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/2026/2025-nian-zong-jie-cong-Copilot-dao-ke-zhi-xing-Agent https://www.hansking.cn/blog/2026/2025-nian-zong-jie-cong-Copilot-dao-ke-zhi-xing-Agent [{"value":"2025 年总结：从 Copilot 到可执行 Agent","url":"#2025-年总结从-copilot-到可执行-agent","depth":1},{"value":"一、Copilot 解决的是“帮我写”，Agent 面对的是“帮我做”","url":"#一copilot-解决的是帮我写agent-面对的是帮我做","depth":2},{"value":"二、工具调用是能力，也是风险入口","url":"#二工具调用是能力也是风险入口","depth":2},{"value":"三、推理模型让“先想清楚再行动”变得可用","url":"#三推理模型让先想清楚再行动变得可用","depth":2},{"value":"四、Agent 需要记忆，但记忆应该由系统保存","url":"#四agent-需要记忆但记忆应该由系统保存","depth":2},{"value":"五、可靠 Agent 必须考虑重复和失败","url":"#五可靠-agent-必须考虑重复和失败","depth":2},{"value":"六、MCP 和统一协议带来了新的连接方式","url":"#六mcp-和统一协议带来了新的连接方式","depth":2},{"value":"七、Agent 评测不能只问“答得像不像”","url":"#七agent-评测不能只问答得像不像","depth":2},{"value":"八、人机协作不是退步，是安全阀","url":"#八人机协作不是退步是安全阀","depth":2},{"value":"九、2025 年留下的三个教训","url":"#九2025-年留下的三个教训","depth":2},{"value":"总结：从辅助生成到值得托付","url":"#总结从辅助生成到值得托付","depth":2}] Sun, 07 Dec 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgentCopilot技术总结 https://www.hansking.cn/blog/2026/AICodingAgent-de-dai-ma-ku-li-jie-yu-bu-ding-sheng-cheng https://www.hansking.cn/blog/2026/AICodingAgent-de-dai-ma-ku-li-jie-yu-bu-ding-sheng-cheng [{"value":"AI Coding Agent 的代码库理解与补丁生成","url":"#ai-coding-agent-的代码库理解与补丁生成","depth":1},{"value":"一、代码库理解不是把所有文件塞进上下文","url":"#一代码库理解不是把所有文件塞进上下文","depth":2},{"value":"二、建立适合检索的代码库索引","url":"#二建立适合检索的代码库索引","depth":2},{"value":"三、先把自然语言任务变成变更计划","url":"#三先把自然语言任务变成变更计划","depth":2},{"value":"四、上下文选择要有排序和预算","url":"#四上下文选择要有排序和预算","depth":2},{"value":"五、补丁生成：最小改动比聪明重构更重要","url":"#五补丁生成最小改动比聪明重构更重要","depth":2},{"value":"六、工具调用必须有明确边界","url":"#六工具调用必须有明确边界","depth":2},{"value":"七、验证不是最后点一下测试按钮","url":"#七验证不是最后点一下测试按钮","depth":2},{"value":"八、用回归测试保护代码库记忆","url":"#八用回归测试保护代码库记忆","depth":2},{"value":"九、让人审查决策，而不是审查一团文字","url":"#九让人审查决策而不是审查一团文字","depth":2},{"value":"总结：Agent 的能力上限取决于证据链","url":"#总结agent-的能力上限取决于证据链","depth":2}] Fri, 01 May 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLMCoding Agent软件工程代码生成 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/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/AI-xi-tong-de-SLO-zhi-liang-yan-chi-ke-yong-xing-yu-an-quan https://www.hansking.cn/blog/2026/AI-xi-tong-de-SLO-zhi-liang-yan-chi-ke-yong-xing-yu-an-quan [{"value":"AI 系统的 SLO：质量、延迟、可用性与安全","url":"#ai-系统的-slo质量延迟可用性与安全","depth":1},{"value":"一、先分清 SLI、SLO 和 SLA","url":"#一先分清-slislo-和-sla","depth":2},{"value":"二、AI 系统至少需要四类 SLO","url":"#二ai-系统至少需要四类-slo","depth":2},{"value":"1. 质量 SLO","url":"#1-质量-slo","depth":3},{"value":"2. 延迟 SLO","url":"#2-延迟-slo","depth":3},{"value":"3. 可用性 SLO","url":"#3-可用性-slo","depth":3},{"value":"4. 安全 SLO","url":"#4-安全-slo","depth":3},{"value":"三、质量指标要有采样和标签","url":"#三质量指标要有采样和标签","depth":2},{"value":"四、错误预算让团队敢于做改进","url":"#四错误预算让团队敢于做改进","depth":2},{"value":"五、降级要提前设计，不要故障时临时发挥","url":"#五降级要提前设计不要故障时临时发挥","depth":2},{"value":"六、发布门禁要同时看离线和在线","url":"#六发布门禁要同时看离线和在线","depth":2},{"value":"七、告警要连接到行动","url":"#七告警要连接到行动","depth":2},{"value":"八、SLO 也要防止被滥用","url":"#八slo-也要防止被滥用","depth":2},{"value":"总结：SLO 是团队对用户体验的共同承诺","url":"#总结slo-是团队对用户体验的共同承诺","depth":2}] Wed, 24 Jun 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLMSLO可观测性可靠性 https://www.hansking.cn/blog/2026/AgentMemory-duan-qi-ji-yi-chang-qi-ji-yi-yu-yong-hu-hua-xiang https://www.hansking.cn/blog/2026/AgentMemory-duan-qi-ji-yi-chang-qi-ji-yi-yu-yong-hu-hua-xiang [{"value":"Agent Memory：短期记忆、长期记忆与用户画像","url":"#agent-memory短期记忆长期记忆与用户画像","depth":1},{"value":"一、先区分四种记忆","url":"#一先区分四种记忆","depth":2},{"value":"1. 短期上下文","url":"#1-短期上下文","depth":3},{"value":"2. 长期事实","url":"#2-长期事实","depth":3},{"value":"3. 用户偏好","url":"#3-用户偏好","depth":3},{"value":"4. 任务记忆","url":"#4-任务记忆","depth":3},{"value":"二、短期记忆首先是上下文预算问题","url":"#二短期记忆首先是上下文预算问题","depth":2},{"value":"三、长期事实必须有来源和置信度","url":"#三长期事实必须有来源和置信度","depth":2},{"value":"四、用户画像不是越详细越好","url":"#四用户画像不是越详细越好","depth":2},{"value":"五、记忆读取也需要权限","url":"#五记忆读取也需要权限","depth":2},{"value":"六、记忆写入不要完全交给模型","url":"#六记忆写入不要完全交给模型","depth":2},{"value":"七、更新要避免旧记忆和新记忆打架","url":"#七更新要避免旧记忆和新记忆打架","depth":2},{"value":"八、过期和删除是记忆的一部分","url":"#八过期和删除是记忆的一部分","depth":2},{"value":"九、记忆检索要评测“帮忙还是添乱”","url":"#九记忆检索要评测帮忙还是添乱","depth":2},{"value":"总结：真正的记忆包含忘记的能力","url":"#总结真正的记忆包含忘记的能力","depth":2}] Sun, 06 Jul 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgentMemory隐私保护 https://www.hansking.cn/blog/2026/Agent-shi-bai-mo-shi-xun-huan-huan-jue-gong-ju-wu-yong-yu-shi-kong https://www.hansking.cn/blog/2026/Agent-shi-bai-mo-shi-xun-huan-huan-jue-gong-ju-wu-yong-yu-shi-kong [{"value":"Agent 失败模式：循环、幻觉、工具误用与失控","url":"#agent-失败模式循环幻觉工具误用与失控","depth":1},{"value":"一、无限循环：任务在动，结果不动","url":"#一无限循环任务在动结果不动","depth":2},{"value":"二、状态漂移：模型记住的和系统发生的不是一回事","url":"#二状态漂移模型记住的和系统发生的不是一回事","depth":2},{"value":"三、幻觉参数：工具调用合法，含义却是假的","url":"#三幻觉参数工具调用合法含义却是假的","depth":2},{"value":"四、工具误用：模型选择了错误能力","url":"#四工具误用模型选择了错误能力","depth":2},{"value":"五、权限越界：最不能靠 Prompt 解决","url":"#五权限越界最不能靠-prompt-解决","depth":2},{"value":"六、无止境重试：把故障放大成事故","url":"#六无止境重试把故障放大成事故","depth":2},{"value":"七、计划漂移：目标在执行中悄悄改变","url":"#七计划漂移目标在执行中悄悄改变","depth":2},{"value":"八、如何设计止损机制","url":"#八如何设计止损机制","depth":2},{"value":"九、把失败变成可观察事件","url":"#九把失败变成可观察事件","depth":2},{"value":"十、用故障注入测试 Agent","url":"#十用故障注入测试-agent","depth":2},{"value":"总结：失控不是模型突然变坏，而是边界没有落地","url":"#总结失控不是模型突然变坏而是边界没有落地","depth":2}] Sun, 20 Jul 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgent可靠性安全工程 https://www.hansking.cn/blog/2026/Agent-shi-shen-me-mo-xing-gong-ju-zhuang-tai-yu-xun-huan https://www.hansking.cn/blog/2026/Agent-shi-shen-me-mo-xing-gong-ju-zhuang-tai-yu-xun-huan [{"value":"Agent 是什么：模型、工具、状态与循环","url":"#agent-是什么模型工具状态与循环","depth":1},{"value":"一、先区分聊天机器人、工作流和 Agent","url":"#一先区分聊天机器人工作流和-agent","depth":2},{"value":"二、Agent 的四个基本零件","url":"#二agent-的四个基本零件","depth":2},{"value":"1. 模型：负责理解和选择","url":"#1-模型负责理解和选择","depth":3},{"value":"2. 工具：连接真实世界","url":"#2-工具连接真实世界","depth":3},{"value":"3. 状态：记住任务走到哪里","url":"#3-状态记住任务走到哪里","depth":3},{"value":"4. 循环：让结果影响下一步","url":"#4-循环让结果影响下一步","depth":3},{"value":"三、把 Agent 写成一个可观察的状态机","url":"#三把-agent-写成一个可观察的状态机","depth":2},{"value":"四、最小 Agent 循环","url":"#四最小-agent-循环","depth":2},{"value":"五、计划不是越详细越好","url":"#五计划不是越详细越好","depth":2},{"value":"六、工具调用时，模型只是在提议","url":"#六工具调用时模型只是在提议","depth":2},{"value":"七、记忆不等于把所有聊天记录塞回去","url":"#七记忆不等于把所有聊天记录塞回去","depth":2},{"value":"八、Agent 的失败不是一种失败","url":"#八agent-的失败不是一种失败","depth":2},{"value":"九、怎么评测一个 Agent","url":"#九怎么评测一个-agent","depth":2},{"value":"十、我的总结：自主性必须和边界一起增长","url":"#十我的总结自主性必须和边界一起增长","depth":2}] Sun, 30 Mar 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgent架构学习top https://www.hansking.cn/blog/2026/Agent-ping-ce-ren-wu-cheng-gong-l-gui-ji-zhi-liang-yu-cheng-ben https://www.hansking.cn/blog/2026/Agent-ping-ce-ren-wu-cheng-gong-l-gui-ji-zhi-liang-yu-cheng-ben [{"value":"Agent 评测：任务成功率、轨迹质量与成本","url":"#agent-评测任务成功率轨迹质量与成本","depth":1},{"value":"一、先定义任务成功而不是模型成功","url":"#一先定义任务成功而不是模型成功","depth":2},{"value":"二、任务成功率只是第一层","url":"#二任务成功率只是第一层","depth":2},{"value":"三、轨迹质量看 Agent 是怎么做的","url":"#三轨迹质量看-agent-是怎么做的","depth":2},{"value":"四、错误预算不能只奖励冒险","url":"#四错误预算不能只奖励冒险","depth":2},{"value":"五、成本要算到成功任务","url":"#五成本要算到成功任务","depth":2},{"value":"六、恢复能力是长任务的分水岭","url":"#六恢复能力是长任务的分水岭","depth":2},{"value":"七、评测集要有难度和风险分层","url":"#七评测集要有难度和风险分层","depth":2},{"value":"八、离线评测和线上反馈要互相连接","url":"#八离线评测和线上反馈要互相连接","depth":2},{"value":"九、评测报告要帮助做决策","url":"#九评测报告要帮助做决策","depth":2},{"value":"总结：Agent 评测是在评估一套行为","url":"#总结agent-评测是在评估一套行为","depth":2}] Sun, 17 Aug 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgent模型评测成本优化top https://www.hansking.cn/blog/2026/BrowserGym-yu-ren-wu-xing-Agent-ji-zhun-ce-shi https://www.hansking.cn/blog/2026/BrowserGym-yu-ren-wu-xing-Agent-ji-zhun-ce-shi [{"value":"BrowserGym 与任务型 Agent 基准测试","url":"#browsergym-与任务型-agent-基准测试","depth":1},{"value":"一、先定义任务的成功状态","url":"#一先定义任务的成功状态","depth":2},{"value":"二、浏览器环境要可复现","url":"#二浏览器环境要可复现","depth":2},{"value":"三、记录完整轨迹","url":"#三记录完整轨迹","depth":2},{"value":"四、指标至少分五类","url":"#四指标至少分五类","depth":2},{"value":"1. 