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-xi-tong-cheng-ben-he-suan-TokenGPU-cun-chu-yu-ren-gong-cheng-ben https://www.hansking.cn/blog/2026/AI-xi-tong-cheng-ben-he-suan-TokenGPU-cun-chu-yu-ren-gong-cheng-ben [{"value":"AI 系统成本核算：Token、GPU、存储与人工成本","url":"#ai-系统成本核算tokengpu存储与人工成本","depth":1},{"value":"一、先算一次“成功任务”的成本","url":"#一先算一次成功任务的成本","depth":2},{"value":"二、Token 成本不只取决于用户说了多少话","url":"#二token-成本不只取决于用户说了多少话","depth":2},{"value":"三、Embedding 和索引是容易被漏算的长期成本","url":"#三embedding-和索引是容易被漏算的长期成本","depth":2},{"value":"四、自建 GPU 不能只除以运行小时","url":"#四自建-gpu-不能只除以运行小时","depth":2},{"value":"五、缓存是成本优化，也是数据边界","url":"#五缓存是成本优化也是数据边界","depth":2},{"value":"六、把人工成本放进模型里","url":"#六把人工成本放进模型里","depth":2},{"value":"七、成本归因要能定位到功能和租户","url":"#七成本归因要能定位到功能和租户","depth":2},{"value":"八、优化顺序：先消除浪费，再换模型","url":"#八优化顺序先消除浪费再换模型","depth":2},{"value":"总结：便宜不是目标，单位价值才是","url":"#总结便宜不是目标单位价值才是","depth":2}] Mon, 15 Jun 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLM成本优化GPU工程实践 https://www.hansking.cn/blog/2026/Guardrails-shu-chu-guo-l-ce-le-pan-duan-yu-ren-gong-dou-di https://www.hansking.cn/blog/2026/Guardrails-shu-chu-guo-l-ce-le-pan-duan-yu-ren-gong-dou-di [{"value":"Guardrails：输出过滤、策略判断与人工兜底","url":"#guardrails输出过滤策略判断与人工兜底","depth":1},{"value":"一、先画出分层护栏","url":"#一先画出分层护栏","depth":2},{"value":"二、输入分类不是给用户贴永久标签","url":"#二输入分类不是给用户贴永久标签","depth":2},{"value":"三、输出过滤要分层，不要只有关键词黑名单","url":"#三输出过滤要分层不要只有关键词黑名单","depth":2},{"value":"四、策略引擎应该独立于模型","url":"#四策略引擎应该独立于模型","depth":2},{"value":"五、工具护栏比回答过滤更重要","url":"#五工具护栏比回答过滤更重要","depth":2},{"value":"六、人工兜底不是把所有问题扔给客服","url":"#六人工兜底不是把所有问题扔给客服","depth":2},{"value":"七、误伤和漏放要用不同指标看","url":"#七误伤和漏放要用不同指标看","depth":2},{"value":"八、拒绝之后要给出安全的下一步","url":"#八拒绝之后要给出安全的下一步","depth":2},{"value":"九、外部内容要被当成数据，不是指令","url":"#九外部内容要被当成数据不是指令","depth":2},{"value":"十、规则发布要像代码发布一样谨慎","url":"#十规则发布要像代码发布一样谨慎","depth":2},{"value":"十一、我的总结：护栏的目标是安全地继续工作","url":"#十一我的总结护栏的目标是安全地继续工作","depth":2}] Sun, 27 Oct 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLM安全Guardrails工程实践 https://www.hansking.cn/blog/2026/LLM-guan-ce-TracingToken-yong-liang-yu-shi-bai-yang-ben https://www.hansking.cn/blog/2026/LLM-guan-ce-TracingToken-yong-liang-yu-shi-bai-yang-ben [{"value":"LLM 观测：Tracing、Token 用量与失败样本","url":"#llm-观测tracingtoken-用量与失败样本","depth":1},{"value":"一、先把一次 LLM 