https://www.hansking.cn/blog A posts created with Next.js and Tailwind.css en-us admin@hansking.cn (HANS) admin@hansking.cn (HANS) Sun, 08 Dec 2024 10:00:00 GMT https://www.hansking.cn/blog/2026/2024-nian-zong-jie-LLM-ying-yong-gong-cheng-hua-jian-cha-qing-dan https://www.hansking.cn/blog/2026/2024-nian-zong-jie-LLM-ying-yong-gong-cheng-hua-jian-cha-qing-dan [{"value":"2024 年总结：LLM 应用工程化检查清单","url":"#2024-年总结llm-应用工程化检查清单","depth":1},{"value":"一、先确认：你解决的是问题，不是展示模型","url":"#一先确认你解决的是问题不是展示模型","depth":2},{"value":"二、模型选型：不要只看榜单第一名","url":"#二模型选型不要只看榜单第一名","depth":2},{"value":"三、Prompt：从“会说话”走向可维护","url":"#三prompt从会说话走向可维护","depth":2},{"value":"四、结构化输出：让字符串接受类型检查","url":"#四结构化输出让字符串接受类型检查","depth":2},{"value":"五、RAG：真正难的是数据和闭环","url":"#五rag真正难的是数据和闭环","depth":2},{"value":"六、多模态：输入变多，验证不能变少","url":"#六多模态输入变多验证不能变少","depth":2},{"value":"七、长上下文：能放下不等于能用好","url":"#七长上下文能放下不等于能用好","depth":2},{"value":"八、微调：先确认 Prompt 和数据真的不够","url":"#八微调先确认-prompt-和数据真的不够","depth":2},{"value":"九、安全：Prompt Injection 不是模型一句话能解决的","url":"#九安全prompt-injection-不是模型一句话能解决的","depth":2},{"value":"十、可观测性：没有 Trace 就没有复盘","url":"#十可观测性没有-trace-就没有复盘","depth":2},{"value":"十一、成本：把一次回答当成完整账单","url":"#十一成本把一次回答当成完整账单","depth":2},{"value":"十二、发布门禁：不要把生产当实验室","url":"#十二发布门禁不要把生产当实验室","depth":2},{"value":"十三、我在 2024 年留下的工程排序","url":"#十三我在-2024-年留下的工程排序","depth":2},{"value":"十四、我的总结：工程化就是把惊喜变成可重复","url":"#十四我的总结工程化就是把惊喜变成可重复","depth":2}] Sun, 08 Dec 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLM工程实践RAG多模态 https://www.hansking.cn/blog/2026/AI-ying-yong-de-yin-si-bao-hu-tuo-min-liu-cun-yu-shu-ju-bian-jie https://www.hansking.cn/blog/2026/AI-ying-yong-de-yin-si-bao-hu-tuo-min-liu-cun-yu-shu-ju-bian-jie [{"value":"AI 应用的隐私保护：脱敏、留存与数据边界","url":"#ai-应用的隐私保护脱敏留存与数据边界","depth":1},{"value":"一、先画数据流，再谈安全","url":"#一先画数据流再谈安全","depth":2},{"value":"二、脱敏不是全局替换字符串","url":"#二脱敏不是全局替换字符串","depth":2},{"value":"脱敏要有上下文","url":"#脱敏要有上下文","depth":3},{"value":"三、脱敏应该放在哪一层","url":"#三脱敏应该放在哪一层","depth":2},{"value":"四、向量库也不是“只存数字”","url":"#四向量库也不是只存数字","depth":2},{"value":"五、留存策略：不是存得越久越有用","url":"#五留存策略不是存得越久越有用","depth":2},{"value":"六、第三方模型调用前，问清楚四个问题","url":"#六第三方模型调用前问清楚四个问题","depth":2},{"value":"七、隐私保护要进入测试和审计","url":"#七隐私保护要进入测试和审计","depth":2},{"value":"总结：数据边界要比模型边界更早建立","url":"#总结数据边界要比模型边界更早建立","depth":2}] Tue, 