任务结果","url":"#1-任务结果","depth":3},{"value":"2. 动作质量","url":"#2-动作质量","depth":3},{"value":"3. 效率","url":"#3-效率","depth":3},{"value":"4. 恢复能力","url":"#4-恢复能力","depth":3},{"value":"5. 安全性","url":"#5-安全性","depth":3},{"value":"五、轨迹评测和最终评测要分开","url":"#五轨迹评测和最终评测要分开","depth":2},{"value":"六、基准任务要覆盖边界","url":"#六基准任务要覆盖边界","depth":2},{"value":"七、不要让 Agent 记住答案","url":"#七不要让-agent-记住答案","depth":2},{"value":"八、工具执行要有沙箱","url":"#八工具执行要有沙箱","depth":2},{"value":"九、失败回放比排行榜更有用","url":"#九失败回放比排行榜更有用","depth":2},{"value":"十、BrowserGym 类基准如何服务真实产品","url":"#十browsergym-类基准如何服务真实产品","depth":2},{"value":"总结：Agent 的路迹也是答案","url":"#总结agent-的路迹也是答案","depth":2}] Sun, 31 Aug 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgentBrowserGym模型评测 https://www.hansking.cn/blog/2026/Chain-of-ThoughtReAct-yu-Tree-of-Thought-dui-bi https://www.hansking.cn/blog/2026/Chain-of-ThoughtReAct-yu-Tree-of-Thought-dui-bi [{"value":"Chain-of-Thought、ReAct 与 Tree-of-Thought 对比","url":"#chain-of-thoughtreact-与-tree-of-thought-对比","depth":1},{"value":"一、先看一张统一的地图","url":"#一先看一张统一的地图","depth":2},{"value":"二、Chain-of-Thought：先把一条路走清楚","url":"#二chain-of-thought先把一条路走清楚","depth":2},{"value":"三、ReAct：让思考遇到真实世界","url":"#三react让思考遇到真实世界","depth":2},{"value":"四、Tree-of-Thought：为什么要保留多个方向","url":"#四tree-of-thought为什么要保留多个方向","depth":2},{"value":"五、三种方法的核心差异","url":"#五三种方法的核心差异","depth":2},{"value":"六、选择方法的四个问题","url":"#六选择方法的四个问题","depth":2},{"value":"问题一：任务是否需要外部事实","url":"#问题一任务是否需要外部事实","depth":3},{"value":"问题二：第一步选错后能否继续修正","url":"#问题二第一步选错后能否继续修正","depth":3},{"value":"问题三：有没有可靠的评估函数","url":"#问题三有没有可靠的评估函数","depth":3},{"value":"问题四：预算和延迟是否允许","url":"#问题四预算和延迟是否允许","depth":3},{"value":"七、不要把三种方法当成 Prompt 技巧","url":"#七不要把三种方法当成-prompt-技巧","depth":2},{"value":"八、用实验而不是感觉做选择","url":"#八用实验而不是感觉做选择","depth":2},{"value":"九、我的总结：多走几步不等于走得更对","url":"#九我的总结多走几步不等于走得更对","depth":2}] Sun, 19 Jan 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLM推理ReActTree-of-Thought https://www.hansking.cn/blog/2026/Claude3-xi-lie-yu-mo-xing-neng-li-ping-ce-fang-fa https://www.hansking.cn/blog/2026/Claude3-xi-lie-yu-mo-xing-neng-li-ping-ce-fang-fa [{"value":"Claude 3 系列与模型能力评测方法","url":"#claude-3-系列与模型能力评测方法","depth":1},{"value":"一、Claude 3 系列为什么值得写进发展史","url":"#一claude-3-系列为什么值得写进发展史","depth":2},{"value":"二、先建立能力矩阵","url":"#二先建立能力矩阵","depth":2},{"value":"三、评测集比一句神奇 Prompt 更重要","url":"#三评测集比一句神奇-prompt-更重要","depth":2},{"value":"四、自动指标和人工指标要分工","url":"#四自动指标和人工指标要分工","depth":2},{"value":"五、不能把长答案当成高质量","url":"#五不能把长答案当成高质量","depth":2},{"value":"六、长上下文要测“找得到”和“用得对”","url":"#六长上下文要测找得到和用得对","depth":2},{"value":"七、多模态评测要看视觉细节","url":"#七多模态评测要看视觉细节","depth":2},{"value":"八、调用参数要固定，结果才可比较","url":"#八调用参数要固定结果才可比较","depth":2},{"value":"九、模型裁判可以用，但不能独裁","url":"#九模型裁判可以用但不能独裁","depth":2},{"value":"十、评测结果要能指导路由","url":"#十评测结果要能指导路由","depth":2},{"value":"十一、上线后继续评测","url":"#十一上线后继续评测","depth":2},{"value":"十二、我的总结：模型评测是地图，不是冠军赛","url":"#十二我的总结模型评测是地图不是冠军赛","depth":2}] Sun, 28 Apr 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLMClaude 3模型评测多模态 https://www.hansking.cn/blog/2026/ComputerUse-liu-lan-qi-cao-zuo-de-zhuang-tai-ji-she-ji https://www.hansking.cn/blog/2026/ComputerUse-liu-lan-qi-cao-zuo-de-zhuang-tai-ji-she-ji [{"value":"Computer Use：浏览器操作的状态机设计","url":"#computer-use浏览器操作的状态机设计","depth":1},{"value":"一、把浏览器任务建模成状态机","url":"#一把浏览器任务建模成状态机","depth":2},{"value":"二、观察结果要包含语义和版本","url":"#二观察结果要包含语义和版本","depth":2},{"value":"三、动作应该是受限的工具集合","url":"#三动作应该是受限的工具集合","depth":2},{"value":"四、等待不能靠固定 sleep","url":"#四等待不能靠固定-sleep","depth":2},{"value":"五、表单填写要分为填充和提交","url":"#五表单填写要分为填充和提交","depth":2},{"value":"六、动作结果要重新观察","url":"#六动作结果要重新观察","depth":2},{"value":"七、导航和域名必须受限","url":"#七导航和域名必须受限","depth":2},{"value":"八、失败恢复要区分可重试和不可重试","url":"#八失败恢复要区分可重试和不可重试","depth":2},{"value":"九、高风险动作必须人工确认","url":"#九高风险动作必须人工确认","depth":2},{"value":"十、评测看任务结果和动作过程","url":"#十评测看任务结果和动作过程","depth":2},{"value":"总结：浏览器操作首先是状态管理","url":"#总结浏览器操作首先是状态管理","depth":2}] Sun, 22 Jun 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMComputer UseAgent浏览器自动化 https://www.hansking.cn/blog/2026/ContextWindow-de-nei-cun-mo-xing-yu-cheng-ben-ji-suan https://www.hansking.cn/blog/2026/ContextWindow-de-nei-cun-mo-xing-yu-cheng-ben-ji-suan [{"value":"Context Window 的内存模型与成本计算","url":"#context-window-的内存模型与成本计算","depth":1},{"value":"一、Context Window 包含哪些东西","url":"#一context-window-包含哪些东西","depth":2},{"value":"二、Token 不是字符，语言和内容会改变比例","url":"#二token-不是字符语言和内容会改变比例","depth":2},{"value":"三、为什么上下文会占显存","url":"#三为什么上下文会占显存","depth":2},{"value":"四、并发让内存问题变成乘法","url":"#四并发让内存问题变成乘法","depth":2},{"value":"五、长上下文不等于模型有效利用全文","url":"#五长上下文不等于模型有效利用全文","depth":2},{"value":"六、什么时候应该用检索而不是全文输入","url":"#六什么时候应该用检索而不是全文输入","depth":2},{"value":"七、对话历史应该如何增长","url":"#七对话历史应该如何增长","depth":2},{"value":"八、成本计算要包含输入和输出","url":"#八成本计算要包含输入和输出","depth":2},{"value":"九、服务端要做上下文预算器","url":"#九服务端要做上下文预算器","depth":2},{"value":"十、长上下文压测应该怎么做","url":"#十长上下文压测应该怎么做","depth":2},{"value":"十一、上下文窗口会影响系统架构","url":"#十一上下文窗口会影响系统架构","depth":2},{"value":"十二、我的总结：窗口是预算，不是宣传数字","url":"#十二我的总结窗口是预算不是宣传数字","depth":2}] Sun, 31 Mar 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLMContext WindowKV Cache成本优化 https://www.hansking.cn/blog/2026/ContinuousBatching-yu-fu-wu-bing-fa https://www.hansking.cn/blog/2026/ContinuousBatching-yu-fu-wu-bing-fa [{"value":"Continuous Batching 与服务并发","url":"#continuous-batching-与服务并发","depth":1},{"value":"一、静态 Batch 为什么浪费","url":"#一静态-batch-为什么浪费","depth":2},{"value":"二、Continuous Batching 如何动态加入请求","url":"#二continuous-batching-如何动态加入请求","depth":2},{"value":"三、Prefill 和 Decode 是两种不同工作","url":"#三prefill-和-decode-是两种不同工作","depth":2},{"value":"四、KV Cache 是并发的硬约束","url":"#四kv-cache-是并发的硬约束","depth":2},{"value":"五、调度目标不只有吞吐","url":"#五调度目标不只有吞吐","depth":2},{"value":"六、队列和背压要尽早生效","url":"#六队列和背压要尽早生效","depth":2},{"value":"七、不同请求不要互相拖累","url":"#七不同请求不要互相拖累","depth":2},{"value":"八、流式输出不能掩盖调度问题","url":"#八流式输出不能掩盖调度问题","depth":2},{"value":"九、压测要模拟真实长度分布","url":"#九压测要模拟真实长度分布","depth":2},{"value":"总结：并发优化是在安排等待","url":"#总结并发优化是在安排等待","depth":2}] Sun, 12 Oct 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMContinuous BatchingGPU并发优化 https://www.hansking.cn/blog/2026/DPO-shi-zhan-xun-lian-yi-ge-geng-fu-he-ye-wu-pian-hao-de-mo-xing https://www.hansking.cn/blog/2026/DPO-shi-zhan-xun-lian-yi-ge-geng-fu-he-ye-wu-pian-hao-de-mo-xing [{"value":"DPO 