\n\n WSAPI --> PIPELINE\n\n subgraph EXTERNAL[\"外部系统依赖\"]\n MILVUS[(\"🗄️ Milvus 2.5
Dense + Sparse
混合检索\")]\n EMBED[\"🧮 BGE-M3
文本→1024维向量\"]\n RERANK[\"⚖️ BGE Reranker
CrossEncoder 精排\"]\n LLM[\"🤖 LLM
qwen-plus
DashScope/OpenAI兼容\"]\n MYSQL[(\"🗃️ MySQL 8.0
历史·摘要·反馈\")]\n end\n\n S4 -.->|\"Dense向量\"| EMBED\n S4 -.->|\"Hybrid Search\"| MILVUS\n S5 -.->|\"Dense向量\"| EMBED\n S5 -.->|\"Hybrid Search\"| MILVUS\n S5 -.->|\"精排打分\"| RERANK\n S7 -.->|\"流式生成\"| LLM\n S8 -.->|\"读写历史\"| MYSQL\n S3 -.->|\"意图分类/改写\"| LLM\n S1 -.->|\"FAQ 精确试探\"| MILVUS\n\n STREAMOUT[\"📤 WebSocket 逐 token 推送
打字机效果 + 来源引用\"]\n S7 --> STREAMOUT\n S8 --> STREAMOUT\n\n style USER fill:#1e3a5f,stroke:#3b82f6,stroke-width:2px\n style WS fill:#1e293b,stroke:#818cf8,stroke-width:2px\n style FAST fill:#064e3b,stroke:#34d399\n style FAQOUT fill:#064e3b,stroke:#34d399\n style DIRECT fill:#713f12,stroke:#fbbf24\n style STREAMOUT fill:#1e3a5f,stroke:#3b82f6,stroke-width:2px\n style S5 fill:#131c2e,stroke:#f59e0b,stroke-width:2px\n style S6 fill:#131c2e,stroke:#f59e0b,stroke-width:2px\n style LLM fill:#500724,stroke:#ec4899\n style MILVUS fill:#0f766e,stroke:#06b6d4\n style EMBED fill:#3b0764,stroke:#a78bfa\n style RERANK fill:#451a03,stroke:#f59e0b\n style MYSQL fill:#1a2e05,stroke:#84cc16" }), "\n", _jsx(Heading, { level: "2", id: "2-%E7%A6%BB%E7%BA%BF%E5%BB%BA%E5%BA%93%E6%B5%81%E7%A8%8B", children: "2. 离线建库流程" }), "\n", _jsx(Mermaid, { chart: "flowchart TD\n START[\"📁 上传文档
PDF / Word / MD / CSV / Excel\"]\n\n subgraph LOAD[\"1. 文档加载 - Document Loaders 注册表\"]\n REGISTRY[\"按后缀匹配 Loader
.pdf→PyPDFLoader
.docx→Docx2txtLoader
.md→TextLoader
.csv/.xlsx→TableLoader\"]\n LOADED[\"Document 对象
page_content + metadata\"]\n REGISTRY --> LOADED\n end\n\n START --> LOAD\n\n subgraph NORMALIZE[\"2. 文档标准化\"]\n META[\"补充元数据
source / file_name / doc_id
scenario_id / kb_version\"]\n DEDUP[\"去重检测
SHA256 文件指纹\"]\n META --> DEDUP\n end\n\n LOAD --> NORMALIZE\n\n subgraph CHUNK[\"3. 文档切分 - Parent-Child Chunking\"]\n HEADER[\"Markdown 按标题切分
保留 h1/h2/h3 层级\"]\n PARENT[\"父块切分
RecursiveCharacterTextSplitter
chunk_size=2000\"]\n CHILD[\"子块切分
RecursiveCharacterTextSplitter
chunk_size=500, overlap=50\"]\n HEADER --> PARENT --> CHILD\n end\n\n NORMALIZE --> CHUNK\n\n subgraph EMBEDDING[\"4. 向量化\"]\n DENSE[\"BGE-M3 → Dense 向量
1024 维 · L2 归一化\"]\n SPARSE[\"Milvus BM25 → Sparse 向量
服务端自动生成\"]\n end\n\n CHILD --> EMBEDDING\n\n subgraph STORE[\"5. 写入 Milvus\"]\n SCHEMA[\"自动创建 Collection
pk / text / dense / sparse
+ source / kb_version\"]\n INDEX[\"创建向量索引
Dense AUTO/HNSW + Sparse BM25\"]\n INSERT[\"insert / upsert
+ flush 持久化\"]\n LOADCOL[\"load_collection()
加载到内存\"]\n SCHEMA --> INDEX --> INSERT --> LOADCOL\n end\n\n EMBEDDING --> STORE\n\n subgraph MANIFEST[\"6. 索引清单 - IndexManifest\"]\n RECORD[\"记录入库文件清单
file_path + chunk_ids\"]\n DIFF[\"增量更新对比
新增/修改/删除文件\"]\n CLEANUP[\"清理旧 chunk
delete by chunk_id\"]\n RECORD --> DIFF --> CLEANUP\n end\n\n STORE --> MANIFEST\n READY[\"✅ 知识就绪
