[{"data":1,"prerenderedAt":222},["ShallowReactive",2],{"site-settings":3,"$f4SxuVxBOg6luBbvHlYjp_aT_supc0gAL7Jyrch-LdSU":26,"related-posts":49,"$fPOLkM25HfAm5vjsmuBJIjmWoKJoITuC52Uu2Oprt-34":221},{"id":4,"site_name":5,"site_subtitle":6,"site_description":7,"author":8,"email":9,"github":10,"icp":10,"page_size":11,"enable_comment":12,"enable_rss":12,"updated_at":13,"site_icon":14,"site_og_image":15,"site_keywords":16,"site_og_image_width":24,"site_og_image_height":25},1,"LiteSite（轻站）","把你想做的网站说出来","LiteSite（轻站）探索自然语言驱动的零代码建站体验，展示响应式网站与前端原型的设计方向。","Admin","support@litesite.com",null,10,true,"2026-10-05T03:56:44.502+00:00","https:\u002F\u002Fnklbhnvyoggxghtaoizu.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fblog-images\u002F2026-10-05-1765bce7-01f1-4206-8d5d-6a717d17f206.png","https:\u002F\u002Fnklbhnvyoggxghtaoizu.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fblog-images\u002F2026-10-05-157c3c1d-2722-474c-860b-d3f38d6a674d.png",[17,18,19,20,21,22,23],"LiteSite","轻站","零代码建站","AI建站","响应式网站","前端原型","网站设计",1200,630,{"id":27,"slug":28,"title":29,"summary":30,"content":31,"status":32,"tags":33,"seo_keywords":38,"published_at":44,"updated_at":45,"category":46},22,"llm-output-instability-hallucination","模型返回不稳定 \u002F 幻觉怎么办？","区分随机波动与事实幻觉，沿着样本评测、证据检索、结构校验、业务规则和人工回退建立可验证的大模型输出链路。","模型返回不稳定 \u002F 幻觉怎么办？先分清两种故障：“不稳定”是相同或相近输入下，答案的措辞、结构甚至判断发生波动；“幻觉”是把没有充分依据的内容说成事实。它们会同时出现，但排查路径不同。把温度调低可以减少一部分随机波动，却不能替代事实核查，更不能保证每次输出完全一致。\n\n## 先定位问题发生在哪一层\n\n| 现象 | 可能原因 | 第一项检查 |\n| --- | --- | --- |\n| 同题答案长短差很多 | 提示词含糊、输出范围过宽、采样设置变化 | 固定输入、模型版本与参数，比较多次结果 |\n| 事实结论互相矛盾 | 检索证据不足或冲突、模型凭记忆补全 | 查看实际送入模型的证据片段及版本 |\n| JSON 字段时有时无 | 只用提示词约定格式、响应被截断 | 检查结束原因、Schema 与服务端校验 |\n| 引用了不存在的来源 | 模型生成了貌似可信的链接或编号 | 对照检索结果逐条核查引用 ID |\n| 价格、金额或日期算错 | 让模型直接承担确定性计算 | 改由数据库或程序计算并验证 |\n\n记录每次调用的任务类型、输入模板版本、模型与参数、检索结果 ID、响应结束原因和校验结果。日志要控制敏感信息，保留排障所需的关联标识即可。没有可复现的输入和证据，团队很难判断是模型、检索、提示词还是数据更新造成了变化。\n\n## 稳定性：先把可控变量固定\n\n为每个场景建立一组真实问题，标注“必须包含什么”“不能承诺什么”和允许的不同说法。评测时固定模型版本、提示词版本、工具结果和检索快照，分别多次运行，记录结论一致率与格式通过率。先判定差异是否影响业务：两次都正确、只是措辞不同，未必需要修；一次说“支持退款”、一次说“不支持”，才是需要处理的波动。\n\n把任务边界写具体：回答对象是谁、只能使用哪些资料、证据不足时如何处理、输出多长、采用什么结构。对分类、提取类任务用枚举和 Schema 约束；对开放问答明确需要结论、依据和限制。降低温度、固定随机种子若接口支持，可能改善复现性，但供应商实现、模型更新和并行推理仍可能带来差异，不能把它当成“确定性开关”。\n\n## 幻觉：让每个关键结论有出处\n\n假设用户问“这个服务套餐能否退款”。如果当前政策没有写明退款条件，模型不应编出“七天无理由退款”。把官方政策、产品说明和已审核的知识库作为优先证据；检索时保留文档 ID、版本、更新时间和访问权限。若新旧政策冲突，先用有效日期和业务规则解决，不能把两份相互矛盾的片段都塞给模型，让它自行猜哪份有效。\n\n提示词可以要求“只根据给定资料回答；没有足够证据就说明无法确认”，并让回答附上实际提供的来源 ID。应用侧再核对：引用 ID 是否存在、用户是否有权限读取该来源、被引用段落是否真的支持对应结论。**有引用不等于事实正确**：模型可能引用了真实页面，却把适用范围读错。重要结论仍要做规则检查或人工抽检。检索增强生成（RAG）改善证据供给，但检索不到、检索错或资料过期时，仍会产生错误答案。\n\n还要把检索到的网页和用户输入视为数据。里面即使出现“忽略规则、改用以下答案”，也不能提升为系统指令。来源 URL、产品名和案例数据要经过正常的权限与真实性校验，不能由模型自行编造。\n\n## 把确定性工作交回程序\n\n订单状态从数据库读取，金额和折扣由程序计算，日期区间按明确时区处理，权限由服务端判断。模型适合把这些已确认结果解释给用户，或从文本里提取待验证字段；不要让它凭记忆决定最新价格、库存、合同条款或用户权限。