实战：训练一个更符合业务偏好的模型","url":"#dpo-实战训练一个更符合业务偏好的模型","depth":1},{"value":"一、先回顾 SFT 和 RLHF 的位置","url":"#一先回顾-sft-和-rlhf-的位置","depth":2},{"value":"二、偏好数据格式很简单，数据含义不简单","url":"#二偏好数据格式很简单数据含义不简单","depth":2},{"value":"三、偏好标注最容易出现什么问题","url":"#三偏好标注最容易出现什么问题","depth":2},{"value":"把长度当质量","url":"#把长度当质量","depth":3},{"value":"把语气当事实","url":"#把语气当事实","depth":3},{"value":"偏好标准互相冲突","url":"#偏好标准互相冲突","depth":3},{"value":"负样本太差","url":"#负样本太差","depth":3},{"value":"四、DPO 的损失直觉","url":"#四dpo-的损失直觉","depth":2},{"value":"五、训练前先处理 Token 和截断","url":"#五训练前先处理-token-和截断","depth":2},{"value":"六、一个简化的训练流程","url":"#六一个简化的训练流程","depth":2},{"value":"七、LoRA 能降低成本，但不能降低评测要求","url":"#七lora-能降低成本但不能降低评测要求","depth":2},{"value":"八、DPO 训练中要盯哪些信号","url":"#八dpo-训练中要盯哪些信号","depth":2},{"value":"九、训练后评测必须包含“不可偏好的”样本","url":"#九训练后评测必须包含不可偏好的样本","depth":2},{"value":"十、如何处理矛盾偏好","url":"#十如何处理矛盾偏好","depth":2},{"value":"十一、什么时候 DPO 不适合","url":"#十一什么时候-dpo-不适合","depth":2},{"value":"十二、我的总结：偏好优化是在教模型“更像你的团队”","url":"#十二我的总结偏好优化是在教模型更像你的团队","depth":2}] Sun, 18 Aug 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLMDPO微调偏好优化 https://www.hansking.cn/blog/2026/DeepSeek-R1-de-xun-lian-si-lu-yu-kai-yuan-ying-xiang https://www.hansking.cn/blog/2026/DeepSeek-R1-de-xun-lian-si-lu-yu-kai-yuan-ying-xiang [{"value":"DeepSeek-R1 的训练思路与开源影响","url":"#deepseek-r1-的训练思路与开源影响","depth":1},{"value":"一、R1-Zero：先让模型在奖励里寻找能力","url":"#一r1-zero先让模型在奖励里寻找能力","depth":2},{"value":"二、为什么冷启动数据仍然重要","url":"#二为什么冷启动数据仍然重要","depth":2},{"value":"三、可验证奖励是整条路线的地基","url":"#三可验证奖励是整条路线的地基","depth":2},{"value":"四、GRPO 为什么成为重要的配套方法","url":"#四grpo-为什么成为重要的配套方法","depth":2},{"value":"五、推理时计算为什么重新成为能力变量","url":"#五推理时计算为什么重新成为能力变量","depth":2},{"value":"六、蒸馏让成果进入更小的模型","url":"#六蒸馏让成果进入更小的模型","depth":2},{"value":"七、开源影响不只是放出一个权重文件","url":"#七开源影响不只是放出一个权重文件","depth":2},{"value":"八、如何避免把 R1 学成几个口号","url":"#八如何避免把-r1-学成几个口号","depth":2},{"value":"九、我的总结：R1 改变的是问题的问法","url":"#九我的总结r1-改变的是问题的问法","depth":2}] Sun, 16 Feb 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMDeepSeek推理模型强化学习 https://www.hansking.cn/blog/2026/FlashAttention-zhu-yi-li-ji-suan-de-IO-you-hua https://www.hansking.cn/blog/2026/FlashAttention-zhu-yi-li-ji-suan-de-IO-you-hua [{"value":"FlashAttention：注意力计算的 IO 优化","url":"#flashattention注意力计算的-io-优化","depth":1},{"value":"一、标准注意力为什么占内存","url":"#一标准注意力为什么占内存","depth":2},{"value":"二、GPU 有很多层存储","url":"#二gpu-有很多层存储","depth":2},{"value":"三、分块计算怎么保持正确结果","url":"#三分块计算怎么保持正确结果","depth":2},{"value":"四、因果 Mask 和变长序列","url":"#四因果-mask-和变长序列","depth":2},{"value":"五、它优化的是 IO，不是把复杂度变成线性","url":"#五它优化的是-io不是把复杂度变成线性","depth":2},{"value":"六、为什么实际速度可能没有论文那么快","url":"#六为什么实际速度可能没有论文那么快","depth":2},{"value":"七、使用时先确认版本和回退路径","url":"#七使用时先确认版本和回退路径","depth":2},{"value":"八、和量化、KV Cache 一起看","url":"#八和量化kv-cache-一起看","depth":2},{"value":"九、从工程角度如何判断值得使用","url":"#九从工程角度如何判断值得使用","depth":2},{"value":"总结：性能优化常常是少搬数据","url":"#总结性能优化常常是少搬数据","depth":2}] Sun, 09 Nov 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMFlashAttentionGPU性能优化 https://www.hansking.cn/blog/2026/GRPO-ru-men-qun-ti-xiang-dui-ce-le-you-hua https://www.hansking.cn/blog/2026/GRPO-ru-men-qun-ti-xiang-dui-ce-le-you-hua [{"value":"GRPO 入门：群体相对策略优化","url":"#grpo-入门群体相对策略优化","depth":1},{"value":"一、先回顾 PPO 在做什么","url":"#一先回顾-ppo-在做什么","depth":2},{"value":"二、GRPO 的核心直觉","url":"#二grpo-的核心直觉","depth":2},{"value":"三、目标函数在优化什么","url":"#三目标函数在优化什么","depth":2},{"value":"四、奖励函数决定模型究竟学会什么","url":"#四奖励函数决定模型究竟学会什么","depth":2},{"value":"五、组怎么采样，比想象中更重要","url":"#五组怎么采样比想象中更重要","depth":2},{"value":"六、长度偏差是一个隐蔽的坑","url":"#六长度偏差是一个隐蔽的坑","depth":2},{"value":"七、一个简化训练流程","url":"#七一个简化训练流程","depth":2},{"value":"八、训练中应该盯哪些指标","url":"#八训练中应该盯哪些指标","depth":2},{"value":"九、GRPO 和 SFT、PPO 如何配合","url":"#九grpo-和-sftppo-如何配合","depth":2},{"value":"十、我的总结：比较可以替代一部分“绝对判断”","url":"#十我的总结比较可以替代一部分绝对判断","depth":2}] Sun, 02 Mar 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLM推理模型GRPO强化学习 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/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/Human-in-the-loop-gao-feng-xian-dong-zuo-qian-de-shen-pi-she-ji https://www.hansking.cn/blog/2026/Human-in-the-loop-gao-feng-xian-dong-zuo-qian-de-shen-pi-she-ji [{"value":"Human-in-the-loop：高风险动作前的审批设计","url":"#human-in-the-loop高风险动作前的审批设计","depth":1},{"value":"一、先给动作做风险分级","url":"#一先给动作做风险分级","depth":2},{"value":"二、确认前先生成可审查的计划","url":"#二确认前先生成可审查的计划","depth":2},{"value":"三、确认必须绑定具体版本","url":"#三确认必须绑定具体版本","depth":2},{"value":"四、确认界面要讲人话","url":"#四确认界面要讲人话","depth":2},{"value":"五、谁来确认是权限问题","url":"#五谁来确认是权限问题","depth":2},{"value":"六、审批状态要进入任务状态机","url":"#六审批状态要进入任务状态机","depth":2},{"value":"七、等待期间也要处理取消和变化","url":"#七等待期间也要处理取消和变化","depth":2},{"value":"八、审批和自动化要有分层策略","url":"#八审批和自动化要有分层策略","depth":2},{"value":"九、审计记录要能还原决定","url":"#九审计记录要能还原决定","depth":2},{"value":"十、用故障和误操作测试审批流程","url":"#十用故障和误操作测试审批流程","depth":2},{"value":"总结：把决定权交给人，也要把信息交给人","url":"#总结把决定权交给人也要把信息交给人","depth":2}] Sun, 03 Aug 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgentHuman-in-the-loop安全设计 https://www.hansking.cn/blog/2026/KVCacheLLM-tui-li-wei-shen-me-hui-bei-xian-cun-ka-zhu https://www.hansking.cn/blog/2026/KVCacheLLM-tui-li-wei-shen-me-hui-bei-xian-cun-ka-zhu [{"value":"KV Cache：LLM 推理为什么会被显存卡住","url":"#kv-cachellm-推理为什么会被显存卡住","depth":1},{"value":"一、先理解自回归生成","url":"#一先理解自回归生成","depth":2},{"value":"二、KV Cache 为什么会占很多显存","url":"#二kv-cache-为什么会占很多显存","depth":2},{"value":"三、Prefill 和 Decode 对 Cache 的使用不同","url":"#三prefill-和-decode-对-cache-的使用不同","depth":2},{"value":"四、为什么连续批处理需要更聪明的 Cache","url":"#四为什么连续批处理需要更聪明的-cache","depth":2},{"value":"五、前缀缓存可以复用，但必须确认边界","url":"#五前缀缓存可以复用但必须确认边界","depth":2},{"value":"六、Cache 量化和压缩要用质量换算","url":"#六cache-量化和压缩要用质量换算","depth":2},{"value":"七、Cache 管理要考虑取消、抢占和恢复","url":"#七cache-管理要考虑取消抢占和恢复","depth":2},{"value":"八、上下文窗口不是无限记忆","url":"#八上下文窗口不是无限记忆","depth":2},{"value":"九、如何压测 KV Cache","url":"#九如何压测-kv-cache","depth":2},{"value":"总结：缓存是速度和空间的交换","url":"#总结缓存是速度和空间的交换","depth":2}] Sun, 28 Sep 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMKV CacheGPU推理优化 https://www.hansking.cn/blog/2026/LLM-an-quan-PromptInjection-yu-yue-quan-gong-ju-diao-yong https://www.hansking.cn/blog/2026/LLM-an-quan-PromptInjection-yu-yue-quan-gong-ju-diao-yong [{"value":"LLM 安全：Prompt Injection 与越权工具调用","url":"#llm-安全prompt-injection-与越权工具调用","depth":1},{"value":"一、直接注入和间接注入","url":"#一直接注入和间接注入","depth":2},{"value":"二、为什么模型很难天然区分指令和数据","url":"#二为什么模型很难天然区分指令和数据","depth":2},{"value":"三、先画清数据流和信任边界","url":"#三先画清数据流和信任边界","depth":2},{"value":"四、最小权限是最有效的防线","url":"#四最小权限是最有效的防线","depth":2},{"value":"五、工具调用要有二次确认","url":"#五工具调用要有二次确认","depth":2},{"value":"六、RAG 的安全重点是检索前后两道门","url":"#六rag-的安全重点是检索前后两道门","depth":2},{"value":"七、浏览器和代码执行是高风险组合","url":"#七浏览器和代码执行是高风险组合","depth":2},{"value":"八、输出过滤挡不住已经发生的副作用","url":"#八输出过滤挡不住已经发生的副作用","depth":2},{"value":"九、检测注入要看行为，不只看文本","url":"#九检测注入要看行为不只看文本","depth":2},{"value":"十、用攻击样本做持续测试","url":"#十用攻击样本做持续测试","depth":2},{"value":"十一、日志和凭证管理决定事故能否收敛","url":"#十一日志和凭证管理决定事故能否收敛","depth":2},{"value":"十二、我的总结：不要试图把模型变成防火墙","url":"#十二我的总结不要试图把模型变成防火墙","depth":2}] Sun, 13 Oct 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLM安全Prompt InjectionAgent https://www.hansking.cn/blog/2026/LLM-wang-guan-she-ji-tong-yi-jian-quan-xian-liu-chong-shi-yu-shen-ji