可被在线问答检索\"]\n MANIFEST --> READY\n\n FAQIN[\"📋 FAQ CSV 入库
并行通道\"]\n FAQIN -->|\"CSV → Milvus FAQ Collection
标准问题 + 标准答案\"| READY\n\n style START fill:#1e3a5f,stroke:#3b82f6,stroke-width:2px\n style READY fill:#064e3b,stroke:#34d399,stroke-width:2px\n style EMBEDDING fill:#3b0764,stroke:#a78bfa\n style STORE fill:#0f766e,stroke:#06b6d4\n style MANIFEST fill:#451a03,stroke:#fbbf24\n style FAQIN fill:#134e4a,stroke:#2dd4bf" }), "\n", _jsx(Heading, { level: "2", id: "3-%E7%B3%BB%E7%BB%9F%E6%80%BB%E8%A7%88%E6%B5%81%E7%A8%8B", children: "3. 系统总览流程" }), "\n", _jsx(Mermaid, { chart: "flowchart TB\n subgraph STARTUP[\"🚀 系统启动\"]\n APP[\"app.py\"]\n PREFLIGHT[\"Preflight Check
LLM · Milvus · MySQL · 模型 · 场景
任一缺失 → 启动失败\"]\n WARM[\"检索栈预热
预加载 Embedding + Reranker 模型\"]\n APP --> PREFLIGHT --> WARM\n end\n\n subgraph ONLINE[\"🟢 在线问答(实时)\"]\n direction TB\n B[\"浏览器\"]\n API2[\"FastAPI
HTTP + WebSocket\"]\n QASVC[\"QAService
编排层\"]\n PIPE[\"RAG Pipeline
7 阶段管线\"]\n B -->|\"提问\"| API2 --> QASVC --> PIPE -->|\"流式答案\"| B\n end\n\n subgraph OFFLINE[\"🔵 离线入库(异步)\"]\n direction TB\n DOC[\"文档上传\"]\n INGEST[\"入库管线
加载→标准化→切分→向量化→存储\"]\n MANIFEST2[\"IndexManifest
增量更新\"]\n DOC --> INGEST --> MANIFEST2\n end\n\n subgraph STORAGE[\"💾 存储层\"]\n MILVUS2[(\"Milvus 2.5
Dense 向量
+ Sparse BM25\")]\n MYSQL2[(\"MySQL 8.0
聊天历史
会话摘要
用户反馈
IndexManifest\")]\n MINIO[(\"MinIO
向量索引文件
binlog\")]\n end\n\n subgraph MODELS[\"🧠 模型层(本地部署)\"]\n BGE3[\"BGE-M3
Embedding
1024维\"]\n RERANKER[\"BGE Reranker
CrossEncoder
精排\"]\n end\n\n subgraph LLM2[\"🤖 LLM 服务\"]\n DASHSCOPE[\"DashScope
qwen-plus
OpenAI 兼容接口\"]\n end\n\n ONLINE -->|\"检索\"| STORAGE\n ONLINE -->|\"向量化\"| MODELS\n ONLINE -->|\"生成\"| LLM2\n OFFLINE -->|\"写入\"| STORAGE\n OFFLINE -->|\"向量化\"| MODELS\n STARTUP -->|\"校验\"| STORAGE\n STARTUP -->|\"校验\"| MODELS\n STARTUP -->|\"校验\"| LLM2\n\n style ONLINE fill:#064e3b,stroke:#34d399,stroke-width:2px\n style OFFLINE fill:#1e3a5f,stroke:#3b82f6,stroke-width:2px\n style STARTUP fill:#451a03,stroke:#f59e0b,stroke-width:2px\n style STORAGE fill:#0f766e,stroke:#06b6d4,stroke-width:2px\n style MODELS fill:#3b0764,stroke:#a78bfa,stroke-width:2px\n style LLM2 fill:#500724,stroke:#ec4899,stroke-width:2px" }), "\n", _jsx(Heading, { level: "2", id: "4-%E4%BB%A3%E7%A0%81%E8%B7%9F%E7%BB%83%E6%B5%81%E7%A8%8B", children: "4. 代码跟练流程" }), "\n", _jsx(Mermaid, { chart: "flowchart LR\n START[\"课程入口
第 01-04 章
环境/RAG/LangChain/Milvus 基础\"]\n\n subgraph ONLINE_COURSE[\"在线问答主链路：第 05-12 章\"]\n CH05[\"05 意图分类与路由入口
Stage 1 decide_route()
direct_answer / faq_exact / retrieval\"]\n CH06[\"06 检索策略与动态计划
Stage 2 检索计划
run_faq / run_doc / top_k / 阈值\"]\n CH07[\"07 查询改写与变体生成
Stage 2 追问改写 + 查询变体
rewrite_query / query_variants\"]\n CH08[\"08 Milvus 混合检索
Stage 3-4
FAQ/doc Hybrid Search + Rerank\"]\n CH09[\"09 QAService 核心编排
服务编排层
把检索能力封装成问答服务\"]\n CH10[\"10 RAG Pipeline 主流程
Stage 0-7 串联
在线问答闭环\"]\n CH11[\"11 Prompt 工程与 Profile
