[{"data":1,"prerenderedAt":1903},["ShallowReactive",2],{"blog-\u002Fblog\u002Fai-agent\u002Fdamoxing-weitiao":3,"blog-related-\u002Fblog\u002Fai-agent\u002Fdamoxing-weitiao":434},{"id":4,"title":5,"author":6,"body":7,"category":403,"cover":404,"date":405,"description":406,"draft":407,"extension":408,"faq":409,"featured":407,"image":404,"keywords":419,"meta":424,"navigation":425,"path":426,"seo":427,"sitemap":428,"stem":429,"tags":430,"updated":405,"__hash__":433},"blog\u002Fblog\u002Fai-agent\u002Fdamoxing-weitiao.md","大模型微调值不值得做","HNREIS",{"type":8,"value":9,"toc":385},"minimark",[10,19,23,26,54,57,60,115,120,123,137,140,145,156,160,168,172,180,184,195,199,267,272,275,307,310,358,361,379],[11,12,13,14,18],"p",{},"通用大模型能力很强，但在特定领域、风格、格式上可能不够。微调能定制，",[15,16,17],"strong",{},"但微调成本高、要数据、要持续维护，不是所有场景都需要，建议先用RAG和提示词。"," 这篇讲清什么情况该微调。",[20,21,22],"h2",{"id":22},"微调解决什么问题",[11,24,25],{},"通用模型不够用的场景：",[27,28,29,36,42,48],"ul",{},[30,31,32,35],"li",{},[15,33,34],{},"领域专业性","：特定行业的专业术语、知识、判断。",[30,37,38,41],{},[15,39,40],{},"固定风格","：品牌语气、特定写作风格。",[30,43,44,47],{},[15,45,46],{},"输出格式","：严格的固定输出结构。",[30,49,50,53],{},[15,51,52],{},"特定任务","：通用模型做不好的专属任务。",[20,55,56],{"id":56},"先试低成本方案",[11,58,59],{},"上微调前，先试：",[61,62,63,79],"table",{},[64,65,66],"thead",{},[67,68,69,73,76],"tr",{},[70,71,72],"th",{},"方案",[70,74,75],{},"成本",[70,77,78],{},"适合",[80,81,82,94,105],"tbody",{},[67,83,84,88,91],{},[85,86,87],"td",{},"提示词优化",[85,89,90],{},"低",[85,92,93],{},"风格、格式、引导",[67,95,96,99,102],{},[85,97,98],{},"RAG",[85,100,101],{},"低到中",[85,103,104],{},"知识、事实、可更新",[67,106,107,110,112],{},[85,108,109],{},"Few-shot示例",[85,111,90],{},[85,113,114],{},"格式、风格示范",[11,116,117],{},[15,118,119],{},"很多场景RAG+提示词就够，不必微调。",[20,121,122],{"id":122},"什么情况值得微调",[27,124,125,128,131,134],{},[30,126,127],{},"RAG和提示词优化后，仍达不到要求。",[30,129,130],{},"需求是稳定的（领域专业、风格、格式），不会频繁变。",[30,132,133],{},"有足够高质量的微调数据。",[30,135,136],{},"投入产出合理（高频、高价值场景）。",[20,138,139],{"id":139},"微调的代价",[141,142,144],"h3",{"id":143},"_1-数据成本","1. 数据成本",[27,146,147,150,153],{},[30,148,149],{},"要高质量标注数据。",[30,151,152],{},"数据量要够。",[30,154,155],{},"标注和清洗成本不小。",[141,157,159],{"id":158},"_2-算力成本","2. 算力成本",[27,161,162,165],{},[30,163,164],{},"GPU训练费用。",[30,166,167],{},"微调方法（全参\u002FLoRA）影响成本。",[141,169,171],{"id":170},"_3-评测成本","3. 评测成本",[27,173,174,177],{},[30,175,176],{},"要评测集验证效果。",[30,178,179],{},"防止\"微调后能力退化\"。",[141,181,183],{"id":182},"_4-持续维护","4. 持续维护",[27,185,186,189,192],{},[30,187,188],{},"模型升级要重新微调。",[30,190,191],{},"持续迭代优化。",[30,193,194],{},"数据更新。",[20,196,198],{"id":197},"微调-vs-rag","微调 vs