https://www.hansking.cn/blog/2026/LLM-wang-guan-she-ji-tong-yi-jian-quan-xian-liu-chong-shi-yu-shen-ji [{"value":"LLM 网关设计：统一鉴权、限流、重试与审计","url":"#llm-网关设计统一鉴权限流重试与审计","depth":1},{"value":"一、网关到底负责什么","url":"#一网关到底负责什么","depth":2},{"value":"二、先定义统一请求协议","url":"#二先定义统一请求协议","depth":2},{"value":"三、鉴权和授权要分开","url":"#三鉴权和授权要分开","depth":2},{"value":"四、模型路由要可解释","url":"#四模型路由要可解释","depth":2},{"value":"五、限流要同时限制请求和资源","url":"#五限流要同时限制请求和资源","depth":2},{"value":"六、重试要防止重复副作用","url":"#六重试要防止重复副作用","depth":2},{"value":"七、熔断和降级要保留安全边界","url":"#七熔断和降级要保留安全边界","depth":2},{"value":"八、审计日志要足够排查，但不要保存一切","url":"#八审计日志要足够排查但不要保存一切","depth":2},{"value":"九、成本和质量要从同一条链路统计","url":"#九成本和质量要从同一条链路统计","depth":2},{"value":"总结：网关不是代理，而是规则的共同入口","url":"#总结网关不是代理而是规则的共同入口","depth":2}] Tue, 13 Jan 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLMAPI Gateway限流可观测性 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/LLM-as-a-Judge-zi-dong-ping-shen-de-pian-cha-yu-jiao-zhun https://www.hansking.cn/blog/2026/LLM-as-a-Judge-zi-dong-ping-shen-de-pian-cha-yu-jiao-zhun [{"value":"LLM-as-a-Judge：自动评审的偏差与校准","url":"#llm-as-a-judge自动评审的偏差与校准","depth":1},{"value":"一、先定义裁判到底要判断什么","url":"#一先定义裁判到底要判断什么","depth":2},{"value":"二、给裁判一份明确的评分标准","url":"#二给裁判一份明确的评分标准","depth":2},{"value":"三、裁判常见的几种偏差","url":"#三裁判常见的几种偏差","depth":2},{"value":"1. 长度偏差","url":"#1-长度偏差","depth":3},{"value":"2. 位置偏差","url":"#2-位置偏差","depth":3},{"value":"3. 自我偏好","url":"#3-自我偏好","depth":3},{"value":"4. 迎合语气","url":"#4-迎合语气","depth":3},{"value":"5. 参考答案锚定","url":"#5-参考答案锚定","depth":3},{"value":"四、先做盲测和一致性检查","url":"#四先做盲测和一致性检查","depth":2},{"value":"五、人工标注用来校准，不是只用来盖章","url":"#五人工标注用来校准不是只用来盖章","depth":2},{"value":"六、让裁判输出证据和解释","url":"#六让裁判输出证据和解释","depth":2},{"value":"七、评审 Prompt 也要版本化","url":"#七评审-prompt-也要版本化","depth":2},{"value":"八、什么时候不要使用 LLM-as-a-Judge","url":"#八什么时候不要使用-llm-as-a-judge","depth":2},{"value":"九、把分数连接到行动","url":"#九把分数连接到行动","depth":2},{"value":"总结：自动裁判要先证明自己值得被信任","url":"#总结自动裁判要先证明自己值得被信任","depth":2}] Sat, 31 Jan 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLMLLM-as-a-Judge模型评测数据质量 https://www.hansking.cn/blog/2026/Llama3-kai-yuan-mo-xing-bu-shu-yu-API-feng-zhuang https://www.hansking.cn/blog/2026/Llama3-kai-yuan-mo-xing-bu-shu-yu-API-feng-zhuang [{"value":"Llama 3 开源模型部署与 API 封装","url":"#llama-3-开源模型部署与-api-封装","depth":1},{"value":"一、开源模型不只是一个权重文件","url":"#一开源模型不只是一个权重文件","depth":2},{"value":"二、先算显存和上下文预算","url":"#二先算显存和上下文预算","depth":2},{"value":"三、Tokenizer 和 Chat Template 必须匹配","url":"#三tokenizer-和-chat-template-必须匹配","depth":2},{"value":"四、选择推理运行时","url":"#四选择推理运行时","depth":2},{"value":"五、先实现一个统一的内部接口","url":"#五先实现一个统一的内部接口","depth":2},{"value":"六、封装 OpenAI 兼容 API","url":"#六封装-openai-兼容-api","depth":2},{"value":"七、生成参数不是越随机越好","url":"#七生成参数不是越随机越好","depth":2},{"value":"八、并发：单用户快不代表多人可用","url":"#八并发单用户快不代表多人可用","depth":2},{"value":"九、鉴权、限流和数据边界不能省","url":"#九鉴权限流和数据边界不能省","depth":2},{"value":"十、健康检查和优雅退出","url":"#十健康检查和优雅退出","depth":2},{"value":"十一、用基准测试决定本地部署是否值得","url":"#十一用基准测试决定本地部署是否值得","depth":2},{"value":"十二、我的总结：开源权重是起点，不是产品","url":"#十二我的总结开源权重是起点不是产品","depth":2}] Sun, 12 May 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLMLlama 3本地部署API https://www.hansking.cn/blog/2026/LoRA-yuan-li-di-zhi-ju-zhen-wei-shen-me-neng-wei-tiao-da-mo-xing https://www.hansking.cn/blog/2026/LoRA-yuan-li-di-zhi-ju-zhen-wei-shen-me-neng-wei-tiao-da-mo-xing [{"value":"LoRA 原理：低秩矩阵为什么能微调大模型","url":"#lora-原理低秩矩阵为什么能微调大模型","depth":1},{"value":"一、全量微调到底在更新什么","url":"#一全量微调到底在更新什么","depth":2},{"value":"二、低秩为什么有机会表达任务变化","url":"#二低秩为什么有机会表达任务变化","depth":2},{"value":"三、参数量到底省了多少","url":"#三参数量到底省了多少","depth":2},{"value":"四、A 和 B 为什么常见地这样初始化","url":"#四a-和-b-为什么常见地这样初始化","depth":2},{"value":"五、LoRA 通常注入哪些层","url":"#五lora-通常注入哪些层","depth":2},{"value":"六、rank 不是越大越好","url":"#六rank-不是越大越好","depth":2},{"value":"七、LoRA 和灾难性遗忘","url":"#七lora-和灾难性遗忘","depth":2},{"value":"八、LoRA 的训练流程","url":"#八lora-的训练流程","depth":2},{"value":"九、学习率和数据量要一起看","url":"#九学习率和数据量要一起看","depth":2},{"value":"十、适配器合并和独立加载","url":"#十适配器合并和独立加载","depth":2},{"value":"十一、LoRA 的能力边界","url":"#十一lora-的能力边界","depth":2},{"value":"十二、如何评测 LoRA 是否真的有效","url":"#十二如何评测-lora-是否真的有效","depth":2},{"value":"十三、我的总结：用小更新保存大能力","url":"#十三我的总结用小更新保存大能力","depth":2}] Sun, 23 Jun 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLMLoRA微调Transformer https://www.hansking.cn/blog/2026/MCPServer-shi-zhan-wei-bo-ke-bao-lu-sou-suo-gong-ju https://www.hansking.cn/blog/2026/MCPServer-shi-zhan-wei-bo-ke-bao-lu-sou-suo-gong-ju [{"value":"MCP Server 实战：为博客暴露搜索工具","url":"#mcp-server-实战为博客暴露搜索工具","depth":1},{"value":"一、先定义工具能做什么","url":"#一先定义工具能做什么","depth":2},{"value":"二、能力声明要让客户端知道边界","url":"#二能力声明要让客户端知道边界","depth":2},{"value":"三、参数 Schema 是第一道边界","url":"#三参数-schema-是第一道边界","depth":2},{"value":"四、Server 端必须重新确认身份和范围","url":"#四server-端必须重新确认身份和范围","depth":2},{"value":"五、搜索结果要带引用，不要只返回一段大文本","url":"#五搜索结果要带引用不要只返回一段大文本","depth":2},{"value":"六、错误返回要帮助 Agent 做正确的下一步","url":"#六错误返回要帮助-agent-做正确的下一步","depth":2},{"value":"七、MCP Server 也需要超时和限流","url":"#七mcp-server-也需要超时和限流","depth":2},{"value":"八、只读工具也要审计","url":"#八只读工具也要审计","depth":2},{"value":"九、和博客 Contentlayer 索引连接","url":"#九和博客-contentlayer-索引连接","depth":2},{"value":"十、测试要覆盖协议和真实行为","url":"#十测试要覆盖协议和真实行为","depth":2},{"value":"总结：暴露能力之前先设计边界","url":"#总结暴露能力之前先设计边界","depth":2}] Sun, 08 Jun 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMMCPAgentTypeScript https://www.hansking.cn/blog/2026/MCP-xie-yi-ru-men-tong-yi-mo-xing-yu-wai-bu-gong-ju-de-lian-jie https://www.hansking.cn/blog/2026/MCP-xie-yi-ru-men-tong-yi-mo-xing-yu-wai-bu-gong-ju-de-lian-jie [{"value":"MCP 协议入门：统一模型与外部工具的连接","url":"#mcp-协议入门统一模型与外部工具的连接","depth":1},{"value":"一、先记住三层：Host、Client、Server","url":"#一先记住三层hostclientserver","depth":2},{"value":"二、MCP 暴露的不是只有工具","url":"#二mcp-暴露的不是只有工具","depth":2},{"value":"1. Tools：让模型发起动作","url":"#1-tools让模型发起动作","depth":3},{"value":"2. Resources：让应用读取上下文","url":"#2-resources让应用读取上下文","depth":3},{"value":"3. Prompts：复用交互模板","url":"#3-prompts复用交互模板","depth":3},{"value":"三、一次工具调用到底发生了什么","url":"#三一次工具调用到底发生了什么","depth":2},{"value":"四、一个最小的 TypeScript Server 骨架","url":"#四一个最小的-typescript-server-骨架","depth":2},{"value":"五、传输方式不是小细节","url":"#五传输方式不是小细节","depth":2},{"value":"六、MCP 不会自动替你解决安全问题","url":"#六mcp-不会自动替你解决安全问题","depth":2},{"value":"七、我研究完 MCP 后留下的判断","url":"#七我研究完-mcp-后留下的判断","depth":2}] Sun, 25 May 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMMCPAgent协议 https://www.hansking.cn/blog/2026/Mixtral-yu-MoE-xi-shu-zhuan-jia-mo-xing-zen-me-sheng-ji-suan https://www.hansking.cn/blog/2026/Mixtral-yu-MoE-xi-shu-zhuan-jia-mo-xing-zen-me-sheng-ji-suan [{"value":"Mixtral 与 MoE：稀疏专家模型怎么省计算","url":"#mixtral-与-moe稀疏专家模型怎么省计算","depth":1},{"value":"一、密集模型和稀疏模型的区别","url":"#一密集模型和稀疏模型的区别","depth":2},{"value":"二、Mixtral 的核心结构直觉","url":"#二mixtral-的核心结构直觉","depth":2},{"value":"三、参数量为什么容易让人误判","url":"#三参数量为什么容易让人误判","depth":2},{"value":"四、为什么路由器不是一个简单分类器","url":"#四为什么路由器不是一个简单分类器","depth":2},{"value":"五、负载均衡和容量因子","url":"#五负载均衡和容量因子","depth":2},{"value":"六、稀疏计算节省了什么","url":"#六稀疏计算节省了什么","depth":2},{"value":"七、专家并行带来通信挑战","url":"#七专家并行带来通信挑战","depth":2},{"value":"八、Mixtral 让开源生态看见了什么","url":"#八mixtral-让开源生态看见了什么","depth":2},{"value":"九、MoE 