17 Mar 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLM隐私保护数据安全RAG https://www.hansking.cn/blog/2026/GraphRAG-zhi-shi-tu-pu-ru-he-bu-zu-xiang-liang-jian-suo https://www.hansking.cn/blog/2026/GraphRAG-zhi-shi-tu-pu-ru-he-bu-zu-xiang-liang-jian-suo [{"value":"GraphRAG：知识图谱如何补足向量检索","url":"#graphrag知识图谱如何补足向量检索","depth":1},{"value":"一、向量检索和图检索各自擅长什么","url":"#一向量检索和图检索各自擅长什么","depth":2},{"value":"二、先定义图谱中的实体和关系","url":"#二先定义图谱中的实体和关系","depth":2},{"value":"三、从文档抽取关系时，模型只是候选生成器","url":"#三从文档抽取关系时模型只是候选生成器","depth":2},{"value":"四、实体归一化比抽取更容易被低估","url":"#四实体归一化比抽取更容易被低估","depth":2},{"value":"五、GraphRAG 的查询流程","url":"#五graphrag-的查询流程","depth":2},{"value":"六、图查询要回答“为什么”，不只是“是什么”","url":"#六图查询要回答为什么不只是是什么","depth":2},{"value":"七、社区发现适合全局问题，但要谨慎解释","url":"#七社区发现适合全局问题但要谨慎解释","depth":2},{"value":"八、权限必须同时存在于图和文本检索","url":"#八权限必须同时存在于图和文本检索","depth":2},{"value":"九、更新和冲突是图谱的长期难题","url":"#九更新和冲突是图谱的长期难题","depth":2},{"value":"十、什么时候不该上 GraphRAG","url":"#十什么时候不该上-graphrag","depth":2},{"value":"十一、如何评测 GraphRAG","url":"#十一如何评测-graphrag","depth":2},{"value":"十二、我的总结：图谱不是向量检索的替代品","url":"#十二我的总结图谱不是向量检索的替代品","depth":2}] Sun, 15 Sep 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLMRAGGraphRAG知识图谱 https://www.hansking.cn/blog/2026/RAG-shu-ju-guan-dao-jie-xi-PDF-wang-ye-yu-biao-ge https://www.hansking.cn/blog/2026/RAG-shu-ju-guan-dao-jie-xi-PDF-wang-ye-yu-biao-ge [{"value":"RAG 数据管道：解析 PDF、网页与表格","url":"#rag-数据管道解析-pdf网页与表格","depth":1},{"value":"一、先承认：不同格式不是同一种文档","url":"#一先承认不同格式不是同一种文档","depth":2},{"value":"二、PDF：先处理布局，再处理文本","url":"#二pdf先处理布局再处理文本","depth":2},{"value":"三、网页：正文只是页面的一小部分","url":"#三网页正文只是页面的一小部分","depth":2},{"value":"四、表格：关系比自然语言更重要","url":"#四表格关系比自然语言更重要","depth":2},{"value":"五、切分：不是越小越好，也不是越大越完整","url":"#五切分不是越小越好也不是越大越完整","depth":2},{"value":"六、清洗要克制：噪声和信息不能一锅端","url":"#六清洗要克制噪声和信息不能一锅端","depth":2},{"value":"七、去重和版本：避免知识库里有十个“同一份答案”","url":"#七去重和版本避免知识库里有十个同一份答案","depth":2},{"value":"八、权限元数据必须跟着每个 chunk 走","url":"#八权限元数据必须跟着每个-chunk-走","depth":2},{"value":"九、质量检查要在入库时就发生","url":"#九质量检查要在入库时就发生","depth":2},{"value":"十、增量处理：文档变化时只做必要工作","url":"#十增量处理文档变化时只做必要工作","depth":2},{"value":"十一、如何评测管道是否真的变好了","url":"#十一如何评测管道是否真的变好了","depth":2},{"value":"十二、我的总结：RAG 的第一模型是数据管道","url":"#十二我的总结rag-的第一模型是数据管道","depth":2}] Sun, 