RAG",[61,200,201,213],{},[64,202,203],{},[67,204,205,208,211],{},[70,206,207],{},"维度",[70,209,210],{},"微调",[70,212,98],{},[80,214,215,226,237,246,257],{},[67,216,217,220,223],{},[85,218,219],{},"解决",[85,221,222],{},"能力\u002F风格\u002F格式",[85,224,225],{},"知识\u002F事实",[67,227,228,231,234],{},[85,229,230],{},"知识更新",[85,232,233],{},"难（重新微调）",[85,235,236],{},"易（更新知识库）",[67,238,239,241,244],{},[85,240,75],{},[85,242,243],{},"高",[85,245,101],{},[67,247,248,251,254],{},[85,249,250],{},"幻觉",[85,252,253],{},"仍可能",[85,255,256],{},"可约束范围",[67,258,259,261,264],{},[85,260,78],{},[85,262,263],{},"稳定能力",[85,265,266],{},"动态知识",[11,268,269],{},[15,270,271],{},"两者互补，按场景选。",[20,273,274],{"id":274},"别踩的坑",[27,276,277,283,289,295,301],{},[30,278,279,282],{},[15,280,281],{},"上来就微调","：很多场景RAG够用。",[30,284,285,288],{},[15,286,287],{},"数据不够质量差","：微调效果差甚至退化。",[30,290,291,294],{},[15,292,293],{},"不评测就上","：不知道效果，可能更差。",[30,296,297,300],{},[15,298,299],{},"忽视持续维护","：模型升级后失效。",[30,302,303,306],{},[15,304,305],{},"微调解决知识问题","：知识用RAG更合适。",[20,308,309],{"id":309},"成本参考",[61,311,312,324],{},[64,313,314],{},[67,315,316,318,321],{},[70,317,72],{},[70,319,320],{},"说明",[70,322,323],{},"成本量级",[80,325,326,337,347],{},[67,327,328,331,334],{},[85,329,330],{},"LoRA轻量微调",[85,332,333],{},"少量数据+轻量算力",[85,335,336],{},"中",[67,338,339,342,345],{},[85,340,341],{},"全参微调",[85,343,344],{},"大量数据+算力",[85,346,243],{},[67,348,349,352,355],{},[85,350,351],{},"微调+评测+维护",[85,353,354],{},"完整流程",[85,356,357],{},"高，持续投入",[20,359,360],{"id":360},"怎么开始",[362,363,364,367,370,373,376],"ol",{},[30,365,366],{},"先优化提示词和用RAG。",[30,368,369],{},"仍不满足，评估是否真需要微调。",[30,371,372],{},"准备高质量微调数据。",[30,374,375],{},"评测验证效果（防退化）。",[30,377,378],{},"规划持续维护。",[380,381,382],"blockquote",{},[11,383,384],{},"广州市汉诺雷斯（HNREIS）帮企业评估和实施大模型微调，从需求评估、数据准备到微调、评测和持续维护。把你的定制需求告诉我们，我们先评估是否真需要微调，给出务实方案。",{"title":386,"searchDepth":387,"depth":387,"links":388},"",2,[389,390,391,392,399,400,401,402],{"id":22,"depth":387,"text":22},{"id":56,"depth":387,"text":56},{"id":122,"depth":387,"text":122},{"id":139,"depth":387,"text":139,"children":393},[394,396,397,398],{"id":143,"depth":395,"text":144},3,{"id":158,"depth":395,"text":159},{"id":170,"depth":395,"text":171},{"id":182,"depth":395,"text":183},{"id":197,"depth":387,"text":198},{"id":274,"depth":387,"text":274},{"id":309,"depth":387,"text":309},{"id":360,"depth":387,"text":360},"ai-agent",null,"2025-11-11","通用大模型不够用时，微调能定制领域\u002F风格\u002F格式。但微调成本高、要数据、要持续维护。本文讲清什么情况该微调，先试RAG和提示词。",false,"md",[410,413,416],{"q":411,"a":412},"什么情况需要微调大模型？","当通用模型在领域专业、特定风格、固定输出格式上达不到要求，且RAG和提示词优化都解决不了时，考虑微调。比如特定行业的专业问答、固定的输出格式、特定的语气风格。但很多场景RAG+提示词就够，不必上来就微调，微调成本高且要持续维护。",{"q":414,"a":415},"微调和RAG有什么区别？","RAG是让模型基于检索到的外部知识回答，知识可随时更新，适合知识频繁变化的场景；微调是把能力\"训练\"进模型，适合稳定的领域专业、风格、格式。很多场景两者互补。建议优先RAG（成本低、灵活），知识稳定的能力需求再考虑微调。",{"q":417,"a":418},"微调要花多少钱？","不便宜。要准备高质量的微调数据（标注成本）、算力（GPU训练）、评测、以及持续的维护和迭代。还可能涉及模型升级后重新微调。