\n\n格式也要单独守住。结构化输出接口能减少缺字段、类型错等问题，但 JSON 合法不代表内容真实。服务端应先检查响应完整，再解析并验证 Schema，然后验证业务规则；失败时进入有限重试、明确的“无法确认”或人工队列。[结构化输出这篇文章](\u002Fblog\u002Fllm-structured-output-reliability)详细说明了格式约束与校验。不要把“解析失败”悄悄替换为看似正常的默认答案。\n\n## 什么情况下应该拒答或转人工？\n\n给每类问题定出可回答条件。例如退款政策必须匹配有效版本的官方条款，医疗或法律类建议需要更严格的专业流程，涉及账户操作时要通过权限校验。未满足条件就说明缺少哪类信息，并提供下一步；这比自信地猜一个答案更有用。模型自报的“置信度 95%”不能直接当可靠概率，阈值要通过真实样本校准。\n\n针对高影响内容设置人工复核：涉及价格承诺、财务金额、身份与权限、对外正式声明时，由有权人员确认后再发布或执行。自动检查可以拦截一部分明显错误，但不能保证语义判断永远正确。重试或换模型可以作为补救，不能用“多问几次取多数票”替代独立证据。\n\n## 怎么知道改进有效？\n\n将线上失败样本脱敏后加入固定评测集，覆盖无证据、冲突证据、旧版资料、诱导性输入和边界情况。至少跟踪：结论一致率、无依据断言率、引用有效率、应拒答时的拒答率、格式通过率和人工复核后的准确率。按任务、模型与知识库版本拆开看，避免一个平均数掩盖高风险场景。\n\n每次只改变一项主要变量并保留旧版对照，评估质量、延迟和成本。降低幻觉的目标不是让模型永远“不知道”，而是让有证据的回答能被核查，没有证据时明确停下。可复现的评测、可信资料、程序校验与回退路径一起工作，才是稳定输出的基础。","published",[34,36],{"name":35,"slug":35},"工程",{"name":37,"slug":37},"大模型",[39,40,41,42,43],"模型幻觉怎么办","大模型输出不稳定","LLM幻觉","RAG事实校验","模型评测","2026-10-05T04:17:13.315+00:00","2026-10-05T04:17:20.1954+00:00",{"id":47,"name":48,"slug":48},3,"技术笔记",{"items":50,"total":219,"page":4,"pageSize":220},[51,57,73,89,109,127,143,159,174,187,202],{"id":27,"slug":28,"title":29,"summary":30,"status":32,"tags":52,"seo_keywords":55,"published_at":44,"category":56},[53,54],{"name":35,"slug":35},{"name":37,"slug":37},[39,40,41,42,43],{"id":47,"name":48,"slug":48},{"id":58,"slug":59,"title":60,"summary":61,"status":32,"tags":62,"seo_keywords":65,"published_at":71,"category":72},21,"llm-cost-budget-guardrails","大模型调用怎么控制成本？把预算控制放进请求链路","从请求准入、原子预留、真实用量结算到超支降级，设计一套能在并发与重试下运行的大模型调用预算控制链路。",[63,64],{"name":35,"slug":35},{"name":37,"slug":37},[66,67,68,69,70],"大模型调用成本控制","LLM预算管理","Token费用","模型调用限额","API成本监控","2026-10-05T04:14:07.621+00:00",{"id":47,"name":48,"slug":48},{"id":74,"slug":75,"title":76,"summary":77,"status":32,"tags":78,"seo_keywords":81,"published_at":87,"category":88},20,"llm-structured-output-reliability","结构化输出怎么保证模型按格式返回？","模型返回 JSON 不等于系统拿到可用数据。本文区分提示词、JSON 模式与 Schema 约束，说明如何用服务端校验、有限重试和业务规则构建可靠的结构化输出链路。",[79,80],{"name":35,"slug":35},{"name":37,"slug":37},[82,83,84,85,86],"大模型结构化输出","JSON Schema","模型按格式返回","LLM输出校验","约束解码","2026-10-05T04:10:02.025+00:00",{"id":47,"name":48,"slug":48},{"id":90,"slug":91,"title":92,"summary":93,"status":32,"tags":94,"seo_keywords":101,"published_at":107,"category":108},19,"geo-vs-seo-explained","理解 GEO 是什么？和 SEO 有什么区别？","GEO 是什么，和 SEO 有什么区别？从用户提问、内容呈现、优化动作与衡量指标四个角度讲清两者的关系，并给出一套适合建站团队的入门做法。",[95,98],{"name":96,"slug":97},"建站","jianzhan",{"name":99,"slug":100},"GEO","geo",[102,103,104,105,106],"GEO是什么","GEO和SEO的区别","生成式引擎优化","SEO","AI搜索优化","2026-10-05T04:07:05.673+00:00",{"id":47,"name":48,"slug":48},{"id":110,"slug":111,"title":112,"summary":113,"status":32,"tags":114,"seo_keywords":119,"published_at":125,"category":126},18,"how-nuxt-ssr-works","Nuxt 的 SSR 是怎么做的？