的质量不是“专家数量越多越高”","url":"#九moe-的质量不是专家数量越多越高","depth":2},{"value":"十、部署时该怎么选","url":"#十部署时该怎么选","depth":2},{"value":"十一、我的总结：把容量和计算分开，是一次重要转向","url":"#十一我的总结把容量和计算分开是一次重要转向","depth":2}] Sun, 26 May 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLMMixtralMoE稀疏模型 https://www.hansking.cn/blog/2026/MoE-de-lu-you-fu-zai-jun-heng-yu-xun-lian-nan-dian https://www.hansking.cn/blog/2026/MoE-de-lu-you-fu-zai-jun-heng-yu-xun-lian-nan-dian [{"value":"MoE 的路由、负载均衡与训练难点","url":"#moe-的路由负载均衡与训练难点","depth":1},{"value":"一、普通 Transformer 为什么需要专家","url":"#一普通-transformer-为什么需要专家","depth":2},{"value":"二、路由器在做什么","url":"#二路由器在做什么","depth":2},{"value":"三、Top-k 为什么既有效又麻烦","url":"#三top-k-为什么既有效又麻烦","depth":2},{"value":"四、负载均衡是 MoE 的核心难题","url":"#四负载均衡是-moe-的核心难题","depth":2},{"value":"五、容量因子和 Token dropping","url":"#五容量因子和-token-dropping","depth":2},{"value":"六、专家为什么会学出不同能力","url":"#六专家为什么会学出不同能力","depth":2},{"value":"七、训练时的通信可能比计算更贵","url":"#七训练时的通信可能比计算更贵","depth":2},{"value":"八、训练稳定性需要多种信号共同判断","url":"#八训练稳定性需要多种信号共同判断","depth":2},{"value":"九、推理部署和训练不是同一笔账","url":"#九推理部署和训练不是同一笔账","depth":2},{"value":"十、如何判断 MoE 是否值得使用","url":"#十如何判断-moe-是否值得使用","depth":2},{"value":"十一、我的总结：MoE 是模型里的调度系统","url":"#十一我的总结moe-是模型里的调度系统","depth":2}] Sun, 09 Jun 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLMMoE稀疏专家深度学习 https://www.hansking.cn/blog/2026/ModelDistillation-shi-zhan-ba-da-mo-xing-neng-li-ya-suo-xia-lai https://www.hansking.cn/blog/2026/ModelDistillation-shi-zhan-ba-da-mo-xing-neng-li-ya-suo-xia-lai [{"value":"Model Distillation 实战：把大模型能力压缩下来","url":"#model-distillation-实战把大模型能力压缩下来","depth":1},{"value":"一、先定义学生模型的任务边界","url":"#一先定义学生模型的任务边界","depth":2},{"value":"二、让教师输出稳定，而不是只追求聪明","url":"#二让教师输出稳定而不是只追求聪明","depth":2},{"value":"三、数据清洗比训练更重要","url":"#三数据清洗比训练更重要","depth":2},{"value":"1. 先做结构校验","url":"#1-先做结构校验","depth":3},{"value":"2. 去重和去污染","url":"#2-去重和去污染","depth":3},{"value":"3. 教师一致性检查","url":"#3-教师一致性检查","depth":3},{"value":"四、两种蒸馏目标","url":"#四两种蒸馏目标","depth":2},{"value":"1. 硬标签蒸馏","url":"#1-硬标签蒸馏","depth":3},{"value":"2. 软目标蒸馏","url":"#2-软目标蒸馏","depth":3},{"value":"五、一个最小训练骨架","url":"#五一个最小训练骨架","depth":2},{"value":"六、评测蒸馏是否值得","url":"#六评测蒸馏是否值得","depth":2},{"value":"七、把不确定性交给大模型","url":"#七把不确定性交给大模型","depth":2},{"value":"总结：蒸馏是在复制能力，也是在复制责任","url":"#总结蒸馏是在复制能力也是在复制责任","depth":2}] Mon, 13 Apr 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLM知识蒸馏模型训练PyTorch https://www.hansking.cn/blog/2026/On-deviceAI-duan-ce-tui-li-de-xing-neng-yu-yin-si-quan-heng https://www.hansking.cn/blog/2026/On-deviceAI-duan-ce-tui-li-de-xing-neng-yu-yin-si-quan-heng [{"value":"On-device AI：端侧推理的性能与隐私权衡","url":"#on-device-ai端侧推理的性能与隐私权衡","depth":1},{"value":"一、先决定哪些任务值得放到设备端","url":"#一先决定哪些任务值得放到设备端","depth":2},{"value":"二、先做内存预算，再谈模型大小","url":"#二先做内存预算再谈模型大小","depth":2},{"value":"三、量化不是免费午餐","url":"#三量化不是免费午餐","depth":2},{"value":"四、推理优化的第一目标是减少等待","url":"#四推理优化的第一目标是减少等待","depth":2},{"value":"五、端侧隐私强，但不是天然安全","url":"#五端侧隐私强但不是天然安全","depth":2},{"value":"六、模型文件也需要保护","url":"#六模型文件也需要保护","depth":2},{"value":"七、端云路由需要可解释","url":"#七端云路由需要可解释","depth":2},{"value":"八、我建议用一张设备矩阵收尾","url":"#八我建议用一张设备矩阵收尾","depth":2},{"value":"总结：端侧 AI 的价值来自克制","url":"#总结端侧-ai-的价值来自克制","depth":2}] Wed, 22 Apr 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLMOn-device AI端侧推理性能优化 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/QLoRA-shi-zhan-zai-xiao-fei-ji-xian-ka-shang-wei-tiao-mo-xing https://www.hansking.cn/blog/2026/QLoRA-shi-zhan-zai-xiao-fei-ji-xian-ka-shang-wei-tiao-mo-xing [{"value":"QLoRA 实战：在消费级显卡上微调模型","url":"#qlora-实战在消费级显卡上微调模型","depth":1},{"value":"一、LoRA 先解决了什么","url":"#一lora-先解决了什么","depth":2},{"value":"二、QLoRA 又增加了什么","url":"#二qlora-又增加了什么","depth":2},{"value":"三、先估算显存，不要直接碰训练命令","url":"#三先估算显存不要直接碰训练命令","depth":2},{"value":"四、环境版本要成套匹配","url":"#四环境版本要成套匹配","depth":2},{"value":"五、数据准备决定微调是否值得","url":"#五数据准备决定微调是否值得","depth":2},{"value":"六、聊天模板和 loss mask 不能错","url":"#六聊天模板和-loss-mask-不能错","depth":2},{"value":"七、一个适合小显存实验的配置思路","url":"#七一个适合小显存实验的配置思路","depth":2},{"value":"八、显存不足时的排查顺序","url":"#八显存不足时的排查顺序","depth":2},{"value":"九、训练中看 loss，训练后看行为","url":"#九训练中看-loss训练后看行为","depth":2},{"value":"十、适配器保存和部署","url":"#十适配器保存和部署","depth":2},{"value":"十一、什么时候 QLoRA 不是好选择","url":"#十一什么时候-qlora-不是好选择","depth":2},{"value":"十二、我的总结：显存门槛降低，工程门槛没有消失","url":"#十二我的总结显存门槛降低工程门槛没有消失","depth":2}] Sun, 07 Jul 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLMQLoRALoRA微调 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/2026/ReActAgent-cong-sou-suo-gong-ju-kai-shi-shi-xian-zhi-neng-ti https://www.hansking.cn/blog/2026/ReActAgent-cong-sou-suo-gong-ju-kai-shi-shi-xian-zhi-neng-ti [{"value":"ReAct Agent：从搜索工具开始实现智能体","url":"#react-agent从搜索工具开始实现智能体","depth":1},{"value":"一、ReAct 到底解决了什么","url":"#一react-到底解决了什么","depth":2},{"value":"二、先定义一个足够窄的搜索工具","url":"#二先定义一个足够窄的搜索工具","depth":2},{"value":"三、把 Agent 状态写出来","url":"#三把-agent-状态写出来","depth":2},{"value":"四、最小循环应该长什么样","url":"#四最小循环应该长什么样","depth":2},{"value":"五、Prompt 应该规定角色边界，而不是写成魔法咒语","url":"#五prompt-应该规定角色边界而不是写成魔法咒语","depth":2},{"value":"六、观察结果时要防止上下文污染","url":"#六观察结果时要防止上下文污染","depth":2},{"value":"七、停止条件决定 Agent 是助手还是黑洞","url":"#七停止条件决定-agent-是助手还是黑洞","depth":2},{"value":"八、用轨迹评测 Agent，而不是只看最后一句话","url":"#八用轨迹评测-agent而不是只看最后一句话","depth":2},{"value":"九、从搜索 Agent 走向真实系统","url":"#九从搜索-agent-走向真实系统","depth":2},{"value":"十、我的总结：循环很简单，边界才是工程","url":"#十我的总结循环很简单边界才是工程","depth":2}] Sun, 13 Apr 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgentReActTypeScript https://www.hansking.cn/blog/2026/SFT-zhi-ling-wei-tiao-shu-ju-ge-shi-qing-xi-yu-xun-lian https://www.hansking.cn/blog/2026/SFT-zhi-ling-wei-tiao-shu-ju-ge-shi-qing-xi-yu-xun-lian [{"value":"SFT 指令微调：数据格式、清洗与训练","url":"#sft-指令微调数据格式清洗与训练","depth":1},{"value":"一、先判断问题是不是真的需要 SFT","url":"#一先判断问题是不是真的需要-sft","depth":2},{"value":"二、SFT 数据的基本形式","url":"#二sft-数据的基本形式","depth":2},{"value":"三、输入输出边界必须一致","url":"#三输入输出边界必须一致","depth":2},{"value":"四、清洗：模型会把脏数据学得很认真","url":"#四清洗模型会把脏数据学得很认真","depth":2},{"value":"去重","url":"#去重","depth":3},{"value":"冲突检查","url":"#冲突检查","depth":3},{"value":"敏感信息","url":"#敏感信息","depth":3},{"value":"模型生成数据","url":"#模型生成数据","depth":3},{"value":"五、数据划分要防止“换句话说的泄漏”","url":"#五数据划分要防止换句话说的泄漏","depth":2},{"value":"六、长度和截断要在训练前算清楚","url":"#六长度和截断要在训练前算清楚","depth":2},{"value":"七、loss 到底应该算在哪些 Token 上","url":"#七loss-到底应该算在哪些-token-上","depth":2},{"value":"八、一个简化的训练配置","url":"#八一个简化的训练配置","depth":2},{"value":"九、训练过程中看什么","url":"#九训练过程中看什么","depth":2},{"value":"十、训练完成后先做离线回归","url":"#十训练完成后先做离线回归","depth":2},{"value":"十一、部署时不要忘记训练模板","url":"#十一部署时不要忘记训练模板","depth":2},{"value":"十二、什么时候训练算成功","url":"#十二什么时候训练算成功","depth":2},{"value":"十三、我的总结：SFT 是数据工程，不只是训练工程","url":"#十三我的总结sft-是数据工程不只是训练工程","depth":2}] Sun, 21 Jul 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLMSFT微调数据工程 https://www.hansking.cn/blog/2026/SpeculativeDecoding-yong-xiao-mo-xing-jia-su-da-mo-xing https://www.hansking.cn/blog/2026/SpeculativeDecoding-yong-xiao-mo-xing-jia-su-da-mo-xing [{"value":"Speculative Decoding：用小模型加速大模型","url":"#speculative-decoding用小模型加速大模型","depth":1},{"value":"一、为什么自回归生成慢","url":"#一为什么自回归生成慢","depth":2},{"value":"二、基本流程是什么","url":"#二基本流程是什么","depth":2},{"value":"三、小模型不需要和大模型一样聪明","url":"#三小模型不需要和大模型一样聪明","depth":2},{"value":"四、速度收益取决于接受率","url":"#四速度收益取决于接受率","depth":2},{"value":"五、为什么结果可以保持目标模型分布","url":"#五为什么结果可以保持目标模型分布","depth":2},{"value":"六、KV Cache 管理决定能不能跑得快","url":"#六kv-cache-管理决定能不能跑得快","depth":2},{"value":"七、不同任务的收益不同","url":"#七不同任务的收益不同","depth":2},{"value":"八、工程上常见的失败原因","url":"#八工程上常见的失败原因","depth":2},{"value":"1. 