相比RAG和提示词，微调是一次性和持续性都较高的投入，建议确实需要时再做。",[420,421,422,423],"大模型微调","模型微调","LoRA微调","定制大模型",{},true,"\u002Fblog\u002Fai-agent\u002Fdamoxing-weitiao",{"title":5,"description":406},{"loc":426},"blog\u002Fai-agent\u002Fdamoxing-weitiao",[210,431,432],"大模型","工程","T6uwOpQ7OZil2912TVszvcXr5eom5mAEnB50Z5Lr4zE",[435,823,1173,1541],{"id":436,"title":437,"author":6,"body":438,"category":403,"cover":404,"date":799,"description":800,"draft":407,"extension":408,"faq":801,"featured":407,"image":404,"keywords":811,"meta":815,"navigation":425,"path":816,"seo":817,"sitemap":818,"stem":819,"tags":820,"updated":799,"__hash__":822},"blog\u002Fblog\u002Fai-agent\u002Fagent-kuangjia.md","主流Agent框架怎么选",{"type":8,"value":439,"toc":786},[440,447,451,454,474,481,484,621,626,629,633,645,649,656,660,663,667,670,672,704,706,762,764,781],[11,441,442,443,446],{},"企业要落地一个AI智能体，市面框架多得让人发懵：LangChain、LlamaIndex、LangGraph、AutoGen、CrewAI、Dify、Coze……选错了要么过度复杂，要么撑不住业务。",[15,444,445],{},"框架不是越新越热门越好，而是要匹配你的场景和团队能力。"," 这篇讲清怎么选。",[20,448,450],{"id":449},"先想清楚你到底需不需要框架","先想清楚：你到底需不需要框架",[11,452,453],{},"很多人一上来就要\"上框架\"，其实很多场景不需要：",[27,455,456,462,468],{},[30,457,458,461],{},[15,459,460],{},"简单问答","：直接调大模型 API，几十行代码搞定，不用框架。",[30,463,464,467],{},[15,465,466],{},"RAG 知识库问答","：用轻量库或框架的检索能力，中等复杂度。",[30,469,470,473],{},[15,471,472],{},"复杂智能体","：多步推理、调用多个工具、多轮对话有记忆、多个智能体协作——这才是框架真正发挥作用的地方。",[11,475,476,477,480],{},"需求复杂度决定框架价值。",[15,478,479],{},"简单需求硬上重框架，是典型的过度设计","。",[20,482,483],{"id":483},"主流框架对比",[61,485,486,505],{},[64,487,488],{},[67,489,490,493,496,499,502],{},[70,491,492],{},"框架",[70,494,495],{},"类型",[70,497,498],{},"核心定位",[70,500,501],{},"适合谁",[70,503,504],{},"主要局限",[80,506,507,524,540,556,572,588,605],{},[67,508,509,512,515,518,521],{},[85,510,511],{},"LangChain",[85,513,514],{},"代码框架",[85,516,517],{},"通用 LLM 应用编排，生态最全",[85,519,520],{},"有研发团队、要深度定制",[85,522,523],{},"抽象较重，早期 API 频繁变动",[67,525,526,529,531,534,537],{},[85,527,528],{},"LangGraph",[85,530,514],{},[85,532,533],{},"基于图的状态机，做可控多步 Agent",[85,535,536],{},"复杂流程、需要稳定状态控制",[85,538,539],{},"学习曲线陡",[67,541,542,545,547,550,553],{},[85,543,544],{},"LlamaIndex",[85,546,514],{},[85,548,549],{},"数据连接与检索（RAG 见长）",[85,551,552],{},"以知识库\u002F文档检索为核心",[85,554,555],{},"Agent 编排不如 LangChain 全",[67,557,558,561,563,566,569],{},[85,559,560],{},"AutoGen",[85,562,514],{},[85,564,565],{},"多智能体对话协作",[85,567,568],{},"多 Agent 角色分工场景",[85,570,571],{},"偏研究，工程化要自己补",[67,573,574,577,579,582,585],{},[85,575,576],{},"CrewAI",[85,578,514],{},[85,580,581],{},"角色化多 Agent 协作，上手快",[85,583,584],{},"快速搭多角色协作原型",[85,586,587],{},"复杂控制力有限",[67,589,590,593,596,599,602],{},[85,591,592],{},"Dify",[85,594,595],{},"低代码平台",[85,597,598],{},"可视化编排 + 知识库 + 应用托管",[85,600,601],{},"快速落地、业务人员参与",[85,603,604],{},"深度定制受平台能力限制",[67,606,607,610,612,615,618],{},[85,608,609],{},"Coze",[85,611,595],{},[85,613,614],{},"字节出品，插件生态丰富",[85,616,617],{},"快速搭建对话型应用",[85,619,620],{},"偏消费侧，企业私有化受限",[11,622,623],{},[15,624,625],{},"没有\"最好\"的框架，只有\"最匹配\"的框架。",[20,627,628],{"id":628},"选型要看的维度",[141,630,632],{"id":631},"_1-编程式还是低代码","1. 编程式还是低代码",[27,634,635,640],{},[30,636,637,639],{},[15,638,514],{},"（LangChain\u002FLlamaIndex 等）：灵活、可控、能深度集成业务系统，但要有研发投入。",[30,641,642,644],{},[15,643,595],{},"（Dify\u002FCoze）：拖拽配置、上手快、适合原型和非纯技术团队，但定制边界明显。",[141,646,648],{"id":647},"_2-单-agent-还是多-agent","2. 