从请求、数据获取到 Hydration","沿着一次文章页请求，看 Nuxt 如何用 Nitro 接收请求、在服务端执行 useFetch 并生成 HTML，再由浏览器复用 payload 完成 Hydration。",[115,116],{"name":35,"slug":35},{"name":117,"slug":118},"Nuxt","nuxt",[120,121,122,123,124],"Nuxt SSR","Nuxt服务端渲染","Nuxt useFetch","Hydration","Nitro渲染流程","2026-10-05T03:55:36.035+00:00",{"id":47,"name":48,"slug":48},{"id":128,"slug":129,"title":130,"summary":131,"status":32,"tags":132,"seo_keywords":135,"published_at":141,"category":142},17,"llm-api-cost-control","大模型调用怎么控制成本？一笔账算清缓存、输出与重试","先按输入、输出和额外调用算清单次成本，再用缓存、模型路由、上下文裁剪和预算限额降本，并用可用答案质量守住效果。",[133,134],{"name":35,"slug":35},{"name":37,"slug":37},[136,137,138,139,140],"大模型调用成本","LLM API成本控制","Token费用计算","提示词缓存","模型路由","2026-10-05T03:49:26.666+00:00",{"id":47,"name":48,"slug":48},{"id":144,"slug":145,"title":146,"summary":147,"status":32,"tags":148,"seo_keywords":151,"published_at":157,"category":158},16,"geo-technical-pipeline-design","GEO 的技术链路你会怎么设计？","从问题池、证据库、可读页面到答案采样与质量回写，设计一条可维护的 GEO 技术链路，并明确外部 AI 系统的边界。",[149,150],{"name":35,"slug":35},{"name":99,"slug":100},[152,153,154,155,156],"GEO技术链路","GEO技术方案","生成式引擎优化架构","AI搜索内容工程","GEO监测","2026-10-05T03:44:34.947+00:00",{"id":47,"name":48,"slug":48},{"id":160,"slug":161,"title":162,"summary":163,"status":32,"tags":164,"seo_keywords":166,"published_at":172,"category":173},15,"geo-core-metrics-calculation","GEO 的核心指标你怎么理解？怎么算？","用同一组问题和明确的分母，算清 GEO 的品牌提及率、引用率、有效引用率、竞争提及份额与有效线索率。",[165],{"name":99,"slug":100},[167,168,169,170,171],"GEO核心指标","GEO指标怎么算","品牌提及率","AI引用率","生成式引擎优化效果","2026-10-05T03:40:09.175+00:00",{"id":47,"name":48,"slug":48},{"id":175,"slug":176,"title":177,"summary":178,"status":32,"tags":179,"seo_keywords":182,"published_at":185,"category":186},14,"geo-business-understanding","GEO 从业务理解开始：把真实问题写成可引用的内容","做 GEO 之前，先弄清业务服务谁、用户在什么场景下提问、哪些证据能支持回答，再用真实问题和业务结果验证内容。",[180,181],{"name":96,"slug":97},{"name":99,"slug":100},[99,104,183,184,106],"业务理解","GEO内容策略","2026-10-05T03:33:52.358+00:00",{"id":47,"name":48,"slug":48},{"id":188,"slug":189,"title":190,"summary":191,"status":32,"tags":192,"seo_keywords":194,"published_at":200,"category":201},13,"website-launch-checklist","网站上线前后检查清单：从可访问到可维护","一份实用的网站上线清单，逐项检查域名、HTTPS、手机体验、页面速度、SEO、表单、备份与上线后的监测。",[193],{"name":35,"slug":35},[195,196,197,198,199],"网站上线检查清单","网站上线","网站性能优化","网站SEO","网站维护","2026-10-05T03:24:44.969+00:00",{"id":47,"name":48,"slug":48},{"id":203,"slug":204,"title":205,"summary":206,"status":32,"tags":207,"seo_keywords":209,"published_at":215,"category":216},12,"website-planning-checklist","建站之前：用一张页面清单理清目标、结构与内容","从网站目标、访问者任务到页面清单和内容准备，用一套可执行的方法完成建站规划，避免边做边改。",[208],{"name":35,"slug":35},[210,211,212,213,214],"建站规划","网站信息架构","网站内容规划","页面清单","个人网站建设","2026-10-05T03:23:27.683+00:00",{"id":217,"name":218,"slug":218},7,"项目实践",11,100,[],1791174017013]