记录项目实践、工程思考与持续学习的个人技术博客。https://project-blog.pages.dev/https://project-blog.pages.dev/projects/knowforge-rag/https://project-blog.pages.dev/projects/knowforge-rag/从可运行的工程全貌切入，说明 RAG 的核心链路、KnowForge RAG 的系统边界与技术栈，并完成 Docker Compose 本地环境的启动验证。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/02-rag-fundamentals/https://project-blog.pages.dev/projects/knowforge-rag/02-rag-fundamentals/聚焦 Embedding、向量相似度、Dense 与 Sparse 检索以及 Reranker，建立混合检索所依赖的数学直觉和工程判断依据。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/03-langchain-ecosystem/https://project-blog.pages.dev/projects/knowforge-rag/03-langchain-ecosystem/梳理 LangChain 组件在 KnowForge RAG 中的实际职责，明确 Runnable、Prompt、消息、解析器及历史记录与业务编排之间的边界。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/04-milvus-index-and-operations/https://project-blog.pages.dev/projects/knowforge-rag/04-milvus-index-and-operations/从服务化向量数据库的选型出发，展开 Milvus 的数据模型、索引机制、检索参数与基础操作，为后续混合检索链路建立可靠的数据底座。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/04-5-vibe-coding-rag-practice/https://project-blog.pages.dev/projects/knowforge-rag/04-5-vibe-coding-rag-practice/以可运行的简易混合检索 RAG 为目标，使用 Codex 推进需求约束、代码实现、验证与修正，串联 Dense、Sparse、HNSW 和 Milvus 等核心能力。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/05-intent-classification/https://project-blog.pages.dev/projects/knowforge-rag/05-intent-classification/构建查询入口的分层决策机制，通过规则候选、BERT 模型增强、决策网关仲裁和知识查询兜底，将用户请求稳定地路由到正确链路。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/06-retrieval-strategy/https://project-blog.pages.dev/projects/knowforge-rag/06-retrieval-strategy/将入口阶段产生的 IntentResult 转换为可执行的 RetrievalPlan，集中定义召回通道、过滤边界、候选规模和重排策略。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/07-query-rewrite-variants/https://project-blog.pages.dev/projects/knowforge-rag/07-query-rewrite-variants/处理多轮对话中的省略与指代，并通过受控的查询变体扩大召回覆盖范围，同时避免改写偏离用户原始意图。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/08-milvus-hybrid-search/https://project-blog.pages.dev/projects/knowforge-rag/08-milvus-hybrid-search/落地 Milvus Dense 与 Sparse Hybrid Search，覆盖数据前置检查、过滤表达式、结果融合、候选去重以及 Reranker 精排。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/09-qaservice-orchestration/https://project-blog.pages.dev/projects/knowforge-rag/09-qaservice-orchestration/以 QAService 作为应用服务门面，组织查询准备、检索和流式生成的调用顺序，让业务编排与底层能力保持清晰分工。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/10-rag-pipeline/https://project-blog.pages.dev/projects/knowforge-rag/10-rag-pipeline/贯通 RAG Pipeline 的八个执行阶段，明确 FAQ 快速路径、上下文筛选与截断、Prompt Profile 选择、变量注入及答案引用增强。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/11-fastapi-async/https://project-blog.pages.dev/projects/knowforge-rag/11-fastapi-async/围绕 KnowForge RAG 的 API 层说明异步请求模型、路由、中间件、依赖注入与 WebSocket 流式输出，呈现接口到核心服务的完整衔接。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/12-app-entry-preflight/https://project-blog.pages.dev/projects/knowforge-rag/12-app-entry-preflight/拆解 FastAPI 应用启动过程，将关键依赖的前置校验与运行态容错分开，确保服务只在配置和基础设施满足条件时接收请求。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/13-knowledge-governance/https://project-blog.pages.dev/projects/knowforge-rag/13-knowledge-governance/以一次完整的知识库发布为主线，串联数据边界、候选版本、增量入库、质量门禁和 active 切换，形成可追踪且可回滚的治理流程。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/14-rag-quality/https://project-blog.pages.dev/projects/knowforge-rag/14-rag-quality/建立覆盖检索与答案的本地评测体系，通过固定样本、JSON 报告、质量门禁和 Bad Case 回流，让系统改动具备可比较的质量依据。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/15-testing-system/https://project-blog.pages.dev/projects/knowforge-rag/15-testing-system/按纯逻辑、核心服务、API 保护和交付验收划分测试层级，用稳定且可重复的检查覆盖 RAG 系统的关键工程边界。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/16-observability-capacity/https://project-blog.pages.dev/projects/knowforge-rag/16-observability-capacity/围绕日志、指标与 Trace 建立线上诊断链路，并将 LangSmith 适配、问题回流、容量评估、压测和监控告警纳入生产治理。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/17-docker-delivery/https://project-blog.pages.dev/projects/knowforge-rag/17-docker-delivery/从交付视角整合镜像、容器、Compose、网络、卷挂载、健康检查与环境变量，给出跨 Windows 和 Linux 的部署、验收及排障路径。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/appendix-a-pydantic/https://project-blog.pages.dev/projects/knowforge-rag/appendix-a-pydantic/集中说明 Pydantic 在请求模型、配置管理和结构化输出中的作用，用类型声明与运行时校验收紧跨模块数据边界。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/appendix-b-sha256-fingerprint/https://project-blog.pages.dev/projects/knowforge-rag/appendix-b-sha256-fingerprint/解析 SHA256 内容摘要如何生成稳定指纹，并说明 IndexManifest 如何据此识别文件变化、跳过未变内容和组织增量入库。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/appendix-c-hnsw-index/https://project-blog.pages.dev/projects/knowforge-rag/appendix-c-hnsw-index/从高维向量暴力搜索的成本出发，拆解 HNSW 的分层小世界图、近似搜索过程，以及 M、efConstruction、ef 等参数的性能权衡。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/appendix-d-crossencoder-reranker/https://project-blog.pages.dev/projects/knowforge-rag/appendix-d-crossencoder-reranker/说明 CrossEncoder 如何对召回候选进行联合语义评分，并结合 BGE Reranker 展开精度收益、推理成本、策略开关与失效场景。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/appendix-e-recursive-splitter/https://project-blog.pages.dev/projects/knowforge-rag/appendix-e-recursive-splitter/解析 RecursiveCharacterTextSplitter 的分隔符递归过程，并量化 chunk_size、chunk_overlap 与文档结构对切分完整性和检索质量的影响。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/appendix-f-embedding-models/https://project-blog.pages.dev/projects/knowforge-rag/appendix-f-embedding-models/从文本向量化的完整过程出发，对比维度、池化、归一化和 MTEB 指标，说明 BGE-M3 等模型在中文 RAG 中的选型与调优依据。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/appendix-g-chunking-strategy/https://project-blog.pages.dev/projects/knowforge-rag/appendix-g-chunking-strategy/围绕 Parent-Child Chunking 的召回粒度与上下文完整性，展开父子块生成、检索后展开、参数选择和不同文档类型的切分策略。Sun, 13 Sep 2026 00:00:00 GMThttps://project-blog.pages.dev/projects/knowforge-rag/appendix-h-tool-foundations/https://project-blog.pages.dev/projects/knowforge-rag/appendix-h-tool-foundations/汇总跨模块复用的基础设施代码，覆盖时间与 JSON 读写、稳定哈希、元数据存储、Pydantic 模型、配置、日志和 MySQL 惰性连接。Sun, 13 Sep 2026 00:00:00 GMT