29 Sep 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLMRAG数据工程文档解析 https://www.hansking.cn/blog/2026/RAG-jin-jie-HybridSearch-yu-Reranker https://www.hansking.cn/blog/2026/RAG-jin-jie-HybridSearch-yu-Reranker [{"value":"RAG 进阶：Hybrid Search 与 Reranker","url":"#rag-进阶hybrid-search-与-reranker","depth":1},{"value":"一、关键词检索和向量检索看的是不同东西","url":"#一关键词检索和向量检索看的是不同东西","depth":2},{"value":"二、Hybrid Search 的基本流程","url":"#二hybrid-search-的基本流程","depth":2},{"value":"三、分数融合的三种思路","url":"#三分数融合的三种思路","depth":2},{"value":"1. 加权分数","url":"#1-加权分数","depth":3},{"value":"2. Reciprocal Rank Fusion","url":"#2-reciprocal-rank-fusion","depth":3},{"value":"3. 学习排序","url":"#3-学习排序","depth":3},{"value":"四、Reranker 为什么还能再排一次","url":"#四reranker-为什么还能再排一次","depth":2},{"value":"五、候选数量和最终 K 怎么选","url":"#五候选数量和最终-k-怎么选","depth":2},{"value":"六、查询改写和混合检索要配合","url":"#六查询改写和混合检索要配合","depth":2},{"value":"七、过滤要在检索层，不要交给 Reranker","url":"#七过滤要在检索层不要交给-reranker","depth":2},{"value":"八、离线评测要拆开看","url":"#八离线评测要拆开看","depth":2},{"value":"九、延迟预算决定是否值得使用 Reranker","url":"#九延迟预算决定是否值得使用-reranker","depth":2},{"value":"十、什么时候 Hybrid Search 仍然不够","url":"#十什么时候-hybrid-search-仍然不够","depth":2},{"value":"十一、我的总结：先扩大机会，再集中判断","url":"#十一我的总结先扩大机会再集中判断","depth":2}] Sun, 01 Sep 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLMRAGHybrid SearchReranker https://www.hansking.cn/blog/2026duo-zu-hu-AI-ying-yong-ge-li-shu-ju-pei-e-he-jian-suo-quan-xian https://www.hansking.cn/blog/2026duo-zu-hu-AI-ying-yong-ge-li-shu-ju-pei-e-he-jian-suo-quan-xian [{"value":"多租户 AI 应用：隔离数据、配额和检索权限","url":"#多租户-ai-应用隔离数据配额和检索权限","depth":1},{"value":"一、先定义清楚：什么叫隔离","url":"#一先定义清楚什么叫隔离","depth":2},{"value":"二、数据隔离：不要只靠开发者记得加 where","url":"#二数据隔离不要只靠开发者记得加-where","depth":2},{"value":"1. 共享表是便宜而实用的起点","url":"#1-共享表是便宜而实用的起点","depth":3},{"value":"2. 用数据库能力做第二道保险","url":"#2-用数据库能力做第二道保险","depth":3},{"value":"三、向量检索：最容易被忽略的租户边界","url":"#三向量检索最容易被忽略的租户边界","depth":2},{"value":"四、缓存隔离：答案缓存也会泄漏","url":"#四缓存隔离答案缓存也会泄漏","depth":2},{"value":"五、配额和限流：公平不是把所有人都拒之门外","url":"#五配额和限流公平不是把所有人都拒之门外","depth":2},{"value":"六、日志和计费：隔离不仅是看不见，还要算得清","url":"#六日志和计费隔离不仅是看不见还要算得清","depth":2},{"value":"七、用测试证明隔离，而不是用信心证明","url":"#七用测试证明隔离而不是用信心证明","depth":2},{"value":"总结：把租户当成系统边界，而不是业务字段","url":"#总结把租户当成系统边界而不是业务字段","depth":2}] Sun, 