请求拆成 Span","url":"#一先把一次-llm-请求拆成-span","depth":2},{"value":"二、Trace ID 不是用户 ID","url":"#二trace-id-不是用户-id","depth":2},{"value":"三、Token 计量不只是为了算账","url":"#三token-计量不只是为了算账","depth":2},{"value":"四、延迟要拆成用户真正等待的阶段","url":"#四延迟要拆成用户真正等待的阶段","depth":2},{"value":"五、错误要分成可行动的类别","url":"#五错误要分成可行动的类别","depth":2},{"value":"六、失败样本是最有价值的训练材料","url":"#六失败样本是最有价值的训练材料","depth":2},{"value":"七、在线质量监控不能只靠用户点赞","url":"#七在线质量监控不能只靠用户点赞","depth":2},{"value":"八、告警要能推动动作","url":"#八告警要能推动动作","depth":2},{"value":"九、隐私与调试之间要划边界","url":"#九隐私与调试之间要划边界","depth":2},{"value":"十、Trace 回放要避免产生真实副作用","url":"#十trace-回放要避免产生真实副作用","depth":2},{"value":"十一、观测数据本身也要版本化","url":"#十一观测数据本身也要版本化","depth":2},{"value":"十二、把观测接到研发闭环","url":"#十二把观测接到研发闭环","depth":2},{"value":"十三、我的总结：观测不是给模型装摄像头","url":"#十三我的总结观测不是给模型装摄像头","depth":2}] Sun, 10 Nov 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLM可观测性Tracing工程实践 https://www.hansking.cn/blog/2026/Prompt-ban-ben-guan-li-yu-hui-gui-ce-shi https://www.hansking.cn/blog/2026/Prompt-ban-ben-guan-li-yu-hui-gui-ce-shi [{"value":"Prompt 版本管理与回归测试","url":"#prompt-版本管理与回归测试","depth":1},{"value":"一、Prompt 先拆成模板和变量","url":"#一prompt-先拆成模板和变量","depth":2},{"value":"二、版本号要能解释变化","url":"#二版本号要能解释变化","depth":2},{"value":"三、先建立回归集，再修改 Prompt","url":"#三先建立回归集再修改-prompt","depth":2},{"value":"四、不同类型的输出用不同断言","url":"#四不同类型的输出用不同断言","depth":2},{"value":"五、控制随机性和测试成本","url":"#五控制随机性和测试成本","depth":2},{"value":"六、灰度发布比直接全量更诚实","url":"#六灰度发布比直接全量更诚实","depth":2},{"value":"七、Prompt 变更要检查安全边界","url":"#七prompt-变更要检查安全边界","depth":2},{"value":"八、回滚要做到一键可用","url":"#八回滚要做到一键可用","depth":2},{"value":"九、把人工反馈变成工程输入","url":"#九把人工反馈变成工程输入","depth":2},{"value":"总结：Prompt 也需要代码应有的纪律","url":"#总结prompt-也需要代码应有的纪律","depth":2}] Wed, 18 Feb 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLMPrompt Engineering回归测试工程实践 https://www.hansking.cn/blog/2026/ToolCalling-de-ke-kao-xing-Schema-chong-shi-yu-mi-deng https://www.hansking.cn/blog/2026/ToolCalling-de-ke-kao-xing-Schema-chong-shi-yu-mi-deng [{"value":"Tool Calling 的可靠性：Schema、重试与幂等","url":"#tool-calling-的可靠性schema重试与幂等","depth":1},{"value":"一、工具描述不是实现，Schema 