单 Agent 还是多 Agent",[11,650,651,652,655],{},"大部分企业场景，",[15,653,654],{},"单个能调用工具的 Agent 就够用","。多智能体协作（AutoGen\u002FCrewAI）适合确实需要多角色分工的复杂流程，比如\"研究员+执行者+审核者\"。别为了炫技上多 Agent。",[141,657,659],{"id":658},"_3-能不能私有化部署","3. 能不能私有化部署",[11,661,662],{},"涉及企业内部数据，通常要求私有化。Dify 支持私有部署，Coze 偏公有云。代码框架本身可私有化，但要自己搭运维。",[141,664,666],{"id":665},"_4-社区活跃度与稳定性","4. 社区活跃度与稳定性",[11,668,669],{},"框架迭代快是双刃剑。优先选社区活跃、有企业背书、更新稳定的，避免踩\"作者弃坑\"的雷。",[20,671,274],{"id":274},[27,673,674,680,686,692,698],{},[30,675,676,679],{},[15,677,678],{},"过度设计","：简单需求上重框架，开发和维护成本翻倍。",[30,681,682,685],{},[15,683,684],{},"被框架锁死","：把提示词、工具、业务逻辑和框架深度耦合，迁移不动。",[30,687,688,691],{},[15,689,690],{},"忽视核心","：框架只是壳，真正决定效果的是提示词、数据质量和知识库。",[30,693,694,697],{},[15,695,696],{},"盲目追新","：新框架未必成熟，企业项目要稳。",[30,699,700,703],{},[15,701,702],{},"忽视可观测","：Agent 行为不确定性高，没有日志和追踪很难调试。",[20,705,309],{"id":309},[61,707,708,718],{},[64,709,710],{},[67,711,712,714,716],{},[70,713,72],{},[70,715,320],{},[70,717,323],{},[80,719,720,730,740,751],{},[67,721,722,725,728],{},[85,723,724],{},"自研（直接调 API）",[85,726,727],{},"简单场景，研发 1-2 周",[85,729,90],{},[67,731,732,735,738],{},[85,733,734],{},"基于代码框架开发",[85,736,737],{},"中等复杂度，研发 3-8 周",[85,739,336],{},[67,741,742,745,748],{},[85,743,744],{},"低代码平台搭建",[85,746,747],{},"快速原型到中等应用",[85,749,750],{},"平台费 + 少量定制",[67,752,753,756,759],{},[85,754,755],{},"复杂多 Agent 系统",[85,757,758],{},"多角色协作 + 工具集成 + 私有化",[85,760,761],{},"较高，定制开发",[20,763,360],{"id":360},[362,765,766,769,772,775,778],{},[30,767,768],{},"明确场景和复杂度（是否真需要框架）。",[30,770,771],{},"评估团队能力（能否驾驭代码框架）。",[30,773,774],{},"对比 2-3 个候选框架，做小范围验证。",[30,776,777],{},"关注私有化、可观测、可迁移。",[30,779,780],{},"把核心资产与框架解耦。",[380,782,783],{},[11,784,785],{},"广州市汉诺雷斯（HNREIS）帮企业做AI智能体落地，从场景评估、框架选型到私有化部署和工具集成。把你的业务场景告诉我们，我们给出务实的选型建议与方案。",{"title":386,"searchDepth":387,"depth":387,"links":787},[788,789,790,796,797,798],{"id":449,"depth":387,"text":450},{"id":483,"depth":387,"text":483},{"id":628,"depth":387,"text":628,"children":791},[792,793,794,795],{"id":631,"depth":395,"text":632},{"id":647,"depth":395,"text":648},{"id":658,"depth":395,"text":659},{"id":665,"depth":395,"text":666},{"id":274,"depth":387,"text":274},{"id":309,"depth":387,"text":309},{"id":360,"depth":387,"text":360},"2024-05-12","LangChain、LlamaIndex、LangGraph、AutoGen、CrewAI、Dify、Coze 等Agent框架各有侧重。本文从企业落地视角对比主流框架，讲清选型维度，帮你按场景选对工具而不踩坑。",[802,805,808],{"q":803,"a":804},"企业做AI智能体一定要用框架吗？","不一定。简单的单轮问答或单次工具调用，直接调大模型API就能实现。框架的价值在复杂场景：多步推理、工具编排、记忆管理、多智能体协作、可观测的流程控制。需求越复杂，框架越能省去重复造轮子。但框架也有学习成本和被锁定的风险，按需选择，不要为了用框架而用框架。",{"q":806,"a":807},"LangChain和Dify怎么选？","看团队和技术诉求。LangChain是代码框架，灵活、可控、可深度定制，适合有研发团队、要和业务系统深度集成的企业；Dify是低代码平台，拖拽编排、可视化，适合快速出原型、让业务人员参与配置。要做底层定制选代码框架，要快速落地和降低门槛选低代码平台。",{"q":809,"a":810},"用框架会被技术锁定吗？","有一定风险。深度依赖某框架的特有抽象，后续迁移成本会高。建议把核心资产——提示词、工具定义、业务数据、知识库——与框架解耦，框架只承担编排层。