08 Mar 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLM多租户RAG权限管理 https://www.hansking.cn/blog/2026sheng-chan-ji-RAG-li-xian-ping-ce-zai-xian-jian-kong-yu-zeng-liang-suo-yin https://www.hansking.cn/blog/2026sheng-chan-ji-RAG-li-xian-ping-ce-zai-xian-jian-kong-yu-zeng-liang-suo-yin [{"value":"生产级 RAG：离线评测、在线监控与增量索引","url":"#生产级-rag离线评测在线监控与增量索引","depth":1},{"value":"一、先把生产级 RAG 想清楚","url":"#一先把生产级-rag-想清楚","depth":2},{"value":"二、离线评测：上线之前先知道它好不好","url":"#二离线评测上线之前先知道它好不好","depth":2},{"value":"1. 先做黄金数据集，而不是先选评测框架","url":"#1-先做黄金数据集而不是先选评测框架","depth":3},{"value":"2. 把检索和生成分开打分","url":"#2-把检索和生成分开打分","depth":3},{"value":"3. 评测必须能比较版本","url":"#3-评测必须能比较版本","depth":3},{"value":"三、在线监控：让系统自己告诉你哪里坏了","url":"#三在线监控让系统自己告诉你哪里坏了","depth":2},{"value":"1. 先监控基础设施，再监控答案质量","url":"#1-先监控基础设施再监控答案质量","depth":3},{"value":"四、增量索引：知识库更新不能每次从头再来","url":"#四增量索引知识库更新不能每次从头再来","depth":2},{"value":"更新过程要可恢复","url":"#更新过程要可恢复","depth":3},{"value":"五、我最后留下的一份上线清单","url":"#五我最后留下的一份上线清单","depth":2}] Fri, 27 Feb 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLMRAG评测工程实践 https://www.hansking.cn/blog/2026chang-wen-dang-wen-da-Map-ReduceRefine-yu-fen-ceng-zhai-yao https://www.hansking.cn/blog/2026chang-wen-dang-wen-da-Map-ReduceRefine-yu-fen-ceng-zhai-yao [{"value":"长文档问答：Map-Reduce、Refine 与分层摘要","url":"#长文档问答map-reducerefine-与分层摘要","depth":1},{"value":"一、先定义长文档问题的类型","url":"#一先定义长文档问题的类型","depth":2},{"value":"二、Map-Reduce：先分开处理，再合并结果","url":"#二map-reduce先分开处理再合并结果","depth":2},{"value":"三、Refine：按顺序逐步改进答案","url":"#三refine按顺序逐步改进答案","depth":2},{"value":"四、分层摘要：先建立地图，再深入细节","url":"#四分层摘要先建立地图再深入细节","depth":2},{"value":"五、三种方法如何选择","url":"#五三种方法如何选择","depth":2},{"value":"六、切分是所有长文档方案的地基","url":"#六切分是所有长文档方案的地基","depth":2},{"value":"七、上下文预算不是只有一个数字","url":"#七上下文预算不是只有一个数字","depth":2},{"value":"八、引用和证据要从第一步保留","url":"#八引用和证据要从第一步保留","depth":2},{"value":"九、Map-Reduce 的并行并不等于无限并发","url":"#九map-reduce-的并行并不等于无限并发","depth":2},{"value":"十、Refine 要防止答案漂移","url":"#十refine-要防止答案漂移","depth":2},{"value":"十一、摘要质量怎么评测","url":"#十一摘要质量怎么评测","depth":2},{"value":"十二、长文档问答不等于长上下文问答","url":"#十二长文档问答不等于长上下文问答","depth":2},{"value":"十三、我的总结：先建立地图，再回答问题","url":"#十三我的总结先建立地图再回答问题","depth":2}] Sun, 14 Apr 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLM长上下文RAG文档问答