也不是护身符","url":"#一工具描述不是实现schema-也不是护身符","depth":2},{"value":"二、把调用拆成准备、确认、执行三个阶段","url":"#二把调用拆成准备确认执行三个阶段","depth":2},{"value":"三、错误分类决定重试策略","url":"#三错误分类决定重试策略","depth":2},{"value":"四、幂等是副作用工具的生命线","url":"#四幂等是副作用工具的生命线","depth":2},{"value":"五、模型错误与系统错误要分开记录","url":"#五模型错误与系统错误要分开记录","depth":2},{"value":"六、返回给模型的错误要能帮助它做下一步","url":"#六返回给模型的错误要能帮助它做下一步","depth":2},{"value":"七、什么时候应该拒绝让模型调用","url":"#七什么时候应该拒绝让模型调用","depth":2},{"value":"八、我的总结：让模型负责提议，让系统负责承诺","url":"#八我的总结让模型负责提议让系统负责承诺","depth":2}] Sun, 27 Apr 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgentTool Calling工程实践 https://www.hansking.cn/blog/2026gou-jian-huang-jin-shu-ju-ji-rang-ping-ce-cong-gan-jue-bian-cheng-zheng-ju https://www.hansking.cn/blog/2026gou-jian-huang-jin-shu-ju-ji-rang-ping-ce-cong-gan-jue-bian-cheng-zheng-ju [{"value":"构建黄金数据集：让评测从感觉变成证据","url":"#构建黄金数据集让评测从感觉变成证据","depth":1},{"value":"一、先定义数据集要回答什么问题","url":"#一先定义数据集要回答什么问题","depth":2},{"value":"二、样本要覆盖真实分布和危险边界","url":"#二样本要覆盖真实分布和危险边界","depth":2},{"value":"三、期望答案要写成可判断的标准","url":"#三期望答案要写成可判断的标准","depth":2},{"value":"四、标注需要规则，也需要人","url":"#四标注需要规则也需要人","depth":2},{"value":"五、时间切分和客户切分防止污染","url":"#五时间切分和客户切分防止污染","depth":2},{"value":"六、版本化和变更记录不能省","url":"#六版本化和变更记录不能省","depth":2},{"value":"七、指标要和数据集结构对应","url":"#七指标要和数据集结构对应","depth":2},{"value":"八、黄金集要服务模型、RAG 和 Agent","url":"#八黄金集要服务模型rag-和-agent","depth":2},{"value":"九、数据集不是建完就结束","url":"#九数据集不是建完就结束","depth":2},{"value":"总结：黄金数据集是一种工程记忆","url":"#总结黄金数据集是一种工程记忆","depth":2}] Mon, 09 Feb 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLM数据集评测工程实践 https://www.hansking.cn/blog/2026ban-quan-yu-xun-lian-shu-ju-gong-cheng-shi-xu-yao-zhi-dao-de-bian-jie https://www.hansking.cn/blog/2026ban-quan-yu-xun-lian-shu-ju-gong-cheng-shi-xu-yao-zhi-dao-de-bian-jie [{"value":"版权与训练数据：工程师需要知道的边界","url":"#版权与训练数据工程师需要知道的边界","depth":1},{"value":"一、先区分四件常被混在一起的事","url":"#一先区分四件常被混在一起的事","depth":2},{"value":"二、数据来源要能追溯","url":"#二数据来源要能追溯","depth":2},{"value":"三、RAG 不等于自动获得版权许可","url":"#三rag-不等于自动获得版权许可","depth":2},{"value":"四、训练集和评测集最容易失控","url":"#四训练集和评测集最容易失控","depth":2},{"value":"五、删除请求要能一路追到底","url":"#五删除请求要能一路追到底","depth":2},{"value":"六、模型供应商也是数据链路的一部分","url":"#六模型供应商也是数据链路的一部分","depth":2},{"value":"七、输出也要做版权和来源审查","url":"#七输出也要做版权和来源审查","depth":2},{"value":"八、我建议团队建立一张“数据护照”","url":"#八我建议团队建立一张数据护照","depth":2},{"value":"总结：未知不是允许，能追溯才有底气","url":"#总结未知不是允许能追溯才有底气","depth":2}] Thu, 26 Mar 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLM版权数据合规工程实践 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