草稿模型太慢","url":"#1-草稿模型太慢","depth":3},{"value":"2. Tokenizer 不兼容","url":"#2-tokenizer-不兼容","depth":3},{"value":"3. 采样参数不一致","url":"#3-采样参数不一致","depth":3},{"value":"4. 并发下显存不足","url":"#4-并发下显存不足","depth":3},{"value":"5. 只测平均值","url":"#5-只测平均值","depth":3},{"value":"九、上线前怎么评估","url":"#九上线前怎么评估","depth":2},{"value":"总结：让小模型猜，让大模型负责","url":"#总结让小模型猜让大模型负责","depth":2}] Sun, 26 Oct 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMSpeculative Decoding推理优化性能优化 https://www.hansking.cn/blog/2026/Test-timeCompute-yong-geng-duo-tui-li-huan-geng-gao-zheng-que-l https://www.hansking.cn/blog/2026/Test-timeCompute-yong-geng-duo-tui-li-huan-geng-gao-zheng-que-l [{"value":"Test-time Compute：用更多推理换更高正确率","url":"#test-time-compute用更多推理换更高正确率","depth":1},{"value":"一、训练时计算和推理时计算有什么不同","url":"#一训练时计算和推理时计算有什么不同","depth":2},{"value":"二、最简单的方式：Best-of-N","url":"#二最简单的方式best-of-n","depth":2},{"value":"三、自洽性不是简单投票","url":"#三自洽性不是简单投票","depth":2},{"value":"四、搜索与分支：把答案当成一个小型搜索问题","url":"#四搜索与分支把答案当成一个小型搜索问题","depth":2},{"value":"五、动态预算比固定长思考更合理","url":"#五动态预算比固定长思考更合理","depth":2},{"value":"六、推理时计算的成本怎么算","url":"#六推理时计算的成本怎么算","depth":2},{"value":"七、质量收益必须通过对照实验证明","url":"#七质量收益必须通过对照实验证明","depth":2},{"value":"八、什么时候不值得使用更多计算","url":"#八什么时候不值得使用更多计算","depth":2},{"value":"九、生产环境必须设置护栏","url":"#九生产环境必须设置护栏","depth":2},{"value":"十、我的总结：把计算当成预算，而不是信仰","url":"#十我的总结把计算当成预算而不是信仰","depth":2}] Sun, 02 Feb 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLM推理模型Test-time Compute性能优化 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/2026/vLLM-tui-li-fu-wu-PagedAttention-yu-tun-tu-you-hua https://www.hansking.cn/blog/2026/vLLM-tui-li-fu-wu-PagedAttention-yu-tun-tu-you-hua [{"value":"vLLM 推理服务：PagedAttention 与吞吐优化","url":"#vllm-推理服务pagedattention-与吞吐优化","depth":1},{"value":"一、先用最小服务确认链路","url":"#一先用最小服务确认链路","depth":2},{"value":"二、PagedAttention 解决什么问题","url":"#二pagedattention-解决什么问题","depth":2},{"value":"三、动态批处理让请求不必互相等待","url":"#三动态批处理让请求不必互相等待","depth":2},{"value":"四、max-model-len 不是越大越好","url":"#四max-model-len-不是越大越好","depth":2},{"value":"五、流式输出要正确处理断开和取消","url":"#五流式输出要正确处理断开和取消","depth":2},{"value":"六、请求调度要考虑公平性","url":"#六请求调度要考虑公平性","depth":2},{"value":"七、监控不能只有 GPU 利用率","url":"#七监控不能只有-gpu-利用率","depth":2},{"value":"八、压测要使用真实长度分布","url":"#八压测要使用真实长度分布","depth":2},{"value":"九、常见优化顺序","url":"#九常见优化顺序","depth":2},{"value":"十、发布和回滚要一起设计","url":"#十发布和回滚要一起设计","depth":2},{"value":"总结：推理服务的核心是调度资源","url":"#总结推理服务的核心是调度资源","depth":2}] Sun, 14 Sep 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMvLLMPagedAttention推理服务 https://www.hansking.cn/blog/2026cong-mo-xing-xuan-xing-dao-neng-li-lu-you-jian-li-zi-ji-de-ModelGateway https://www.hansking.cn/blog/2026cong-mo-xing-xuan-xing-dao-neng-li-lu-you-jian-li-zi-ji-de-ModelGateway [{"value":"从模型选型到能力路由：建立自己的 Model Gateway","url":"#从模型选型到能力路由建立自己的-model-gateway","depth":1},{"value":"一、先建立模型注册表","url":"#一先建立模型注册表","depth":2},{"value":"二、先描述任务，再选择模型","url":"#二先描述任务再选择模型","depth":2},{"value":"三、路由是过滤、评分和降级","url":"#三路由是过滤评分和降级","depth":2},{"value":"四、评分函数要能解释","url":"#四评分函数要能解释","depth":2},{"value":"五、能力路由比静态配置更灵活","url":"#五能力路由比静态配置更灵活","depth":2},{"value":"六、统一适配供应商差异","url":"#六统一适配供应商差异","depth":2},{"value":"七、故障时要有健康评分和熔断","url":"#七故障时要有健康评分和熔断","depth":2},{"value":"八、灰度和回滚是路由系统的基本能力","url":"#八灰度和回滚是路由系统的基本能力","depth":2},{"value":"九、模型选型必须回到自己的基准集","url":"#九模型选型必须回到自己的基准集","depth":2},{"value":"总结：模型选择最后是系统选择","url":"#总结模型选择最后是系统选择","depth":2}] Sun, 04 Jan 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLMModel Gateway模型选型系统架构 https://www.hansking.cn/blog/2026cong-ling-da-jian-LLM-ying-yong-CI-ce-shi-ping-ce-yu-fa-bu-men-jin https://www.hansking.cn/blog/2026cong-ling-da-jian-LLM-ying-yong-CI-ce-shi-ping-ce-yu-fa-bu-men-jin [{"value":"从零搭建 LLM 应用 CI：测试、评测与发布门禁","url":"#从零搭建-llm-应用-ci测试评测与发布门禁","depth":1},{"value":"一、先把测试分成四层","url":"#一先把测试分成四层","depth":2},{"value":"二、代码层先测试确定性行为","url":"#二代码层先测试确定性行为","depth":2},{"value":"三、链路层用可控替身测试 Agent","url":"#三链路层用可控替身测试-agent","depth":2},{"value":"四、黄金数据集是 LLM 应用的回归测试","url":"#四黄金数据集是-llm-应用的回归测试","depth":2},{"value":"五、不要用精确字符串比较所有模型答案","url":"#五不要用精确字符串比较所有模型答案","depth":2},{"value":"六、把 Prompt 和模型版本当成代码","url":"#六把-prompt-和模型版本当成代码","depth":2},{"value":"七、CI 流水线可以这样分段","url":"#七ci-流水线可以这样分段","depth":2},{"value":"八、发布门禁要防止回归","url":"#八发布门禁要防止回归","depth":2},{"value":"九、让 CI 本身也可靠","url":"#九让-ci-本身也可靠","depth":2},{"value":"总结：绿色流水线应该代表值得信任","url":"#总结绿色流水线应该代表值得信任","depth":2}] Fri, 03 Jul 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLMCI/CD自动化测试评测 https://www.hansking.cn/blog/2026pian-hao-you-hua-cong-RLHF-dao-DPO https://www.hansking.cn/blog/2026pian-hao-you-hua-cong-RLHF-dao-DPO [{"value":"偏好优化：从 RLHF 到 DPO","url":"#偏好优化从-rlhf-到-dpo","depth":1},{"value":"一、预训练和 SFT 解决了什么","url":"#一预训练和-sft-解决了什么","depth":2},{"value":"二、偏好数据把绝对打分改成相对比较","url":"#二偏好数据把绝对打分改成相对比较","depth":2},{"value":"三、奖励模型：把偏好变成一个分数","url":"#三奖励模型把偏好变成一个分数","depth":2},{"value":"四、PPO 为什么被用于 RLHF","url":"#四ppo-为什么被用于-rlhf","depth":2},{"value":"五、DPO 的直觉：直接让模型偏爱 chosen","url":"#五dpo-的直觉直接让模型偏爱-chosen","depth":2},{"value":"六、三种路线放在一起看","url":"#六三种路线放在一起看","depth":2},{"value":"七、为什么数据质量比算法名字更重要","url":"#七为什么数据质量比算法名字更重要","depth":2},{"value":"偏好差异不清楚","url":"#偏好差异不清楚","depth":3},{"value":"chosen 只是更长","url":"#chosen-只是更长","depth":3},{"value":"rejected 过于糟糕","url":"#rejected-过于糟糕","depth":3},{"value":"任务标准不一致","url":"#任务标准不一致","depth":3},{"value":"八、偏好优化会带来哪些副作用","url":"#八偏好优化会带来哪些副作用","depth":2},{"value":"九、KL 约束为什么重要","url":"#九kl-约束为什么重要","depth":2},{"value":"十、离线偏好和在线反馈如何结合","url":"#十离线偏好和在线反馈如何结合","depth":2},{"value":"十一、什么时候选 RLHF，什么时候选 DPO","url":"#十一什么时候选-rlhf什么时候选-dpo","depth":2},{"value":"十二、我的总结：对齐是持续校准，不是一次训练","url":"#十二我的总结对齐是持续校准不是一次训练","depth":2}] Sun, 04 Aug 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLMRLHFDPO对齐 https://www.hansking.cn/blog/2026fu-xian-lun-wen-de-gong-cheng-fang-fa-huan-jing-shu-ju-zhi-biao-yu-wu-cha https://www.hansking.cn/blog/2026fu-xian-lun-wen-de-gong-cheng-fang-fa-huan-jing-shu-ju-zhi-biao-yu-wu-cha [{"value":"复现论文的工程方法：环境、数据、指标与误差","url":"#复现论文的工程方法环境数据指标与误差","depth":1},{"value":"一、先定义你要复现到什么程度","url":"#一先定义你要复现到什么程度","depth":2},{"value":"二、把环境变成可记录的实验输入","url":"#二把环境变成可记录的实验输入","depth":2},{"value":"三、核对数据：数量对上不代表内容对上","url":"#三核对数据数量对上不代表内容对上","depth":2},{"value":"四、重新实现指标之前先读清定义","url":"#四重新实现指标之前先读清定义","depth":2},{"value":"五、从最小实验开始逐步放大","url":"#五从最小实验开始逐步放大","depth":2},{"value":"六、随机性要被测量，而不是被假装消除","url":"#六随机性要被测量而不是被假装消除","depth":2},{"value":"七、误差分析要回答“为什么不一样”","url":"#七误差分析要回答为什么不一样","depth":2},{"value":"八、复现报告要诚实区分事实和推测","url":"#八复现报告要诚实区分事实和推测","depth":2},{"value":"九、把复现变成个人能力，而不是一次性作业","url":"#九把复现变成个人能力而不是一次性作业","depth":2},{"value":"总结：复现不是复制数字，而是重建证据","url":"#总结复现不是复制数字而是重建证据","depth":2}] Tue, 21 Jul 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLM论文复现实验工程机器学习 https://www.hansking.cn/blog/2026duo-zhi-neng-ti-xie-zuo-shen-me-shi-hou-zhi-de-chai-cheng-duo-ge-Agent https://www.hansking.cn/blog/2026duo-zhi-neng-ti-xie-zuo-shen-me-shi-hou-zhi-de-chai-cheng-duo-ge-Agent [{"value":"多智能体协作：什么时候值得拆成多个 Agent","url":"#多智能体协作什么时候值得拆成多个-agent","depth":1},{"value":"一、先问一个不太讨喜的问题：单 Agent 真的不够吗","url":"#一先问一个不太讨喜的问题单-agent-真的不够吗","depth":2},{"value":"二、什么情况下拆分确实有价值","url":"#二什么情况下拆分确实有价值","depth":2},{"value":"1. 