同时优先选开源、社区活跃、更新稳定的框架，降低被单一厂商锁定或项目停更的风险。",[812,511,592,813,814],"Agent框架","AI智能体开发","框架选型",{},"\u002Fblog\u002Fai-agent\u002Fagent-kuangjia",{"title":437,"description":800},{"loc":816},"blog\u002Fai-agent\u002Fagent-kuangjia",[821,814,431],"Agent","yYDRGbvFeaR227ajWv5qiaN844ebZL6NwOSs4iOxEcE",{"id":824,"title":825,"author":6,"body":826,"category":403,"cover":404,"date":1145,"description":1146,"draft":407,"extension":408,"faq":1147,"featured":407,"image":404,"keywords":1157,"meta":1163,"navigation":425,"path":1164,"seo":1165,"sitemap":1166,"stem":1167,"tags":1168,"updated":1145,"__hash__":1172},"blog\u002Fblog\u002Fai-agent\u002Fai-agent-vs-chatbot.md","AI Agent 和传统聊天机器人有什么区别？别再被忽悠",{"type":8,"value":827,"toc":1136},[828,835,838,852,858,861,963,966,971,984,987,990,993,1007,1010,1014,1017,1049,1052,1098,1102,1105,1131],[11,829,830,831,834],{},"很多企业几年前上过\"智能客服\"，结果发现它\"一点都不智能\"——答非所问、死板的流程树、动不动就\"转人工\"。于是对\"AI 客服\"有阴影。",[15,832,833],{},"但现在的 AI Agent 和当年的聊天机器人，是完全不同的两种东西。"," 这篇文章讲清它们的本质区别，帮你判断企业该上哪种、值不值得换。",[20,836,837],{"id":837},"一句话区分",[27,839,840,846],{},[30,841,842,845],{},[15,843,844],{},"传统聊天机器人","：基于关键词匹配和预设流程图。用户说\"退款\"，它跳到\"退款流程\"；用户换个说法\"我不要了\"，它就懵了。",[30,847,848,851],{},[15,849,850],{},"AI Agent","：基于大语言模型，能理解意图、规划任务、调用工具、记住上下文。用户说\"我上周买的那个订单想退\"，它能自己查订单、判断是否符合退款规则、发起退款流程。",[11,853,854,855],{},"本质区别：",[15,856,857],{},"前者是\"执行固定脚本的程序\"，后者是\"能理解并完成任务的助手\"。",[20,859,860],{"id":860},"六维对比表",[61,862,863,874],{},[64,864,865],{},[67,866,867,870,872],{},[70,868,869],{},"能力",[70,871,844],{},[70,873,850],{},[80,875,876,887,898,909,920,931,942,952],{},[67,877,878,881,884],{},[85,879,880],{},"对话方式",[85,882,883],{},"关键词 + 流程图",[85,885,886],{},"大模型理解自然语言",[67,888,889,892,895],{},[85,890,891],{},"理解能力",[85,893,894],{},"死板，换个说法就懵",[85,896,897],{},"灵活，懂口语、省略、上下文",[67,899,900,903,906],{},[85,901,902],{},"自主性",[85,904,905],{},"被动应答",[85,907,908],{},"主动规划、多步执行",[67,910,911,914,917],{},[85,912,913],{},"工具调用",[85,915,916],{},"几乎不能",[85,918,919],{},"能查库、调 API、发通知、生成文件",[67,921,922,925,928],{},[85,923,924],{},"记忆",[85,926,927],{},"无或当次会话",[85,929,930],{},"长期记忆（记住用户历史）",[67,932,933,936,939],{},[85,934,935],{},"复杂任务",[85,937,938],{},"只能引导转人工",[85,940,941],{},"可独立完成跨系统任务",[67,943,944,946,949],{},[85,945,230],{},[85,947,948],{},"改流程图\u002F规则",[85,950,951],{},"喂新文档即可",[67,953,954,957,960],{},[85,955,956],{},"用户体验",[85,958,959],{},"机械、易激怒用户",[85,961,962],{},"接近真人，体验好",[20,964,965],{"id":965},"真实场景对比",[11,967,968],{},[15,969,970],{},"场景：用户问\"我上周买的那个能退吗\"",[27,972,973,979],{},[30,974,975,978],{},[15,976,977],{},"传统机器人","：抓不到关键词，回\"请问您要咨询什么\"，或跳到通用退款说明页。用户烦躁，转人工。",[30,980,981,983],{},[15,982,850],{},"：理解意图（查订单+判断退款）→ 调用订单系统查到上周的订单 → 判断商品是否符合退款规则 → 回答\"您的订单 xxx 符合 7 天无理由退款，我帮您发起申请，确认吗？\" → 用户确认 → 调用退款 API → 完成。",[11,985,986],{},"整个流程 AI Agent 自己做完，用户感觉\"这个客服真懂\"。",[20,988,989],{"id":989},"为什么传统机器人体验差",[11,991,992],{},"传统机器人的工作机制是\"if 用户说 X，then 回 Y\"。