Stage 5-6
上下文组织 + 回答口径\"]\n CH12[\"12 FastAPI 服务入口
WebSocket / HTTP
浏览器可调用\"]\n end\n\n subgraph ENGINEERING_COURSE[\"工程化与治理：第 13-19 章\"]\n CH13[\"13 启动前置校验
Preflight Check
依赖缺失直接失败\"]\n CH14[\"14 知识库版本管理
Stage 0 KB 版本
查询绑定有效版本\"]\n CH15[\"15 数据隔离
场景 / 数据域 / source filter
避免跨域回答\"]\n CH16[\"16 文档入库链路
离线入库流程
加载→切分→向量化→写入 Milvus\"]\n CH17[\"17 质量评估
问答评测集
衡量召回和回答质量\"]\n CH18[\"18 测试体系与门禁
自动化测试 / 接口验收
防止链路回退\"]\n CH19[\"19 可观测性与 Trace
Stage 7 Trace / 日志
定位每一步耗时和错误\"]\n end\n\n DONE[\"最终项目闭环
可启动 / 可入库 / 可问答 / 可测试 / 可观测\"]\n\n START --> CH05 --> CH06 --> CH07 --> CH08 --> CH09 --> CH10 --> CH11 --> CH12\n CH12 --> CH13 --> CH14 --> CH15 --> CH16 --> CH17 --> CH18 --> CH19 --> DONE\n\n CH05 -.对应.-> ONLINE_STAGE1[\"动画在线流程 Stage 1
decide_route()\"]\n CH06 -.对应.-> ONLINE_STAGE2A[\"动画在线流程 Stage 2
检索计划\"]\n CH07 -.对应.-> ONLINE_STAGE2B[\"动画在线流程 Stage 2
追问改写/查询变体\"]\n CH08 -.对应.-> ONLINE_STAGE34[\"动画在线流程 Stage 3-4
FAQ/文档检索\"]\n CH10 -.对应.-> ONLINE_STAGE07[\"动画在线流程 Stage 0-7
主链路串联\"]\n CH16 -.对应.-> OFFLINE_FLOW[\"动画离线入库流程
文档/FAQ 入库\"]\n CH19 -.对应.-> TRACE_FLOW[\"动画在线流程 Stage 7
Trace/end 事件\"]\n\n style START fill:#1e3a5f,stroke:#3b82f6,stroke-width:2px\n style DONE fill:#064e3b,stroke:#34d399,stroke-width:2px\n style CH05 fill:#713f12,stroke:#fbbf24,stroke-width:2px\n style CH06 fill:#713f12,stroke:#fbbf24,stroke-width:2px\n style CH07 fill:#713f12,stroke:#fbbf24,stroke-width:2px\n style CH08 fill:#0f766e,stroke:#06b6d4,stroke-width:2px\n style CH09 fill:#1e293b,stroke:#818cf8,stroke-width:2px\n style CH10 fill:#1e293b,stroke:#818cf8,stroke-width:2px\n style CH11 fill:#500724,stroke:#ec4899,stroke-width:2px\n style CH12 fill:#1e3a5f,stroke:#3b82f6,stroke-width:2px\n style CH13 fill:#451a03,stroke:#f59e0b\n style CH14 fill:#0f766e,stroke:#06b6d4\n style CH15 fill:#0f766e,stroke:#06b6d4\n style CH16 fill:#1e3a5f,stroke:#3b82f6\n style CH17 fill:#3b0764,stroke:#a78bfa\n style CH18 fill:#3b0764,stroke:#a78bfa\n style CH19 fill:#451a03,stroke:#f59e0b\n style ONLINE_STAGE1 fill:#1e293b,stroke:#fbbf24\n style ONLINE_STAGE2A fill:#1e293b,stroke:#fbbf24\n style ONLINE_STAGE2B fill:#1e293b,stroke:#fbbf24\n style ONLINE_STAGE34 fill:#0f766e,stroke:#06b6d4\n style ONLINE_STAGE07 fill:#1e293b,stroke:#818cf8\n style OFFLINE_FLOW fill:#1e3a5f,stroke:#3b82f6\n style TRACE_FLOW fill:#451a03,stroke:#f59e0b" })] }); } function MDXContent(props = {}) { const {wrapper: MDXLayout} = { ..._provideComponents(), ...props.components }; return MDXLayout ? _jsx(MDXLayout, { ...props, children: _jsx(_createMdxContent, { ...props }) }) : _createMdxContent(props); } return { default: MDXContent }; function _missingMdxReference(id, component) { throw new Error("Expected " + (component ? "component" : "object") + " `" + id + "` to be defined: you likely forgot to import, pass, or provide it."); }

​Graph RAG

​1. 在线提问流程

​2. 离线建库流程

​3. 系统总览流程

​4. 代码跟练流程

Graph RAG

1. 在线提问流程

2. 离线建库流程

3. 系统总览流程

4. 代码跟练流程