任务有天然的专业边界","url":"#1-任务有天然的专业边界","depth":3},{"value":"2. 不同环节需要不同权限","url":"#2-不同环节需要不同权限","depth":3},{"value":"3. 某些工作可以并行","url":"#3-某些工作可以并行","depth":3},{"value":"4. 需要相互质疑，而不是只需要多写一遍","url":"#4-需要相互质疑而不是只需要多写一遍","depth":3},{"value":"三、四种常见协作模式","url":"#三四种常见协作模式","depth":2},{"value":"顺序流水线：一个接一个传递结果","url":"#顺序流水线一个接一个传递结果","depth":3},{"value":"并行专家：多个 Agent 同时给出独立意见","url":"#并行专家多个-agent-同时给出独立意见","depth":3},{"value":"监督者—执行者：负责规划与执行分离","url":"#监督者执行者负责规划与执行分离","depth":3},{"value":"辩论与裁决：不同假设之间互相挑战","url":"#辩论与裁决不同假设之间互相挑战","depth":3},{"value":"四、协作的核心不是聊天，而是共享状态","url":"#四协作的核心不是聊天而是共享状态","depth":2},{"value":"五、消息协议要小，结果要能验证","url":"#五消息协议要小结果要能验证","depth":2},{"value":"六、失败恢复比角色数量更值得设计","url":"#六失败恢复比角色数量更值得设计","depth":2},{"value":"七、如何判断拆分没有成功","url":"#七如何判断拆分没有成功","depth":2},{"value":"八、我的结论：把 Agent 当成有边界的同事","url":"#八我的结论把-agent-当成有边界的同事","depth":2}] Sun, 11 May 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgent多智能体架构 https://www.hansking.cn/blog/2026duo-mo-xing-lu-you-an-zhi-liang-yan-chi-he-jia-ge-xuan-ze-mo-xing https://www.hansking.cn/blog/2026duo-mo-xing-lu-you-an-zhi-liang-yan-chi-he-jia-ge-xuan-ze-mo-xing [{"value":"多模型路由：按质量、延迟和价格选择模型","url":"#多模型路由按质量延迟和价格选择模型","depth":1},{"value":"一、先把模型调用从业务代码里拿出来","url":"#一先把模型调用从业务代码里拿出来","depth":2},{"value":"二、先定义“什么叫合适”","url":"#二先定义什么叫合适","depth":2},{"value":"三、路由策略从简单规则开始","url":"#三路由策略从简单规则开始","depth":2},{"value":"四、加权评分适合处理多个目标","url":"#四加权评分适合处理多个目标","depth":2},{"value":"五、故障时的降级不是“随便换一个模型”","url":"#五故障时的降级不是随便换一个模型","depth":2},{"value":"六、熔断器保护整个服务","url":"#六熔断器保护整个服务","depth":2},{"value":"七、质量路由需要真实评测数据","url":"#七质量路由需要真实评测数据","depth":2},{"value":"八、成本控制要看“每个成功任务”","url":"#八成本控制要看每个成功任务","depth":2},{"value":"九、数据边界应该在路由前执行","url":"#九数据边界应该在路由前执行","depth":2},{"value":"十、流式响应让路由更难一点","url":"#十流式响应让路由更难一点","depth":2},{"value":"十一、不要过早训练一个“智能路由器”","url":"#十一不要过早训练一个智能路由器","depth":2},{"value":"十二、我的总结：路由系统是在管理取舍","url":"#十二我的总结路由系统是在管理取舍","depth":2}] Sun, 24 Nov 2024 10:00:00 GMT admin@hansking.cn (HANS) AILLM模型路由架构成本优化 https://www.hansking.cn/blog/2026duo-mo-tai-Agent-rang-mo-xing-du-tu-du-biao-du-ping-mu https://www.hansking.cn/blog/2026duo-mo-tai-Agent-rang-mo-xing-du-tu-du-biao-du-ping-mu [{"value":"多模态 Agent：让模型读图、读表、读屏幕","url":"#多模态-agent让模型读图读表读屏幕","depth":1},{"value":"一、先区分三类多模态任务","url":"#一先区分三类多模态任务","depth":2},{"value":"1. 视觉问答","url":"#1-视觉问答","depth":3},{"value":"2. 视觉结构化","url":"#2-视觉结构化","depth":3},{"value":"3. 视觉行动","url":"#3-视觉行动","depth":3},{"value":"二、图像预处理决定模型能不能看清","url":"#二图像预处理决定模型能不能看清","depth":2},{"value":"三、表格不能只当成一张图片","url":"#三表格不能只当成一张图片","depth":2},{"value":"四、结构化输出要有证据和校验","url":"#四结构化输出要有证据和校验","depth":2},{"value":"五、屏幕理解的关键是状态，不是坐标","url":"#五屏幕理解的关键是状态不是坐标","depth":2},{"value":"六、Agent 要知道什么时候自己不确定","url":"#六agent-要知道什么时候自己不确定","depth":2},{"value":"七、评测多模态 Agent 要测整条链路","url":"#七评测多模态-agent-要测整条链路","depth":2},{"value":"八、我最后留下的工程原则","url":"#八我最后留下的工程原则","depth":2}] Tue, 19 May 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLM多模态AgentComputer Usetop 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/2026da-mo-xing-fa-zhan-shi-yu-100-pian-ji-shu-xue-xi-xie-zuo-ji-hua https://www.hansking.cn/blog/2026da-mo-xing-fa-zhan-shi-yu-100-pian-ji-shu-xue-xi-xie-zuo-ji-hua [{"value":"2023—2026 大模型发展史：100篇技术与学习教程写作计划","url":"#20232026-大模型发展史100篇技术与学习教程写作计划","depth":1},{"value":"2023：从 ChatGPT 到 LLM 应用原型","url":"#2023从-chatgpt-到-llm-应用原型","depth":2},{"value":"2024：多模态、长上下文与应用工程化","url":"#2024多模态长上下文与应用工程化","depth":2},{"value":"2025：推理模型、智能体与推理基础设施","url":"#2025推理模型智能体与推理基础设施","depth":2},{"value":"2026：当下值得学习的通用能力","url":"#2026当下值得学习的通用能力","depth":2},{"value":"推荐的单篇文章模板","url":"#推荐的单篇文章模板","depth":2}] Sat, 09 Sep 2023 09:00:00 GMT admin@hansking.cn (HANS) AILLM学习路线写作计划 https://www.hansking.cn/blog/2026da-mo-xing-fa-zhan-shi-zong-fu-pan-cong-sheng-cheng-wen-ben-dao-ke-zhi-xing-xi-tong https://www.hansking.cn/blog/2026da-mo-xing-fa-zhan-shi-zong-fu-pan-cong-sheng-cheng-wen-ben-dao-ke-zhi-xing-xi-tong [{"value":"大模型发展史总复盘：从生成文本到可执行系统","url":"#大模型发展史总复盘从生成文本到可执行系统","depth":1},{"value":"一、2023：模型走出研究论文，进入开发者工具箱","url":"#一2023模型走出研究论文进入开发者工具箱","depth":2},{"value":"二、2024：从“调用模型”走向“组织上下文”","url":"#二2024从调用模型走向组织上下文","depth":2},{"value":"三、2025：模型开始规划、推理和调用工具","url":"#三2025模型开始规划推理和调用工具","depth":2},{"value":"四、2026：真正的重点是系统能不能长期运行","url":"#四2026真正的重点是系统能不能长期运行","depth":2},{"value":"五、四条一直没有改变的主线","url":"#五四条一直没有改变的主线","depth":2},{"value":"1. 数据比想象中更重要","url":"#1-数据比想象中更重要","depth":3},{"value":"2. 评测比感觉更重要","url":"#2-评测比感觉更重要","depth":3},{"value":"3. 边界比能力更重要","url":"#3-边界比能力更重要","depth":3},{"value":"4. 系统比模型更重要","url":"#4-系统比模型更重要","depth":3},{"value":"六、我会怎样继续学习大模型","url":"#六我会怎样继续学习大模型","depth":2},{"value":"七、100 篇之后，真正值得留下的东西","url":"#七100-篇之后真正值得留下的东西","depth":2},{"value":"总结：从“会生成”到“值得托付”","url":"#总结从会生成到值得托付","depth":2}] Sat, 08 Aug 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLM技术总结Agent学习路线 https://www.hansking.cn/blog/2026shi-shi-yu-yin-Agent-yan-chi-yu-suan-yu-da-duan-chu-li https://www.hansking.cn/blog/2026shi-shi-yu-yin-Agent-yan-chi-yu-suan-yu-da-duan-chu-li [{"value":"实时语音 Agent：延迟预算与打断处理","url":"#实时语音-agent延迟预算与打断处理","depth":1},{"value":"一、先画出完整音频链路","url":"#一先画出完整音频链路","depth":2},{"value":"二、用延迟预算替代一句“要快一点”","url":"#二用延迟预算替代一句要快一点","depth":2},{"value":"三、VAD 决定系统什么时候接话","url":"#三vad-决定系统什么时候接话","depth":2},{"value":"四、ASR 的 partial 结果不能直接当最终事实","url":"#四asr-的-partial-结果不能直接当最终事实","depth":2},{"value":"五、LLM 和 TTS 要按可播放单元流式协作","url":"#五llm-和-tts-要按可播放单元流式协作","depth":2},{"value":"六、打断是核心能力，不是异常分支","url":"#六打断是核心能力不是异常分支","depth":2},{"value":"七、状态机比一堆布尔变量可靠","url":"#七状态机比一堆布尔变量可靠","depth":2},{"value":"八、实时语音也要有隐私边界","url":"#八实时语音也要有隐私边界","depth":2},{"value":"总结：快不是催出来的，是每一层都愿意让路","url":"#总结快不是催出来的是每一层都愿意让路","depth":2}] Thu, 28 May 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLM语音交互AgentWebRTC https://www.hansking.cn/blog/2026xiao-mo-xing-lu-xian-zheng-liu-jian-zhi-yu-ren-wu-zhuan-yong-mo-xing https://www.hansking.cn/blog/2026xiao-mo-xing-lu-xian-zheng-liu-jian-zhi-yu-ren-wu-zhuan-yong-mo-xing [{"value":"小模型路线：蒸馏、剪枝与任务专用模型","url":"#小模型路线蒸馏剪枝与任务专用模型","depth":1},{"value":"一、先判断：你到底需要一个什么模型","url":"#一先判断你到底需要一个什么模型","depth":2},{"value":"二、四条常见的小模型路线","url":"#二四条常见的小模型路线","depth":2},{"value":"1. 量化：先减少每个参数占多少空间","url":"#1-量化先减少每个参数占多少空间","depth":3},{"value":"2. 剪枝：删除不重要的结构","url":"#2-剪枝删除不重要的结构","depth":3},{"value":"3. 知识蒸馏：让学生学习老师的行为","url":"#3-知识蒸馏让学生学习老师的行为","depth":3},{"value":"4. 