问题在于：",[27,994,995,998,1001,1004],{},[30,996,997],{},"用户的说法千变万化，关键词覆盖不全。",[30,999,1000],{},"业务一复杂，流程图爆炸，维护噩梦。",[30,1002,1003],{},"没法理解上下文，每句都从头开始。",[30,1005,1006],{},"没法真正\"做事\"，只能\"指路\"。",[11,1008,1009],{},"所以传统机器人最后基本都退化成\"转人工的入口\"，用户也学会了\"上来就转人工\"。",[20,1011,1013],{"id":1012},"ai-agent-为什么能真正解决问题","AI Agent 为什么能真正解决问题",[11,1015,1016],{},"AI Agent 的核心能力：",[27,1018,1019,1025,1031,1037,1043],{},[30,1020,1021,1024],{},[15,1022,1023],{},"理解自然语言","：基于大模型，用户怎么说都能懂。",[30,1026,1027,1030],{},[15,1028,1029],{},"调用工具","：接业务系统（CRM、订单、知识库），能真正查数据、改数据。",[30,1032,1033,1036],{},[15,1034,1035],{},"规划任务","：把复杂目标拆成步骤，自己执行。",[30,1038,1039,1042],{},[15,1040,1041],{},"长期记忆","：记住用户历史，体验连贯。",[30,1044,1045,1048],{},[15,1046,1047],{},"知识可更新","：喂新文档即学习，不用改代码。",[20,1050,1051],{"id":1051},"企业该上哪种",[61,1053,1054,1064],{},[64,1055,1056],{},[67,1057,1058,1061],{},[70,1059,1060],{},"场景",[70,1062,1063],{},"建议",[80,1065,1066,1074,1082,1090],{},[67,1067,1068,1071],{},[85,1069,1070],{},"客服量大、问题多样",[85,1072,1073],{},"AI Agent（体验好、省人工）",[67,1075,1076,1079],{},[85,1077,1078],{},"问题高度标准化、简单",[85,1080,1081],{},"传统机器人够用，或 AI Agent 低配版",[67,1083,1084,1087],{},[85,1085,1086],{},"需要跨系统办事（查单、改单、退款）",[85,1088,1089],{},"AI Agent（能调工具）",[67,1091,1092,1095],{},[85,1093,1094],{},"预算极有限、想先试水",[85,1096,1097],{},"AI Agent MVP（先接知识库做问答）",[20,1099,1101],{"id":1100},"怎么从传统机器人升级到-ai-agent","怎么从传统机器人升级到 AI Agent",[11,1103,1104],{},"不用一次性推翻：",[362,1106,1107,1113,1119,1125],{},[30,1108,1109,1112],{},[15,1110,1111],{},"先接知识库","：把产品手册、FAQ、SOP 喂给 AI，让它能准确回答咨询类问题。",[30,1114,1115,1118],{},[15,1116,1117],{},"再接业务系统","：让 AI 能查订单、查库存、发通知，真正\"办事\"。",[30,1120,1121,1124],{},[15,1122,1123],{},"加人工兜底","：AI 把握不大时转人工，持续优化。",[30,1126,1127,1130],{},[15,1128,1129],{},"逐步替换","：AI 能力稳定后，逐步下线老机器人的场景。",[380,1132,1133],{},[11,1134,1135],{},"广州市汉诺雷斯（HNREIS）提供 AI 智能客服定制，支持从传统机器人渐进式升级。告诉我们你现在的客服痛点和量级，我们帮你评估 AI Agent 能解决多少、ROI 多少。",{"title":386,"searchDepth":387,"depth":387,"links":1137},[1138,1139,1140,1141,1142,1143,1144],{"id":837,"depth":387,"text":837},{"id":860,"depth":387,"text":860},{"id":965,"depth":387,"text":965},{"id":989,"depth":387,"text":989},{"id":1012,"depth":387,"text":1013},{"id":1051,"depth":387,"text":1051},{"id":1100,"depth":387,"text":1101},"2024-05-26","传统聊天机器人基于关键词和流程图，只能被动应答；AI Agent 基于大模型，能理解意图、规划任务、调用工具、长期记忆。本文用对比表和真实场景讲清两者的本质区别，帮你判断企业该上哪种。",[1148,1151,1154],{"q":1149,"a":1150},"AI Agent 比传统聊天机器人贵很多吗？","一次性投入是的，但要看性价比。传统聊天机器人便宜（几千到一两万），但只能回答固定问题，用户体验差、转人工率高；AI Agent 投入更高（几万起），但能理解自然语言、调用业务系统、真正解决问题，长期看能省更多人工成本。如果客服量大，AI Agent 的回报更快。",{"q":1152,"a":1153},"我们已经有客服机器人了，要全部换掉吗？","不必全部换。可以分步：先用 AI Agent 接管高频、标准的问题（产品咨询、订单查询、售后流程），复杂问题继续走老系统或人工。我们支持渐进式升级，避免一次性推翻重来。",{"q":1155,"a":1156},"AI Agent 会不会答错或者说胡话？","会，但有办法控制。关键是 RAG（检索增强生成）——让 AI 先在企业知识库里查到准确信息再回答，而不是凭空生成。再配上人工兜底机制（AI 把握不大时转人工）和回答质量回流，准确率可以做到很高。",[1158,1159,1160,1161,1162],"AI Agent和聊天机器人区别","智能客服","AI客服机器人","传统客服机器人","AI智能体",{},"\u002Fblog\u002Fai-agent\u002Fai-agent-vs-chatbot",{"title":825,"description":1146},{"loc":1164},"blog\u002Fai-agent\u002Fai-agent-vs-chatbot",[1169,1170,1171],"对比","AI","概念","8qKIdcO08ytoPMy5eHwEUZP4TpzwR1mih2PoU1dQnm0",{"id":1174,"title":1175,"author":6,"body":1176,"category":403,"cover":404,"date":1514,"description":1515,"draft":407,"extension":408,"faq":1516,"featured":407,"image":404,"keywords":1526,"meta":1531,"navigation":425,"path":1532,"seo":1533,"sitemap":1534,"stem":1535,"tags":1536,"updated":1514,"__hash__":1540},"blog\u002Fblog\u002Fai-agent\u002Fai-caiwu-baobiao.md","AI辅助财务报表和数据分析怎么做",{"type":8,"value":1177,"toc":1495},[1178,1185,1189,1193,1204,1208,1219,1223,1234,1238,1246,1250,1276,1281,1285,1288,1334,1339,1342,1346,1357,1361,1369,1373,1381,1385,1393,1395,1442,1445,1471,1473,1490],[11,1179,1180,1181,1184],{},"财务部门每天和报表、数据打交道，重复分析多、对数据敏感度要求极高。",[15,1182,1183],{},"AI能帮财务提效，但财务数据的特殊性决定了\"怎么用\"比\"用不用\"更重要。"," 这篇讲清AI辅助财务分析的边界、做法和安全要点。",[20,1186,1188],{"id":1187},"ai在财务分析里能做什么","AI在财务分析里能做什么",[141,1190,1192],{"id":1191},"_1-报表解读与可视化分析","1. 报表解读与可视化分析",[27,1194,1195,1198,1201],{},[30,1196,1197],{},"把利润表、资产负债表、现金流量表的关键指标提炼成通俗解读。",[30,1199,1200],{},"自动生成同比、环比、趋势的文字说明。",[30,1202,1203],{},"辅助管理层快速理解\"这份报表说明了什么\"。",[141,1205,1207],{"id":1206},"_2-异常检测","2. 异常检测",[27,1209,1210,1213,1216],{},[30,1211,1212],{},"识别费用、成本、应收应付的异常波动。",[30,1214,1215],{},"标记可能的问题（如某类支出突增、账期异常）。",[30,1217,1218],{},"帮财务聚焦风险点，而不是大海捞针。",[141,1220,1222],{"id":1221},"_3-辅助预测","3. 辅助预测",[27,1224,1225,1228,1231],{},[30,1226,1227],{},"基于历史数据做现金流、收入、成本的短期预测。",[30,1229,1230],{},"结合业务假设做情景分析。",[30,1232,1233],{},"预测是辅助参考，不是承诺。",[141,1235,1237],{"id":1236},"_4-数据问答","4. 数据问答",[27,1239,1240,1243],{},[30,1241,1242],{},"用自然语言问数据：\"上季度哪个产品线毛利率最高\"。",[30,1244,1245],{},"AI转为查询，返回结果——类似 BI 工具的自然语言版。",[20,1247,1249],{"id":1248},"ai不能替代的","AI不能替代的",[27,1251,1252,1258,1264,1270],{},[30,1253,1254,1257],{},[15,1255,1256],{},"记账与出表","：强规则、强合规，应以专业财务软件为准。",[30,1259,1260,1263],{},[15,1261,1262],{},"对外财报","：必须人工编制和审计，AI不可直接生成。",[30,1265,1266,1269],{},[15,1267,1268],{},"专业判断","：税务筹划、投资决策、合规判断，需要财务专业人员。",[30,1271,1272,1275],{},[15,1273,1274],{},"责任承担","：AI出错无人担责，最终决策和签字在人。",[11,1277,1278],{},[15,1279,1280],{},"AI是财务的助手，不是替代者。",[20,1282,1284],{"id":1283},"数据安全财务ai的红线","数据安全：财务AI的红线",[11,1286,1287],{},"财务数据敏感度极高，这是用AI最大的顾虑：",[61,1289,1290,1300],{},[64,1291,1292],{},[67,1293,1294,1297],{},[70,1295,1296],{},"风险点",[70,1298,1299],{},"应对做法",[80,1301,1302,1310,1318,1326],{},[67,1303,1304,1307],{},[85,1305,1306],{},"数据泄露",[85,1308,1309],{},"敏感数据不上公有云，或先脱敏",[67,1311,1312,1315],{},[85,1313,1314],{},"合规要求",[85,1316,1317],{},"涉及上市公司\u002F监管要求，必须私有化",[67,1319,1320,1323],{},[85,1321,1322],{},"模型幻觉",[85,1324,1325],{},"计算和分析结果必须人工复核",[67,1327,1328,1331],{},[85,1329,1330],{},"权限控制",[85,1332,1333],{},"AI应用也要做数据权限隔离",[11,1335,1336],{},[15,1337,1338],{},"核心原则：敏感财务数据优先私有化部署，脱敏后再分析。",[20,1340,1341],{"id":1341},"实施路径",[141,1343,1345],{"id":1344},"_1-先从低风险场景切入","1. 