任务专用微调：只学真正需要的能力","url":"#4-任务专用微调只学真正需要的能力","depth":3},{"value":"三、蒸馏数据怎么做才不把错误放大","url":"#三蒸馏数据怎么做才不把错误放大","depth":2},{"value":"四、评测不能只看平均准确率","url":"#四评测不能只看平均准确率","depth":2},{"value":"五、路由比二选一更现实","url":"#五路由比二选一更现实","depth":2},{"value":"六、部署时别忘了真实硬件","url":"#六部署时别忘了真实硬件","depth":2},{"value":"总结：小模型不是妥协，而是边界清楚","url":"#总结小模型不是妥协而是边界清楚","depth":2}] Sat, 04 Apr 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLM小模型模型压缩部署优化 https://www.hansking.cn/blog/2026tui-li-mo-xing-xing-qi-wei-shen-me-ce-shi-shi-ji-suan-kai-shi-bian-zhong-yao https://www.hansking.cn/blog/2026tui-li-mo-xing-xing-qi-wei-shen-me-ce-shi-shi-ji-suan-kai-shi-bian-zhong-yao [{"value":"推理模型兴起：为什么测试时计算开始变重要","url":"#推理模型兴起为什么测试时计算开始变重要","depth":1},{"value":"一、为什么单次生成开始遇到瓶颈","url":"#一为什么单次生成开始遇到瓶颈","depth":2},{"value":"二、测试时计算到底在增加什么","url":"#二测试时计算到底在增加什么","depth":2},{"value":"三、推理模型为什么看起来像是在思考","url":"#三推理模型为什么看起来像是在思考","depth":2},{"value":"四、从 Chain-of-Thought 到可验证推理","url":"#四从-chain-of-thought-到可验证推理","depth":2},{"value":"五、为什么这会改变模型训练","url":"#五为什么这会改变模型训练","depth":2},{"value":"六、推理预算是一种新的服务资源","url":"#六推理预算是一种新的服务资源","depth":2},{"value":"七、质量曲线比单点榜单更重要","url":"#七质量曲线比单点榜单更重要","depth":2},{"value":"八、它对模型产品意味着什么","url":"#八它对模型产品意味着什么","depth":2},{"value":"九、它没有解决什么问题","url":"#九它没有解决什么问题","depth":2},{"value":"十、作为开发者应该怎样学习","url":"#十作为开发者应该怎样学习","depth":2},{"value":"十一、我的总结：模型的能力开始有了“运行时维度”","url":"#十一我的总结模型的能力开始有了运行时维度","depth":2}] Sun, 05 Jan 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLM推理模型Test-time Compute发展史 https://www.hansking.cn/blog/2026tui-li-mo-xing-de-si-wei-lian-ping-ce-yu-jie-guo-jiao-yan https://www.hansking.cn/blog/2026tui-li-mo-xing-de-si-wei-lian-ping-ce-yu-jie-guo-jiao-yan [{"value":"推理模型的思维链评测与结果校验","url":"#推理模型的思维链评测与结果校验","depth":1},{"value":"一、先分清三个概念","url":"#一先分清三个概念","depth":2},{"value":"最终结果监督：答案对不对","url":"#最终结果监督答案对不对","depth":3},{"value":"过程监督：中间步骤是否合理","url":"#过程监督中间步骤是否合理","depth":3},{"value":"可验证中间结果：能否由系统复查","url":"#可验证中间结果能否由系统复查","depth":3},{"value":"二、不要把“展示思维链”当成评测的终点","url":"#二不要把展示思维链当成评测的终点","depth":2},{"value":"三、先设计可评测的任务，而不是先挑模型","url":"#三先设计可评测的任务而不是先挑模型","depth":2},{"value":"四、自动校验器是推理评测的核心资产","url":"#四自动校验器是推理评测的核心资产","depth":2},{"value":"数值与公式","url":"#数值与公式","depth":3},{"value":"代码","url":"#代码","depth":3},{"value":"SQL 与数据操作","url":"#sql-与数据操作","depth":3},{"value":"结构化计划","url":"#结构化计划","depth":3},{"value":"五、如何评测中间步骤","url":"#五如何评测中间步骤","depth":2},{"value":"六、自洽采样有用，但不能迷信多数票","url":"#六自洽采样有用但不能迷信多数票","depth":2},{"value":"七、裁判模型只能当辅助裁判","url":"#七裁判模型只能当辅助裁判","depth":2},{"value":"八、生产环境要校验结果，而不只是评测结果","url":"#八生产环境要校验结果而不只是评测结果","depth":2},{"value":"九、我在评测中最常见的三个误区","url":"#九我在评测中最常见的三个误区","depth":2},{"value":"十、我的总结：让结论接受事实的审判","url":"#十我的总结让结论接受事实的审判","depth":2}] Sun, 16 Mar 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLM推理模型评测思维链 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 https://www.hansking.cn/blog/2026jie-gou-hua-sheng-cheng-de-zhong-ji-shi-jian-JSONSchema-yu-lei-xing-an-quan https://www.hansking.cn/blog/2026jie-gou-hua-sheng-cheng-de-zhong-ji-shi-jian-JSONSchema-yu-lei-xing-an-quan [{"value":"结构化生成的终极实践：JSON Schema 与类型安全","url":"#结构化生成的终极实践json-schema-与类型安全","depth":1},{"value":"一、先设计数据契约，再写 Prompt","url":"#一先设计数据契约再写-prompt","depth":2},{"value":"二、TypeScript 类型和运行时校验是两回事","url":"#二typescript-类型和运行时校验是两回事","depth":2},{"value":"三、输出协议要包含版本和状态","url":"#三输出协议要包含版本和状态","depth":2},{"value":"四、Prompt 仍然重要，但不再承担全部责任","url":"#四prompt-仍然重要但不再承担全部责任","depth":2},{"value":"五、解析失败时要有分层修复策略","url":"#五解析失败时要有分层修复策略","depth":2},{"value":"六、语义校验比形状校验更难","url":"#六语义校验比形状校验更难","depth":2},{"value":"七、批量处理要记录坏样本","url":"#七批量处理要记录坏样本","depth":2},{"value":"八、从 Schema 生成类型和文档","url":"#八从-schema-生成类型和文档","depth":2},{"value":"九、评测结构化生成要看失败类型","url":"#九评测结构化生成要看失败类型","depth":2},{"value":"总结：结构化生成是协议设计，不是格式许愿","url":"#总结结构化生成是协议设计不是格式许愿","depth":2}] Thu, 22 Jan 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLMJSON SchemaTypeScript结构化输出 https://www.hansking.cn/blog/2026du-lun-wen-fang-fa-ru-he-chai-jie-yi-pian-da-mo-xing-lun-wen https://www.hansking.cn/blog/2026du-lun-wen-fang-fa-ru-he-chai-jie-yi-pian-da-mo-xing-lun-wen [{"value":"读论文方法：如何拆解一篇大模型论文","url":"#读论文方法如何拆解一篇大模型论文","depth":1},{"value":"一、第一遍只回答三个问题","url":"#一第一遍只回答三个问题","depth":2},{"value":"二、把“贡献”拆成可验证的主张","url":"#二把贡献拆成可验证的主张","depth":2},{"value":"三、第二遍读方法：先看数据流，再看公式","url":"#三第二遍读方法先看数据流再看公式","depth":2},{"value":"四、数据决定了结论能走多远","url":"#四数据决定了结论能走多远","depth":2},{"value":"五、基线和消融决定实验是否有说服力","url":"#五基线和消融决定实验是否有说服力","depth":2},{"value":"六、读图表时先看坐标和分母","url":"#六读图表时先看坐标和分母","depth":2},{"value":"七、论文的限制部分通常藏着真正的边界","url":"#七论文的限制部分通常藏着真正的边界","depth":2},{"value":"八、复现之前先写资源清单","url":"#八复现之前先写资源清单","depth":2},{"value":"九、我自己的五问法","url":"#九我自己的五问法","depth":2},{"value":"总结：读懂论文不是背下更多名词","url":"#总结读懂论文不是背下更多名词","depth":2}] Sun, 12 Jul 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLM论文阅读学习方法科研实践 https://www.hansking.cn/blog/2026ruan-jian-gong-cheng-Agent-cong-Issue-dao-ke-shen-cha-PullRequest https://www.hansking.cn/blog/2026ruan-jian-gong-cheng-Agent-cong-Issue-dao-ke-shen-cha-PullRequest [{"value":"软件工程 Agent：从 Issue 到可审查 Pull Request","url":"#软件工程-agent从-issue-到可审查-pull-request","depth":1},{"value":"一、Issue 不是需求，先把它变成验收条件","url":"#一issue-不是需求先把它变成验收条件","depth":2},{"value":"二、先建立仓库基线","url":"#二先建立仓库基线","depth":2},{"value":"三、工作区隔离比自动提交更重要","url":"#三工作区隔离比自动提交更重要","depth":2},{"value":"四、计划要短，但必须能被验证","url":"#四计划要短但必须能被验证","depth":2},{"value":"五、实现阶段坚持最小补丁","url":"#五实现阶段坚持最小补丁","depth":2},{"value":"六、测试要围绕用户行为，而不是只追求覆盖率","url":"#六测试要围绕用户行为而不是只追求覆盖率","depth":2},{"value":"七、处理失败：停止比继续改更聪明","url":"#七处理失败停止比继续改更聪明","depth":2},{"value":"八、生成 PR 时交付证据","url":"#八生成-pr-时交付证据","depth":2},{"value":"九、合并前保留人的判断","url":"#九合并前保留人的判断","depth":2},{"value":"总结：自动化的终点是可验证，不是无人负责","url":"#总结自动化的终点是可验证不是无人负责","depth":2}] Sun, 10 May 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLMCoding AgentGit软件工程 https://www.hansking.cn/blog/2026liang-hua-shi-jian-GPTQAWQGGUF-ru-he-xuan-ze https://www.hansking.cn/blog/2026liang-hua-shi-jian-GPTQAWQGGUF-ru-he-xuan-ze [{"value":"量化实践：GPTQ、AWQ、GGUF 如何选择","url":"#量化实践gptqawqgguf-如何选择","depth":1},{"value":"一、量化到底在损失什么","url":"#一量化到底在损失什么","depth":2},{"value":"二、GPTQ：面向权重的后训练量化","url":"#二gptq面向权重的后训练量化","depth":2},{"value":"三、AWQ：优先保护重要权重","url":"#三awq优先保护重要权重","depth":2},{"value":"四、GGUF：更像一条本地和跨平台生态路径","url":"#四gguf更像一条本地和跨平台生态路径","depth":2},{"value":"五、先按部署目标选择","url":"#五先按部署目标选择","depth":2},{"value":"六、校准数据决定量化后的边界","url":"#六校准数据决定量化后的边界","depth":2},{"value":"七、评测不要只看模型能不能加载","url":"#七评测不要只看模型能不能加载","depth":2},{"value":"八、KV Cache 和上下文会改变结论","url":"#八kv-cache-和上下文会改变结论","depth":2},{"value":"九、常见错误和排查顺序","url":"#九常见错误和排查顺序","depth":2},{"value":"总结：量化选择是部署选择","url":"#总结量化选择是部署选择","depth":2}] Sun, 23 Nov 2025 10:00:00 GMT admin@hansking.cn (HANS) AILLM量化GPTQAWQGGUF https://www.hansking.cn/blog/2026chang-ren-wu-Agent-jian-cha-dian-hui-fu-yu-mi-deng-zhi-xing https://www.hansking.cn/blog/2026chang-ren-wu-Agent-jian-cha-dian-hui-fu-yu-mi-deng-zhi-xing [{"value":"长任务 Agent：检查点、恢复与幂等执行","url":"#长任务-agent检查点恢复与幂等执行","depth":1},{"value":"一、先把任务从对话里拿出来","url":"#一先把任务从对话里拿出来","depth":2},{"value":"二、用状态机表达任务生命周期","url":"#二用状态机表达任务生命周期","depth":2},{"value":"三、检查点要保存什么","url":"#三检查点要保存什么","depth":2},{"value":"四、最危险的是“成功了，但我没收到成功响应”","url":"#四最危险的是成功了但我没收到成功响应","depth":2},{"value":"五、重试不是遇到错误就再来一次","url":"#五重试不是遇到错误就再来一次","depth":2},{"value":"六、恢复时不要盲目重跑整个计划","url":"#六恢复时不要盲目重跑整个计划","depth":2},{"value":"七、进度要让用户看得懂","url":"#七进度要让用户看得懂","depth":2},{"value":"八、检查点本身也要版本化和保护","url":"#八检查点本身也要版本化和保护","depth":2},{"value":"九、可靠性评测要模拟真实故障","url":"#九可靠性评测要模拟真实故障","depth":2},{"value":"总结：长任务 Agent 的记忆应该属于系统","url":"#总结长任务-agent-的记忆应该属于系统","depth":2}] Sat, 06 Jun 2026 10:00:00 GMT admin@hansking.cn (HANS) AILLMAgent工作流引擎可靠性 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