先从低风险场景切入",[27,1347,1348,1351,1354],{},[30,1349,1350],{},"用 AI 解读已公开\u002F已汇总的报表数据。",[30,1352,1353],{},"做趋势分析、异常提示等辅助工作。",[30,1355,1356],{},"不碰原始敏感凭证。",[141,1358,1360],{"id":1359},"_2-数据脱敏与隔离","2. 数据脱敏与隔离",[27,1362,1363,1366],{},[30,1364,1365],{},"分析前对客户名、金额等敏感字段脱敏。",[30,1367,1368],{},"AI 应用与核心财务系统数据隔离。",[141,1370,1372],{"id":1371},"_3-私有化部署","3. 私有化部署",[27,1374,1375,1378],{},[30,1376,1377],{},"有条件的企业部署私有化模型。",[30,1379,1380],{},"数据不出内网。",[141,1382,1384],{"id":1383},"_4-人机协作流程","4. 人机协作流程",[27,1386,1387,1390],{},[30,1388,1389],{},"AI 出初稿和提示，财务复核确认。",[30,1391,1392],{},"关键决策必须人工签字。",[20,1394,309],{"id":309},[61,1396,1397,1407],{},[64,1398,1399],{},[67,1400,1401,1403,1405],{},[70,1402,72],{},[70,1404,320],{},[70,1406,323],{},[80,1408,1409,1420,1431],{},[67,1410,1411,1414,1417],{},[85,1412,1413],{},"轻量分析工具",[85,1415,1416],{},"接入脱敏数据做报表解读",[85,1418,1419],{},"较低",[67,1421,1422,1425,1428],{},[85,1423,1424],{},"私有化AI分析",[85,1426,1427],{},"本地部署模型 + 财务数据集成",[85,1429,1430],{},"中到高",[67,1432,1433,1436,1439],{},[85,1434,1435],{},"完整智能财务中台",[85,1437,1438],{},"多模块 + 权限 + 审计 + 私有化",[85,1440,1441],{},"高，定制开发",[20,1443,1444],{"id":1444},"常见误区",[27,1446,1447,1453,1459,1465],{},[30,1448,1449,1452],{},[15,1450,1451],{},"让AI直接做账出表","：合规风险高，不可取。",[30,1454,1455,1458],{},[15,1456,1457],{},"原始财务数据直接传公有云","：泄露和合规风险。",[30,1460,1461,1464],{},[15,1462,1463],{},"盲目相信AI分析结论","：不复核直接用，可能出错。",[30,1466,1467,1470],{},[15,1468,1469],{},"忽视权限","：AI能看的数据范围要和岗位职责匹配。",[20,1472,360],{"id":360},[362,1474,1475,1478,1481,1484,1487],{},[30,1476,1477],{},"梳理可被AI辅助的重复分析工作。",[30,1479,1480],{},"评估数据敏感度和合规要求。",[30,1482,1483],{},"选私有化或脱敏方案。",[30,1485,1486],{},"从低风险场景试点。",[30,1488,1489],{},"建立人机协作和复核流程。",[380,1491,1492],{},[11,1493,1494],{},"广州市汉诺雷斯（HNREIS）帮企业搭建私有化的AI数据分析应用，在数据安全和合规前提下，让财务和业务团队用自然语言分析数据。把你的数据分析需求告诉我们，我们给出安全合规的方案。",{"title":386,"searchDepth":387,"depth":387,"links":1496},[1497,1503,1504,1505,1511,1512,1513],{"id":1187,"depth":387,"text":1188,"children":1498},[1499,1500,1501,1502],{"id":1191,"depth":395,"text":1192},{"id":1206,"depth":395,"text":1207},{"id":1221,"depth":395,"text":1222},{"id":1236,"depth":395,"text":1237},{"id":1248,"depth":387,"text":1249},{"id":1283,"depth":387,"text":1284},{"id":1341,"depth":387,"text":1341,"children":1506},[1507,1508,1509,1510],{"id":1344,"depth":395,"text":1345},{"id":1359,"depth":395,"text":1360},{"id":1371,"depth":395,"text":1372},{"id":1383,"depth":395,"text":1384},{"id":309,"depth":387,"text":309},{"id":1444,"depth":387,"text":1444},{"id":360,"depth":387,"text":360},"2024-06-03","AI能帮财务做报表解读、异常检测、趋势预测，但财务数据高度敏感。本文讲清AI辅助财务分析能做什么、不能做什么，以及数据安全和私有化的关键考量。",[1517,1520,1523],{"q":1518,"a":1519},"AI能直接帮我做账出报表吗？","目前不适合让AI直接生成对外财报。AI擅长的是辅助：解读已有报表、做数据透视、发现异常、辅助预测分析。记账、出表这类强规则、强合规的工作，仍应以专业财务软件和会计为准，AI作为辅助分析工具。涉及对外披露的报表必须人工复核把关。",{"q":1521,"a":1522},"财务数据这么敏感，能用云端AI吗？","要谨慎。财务数据涉及商业机密和合规要求，一般不建议直接传公有云大模型。可行做法：敏感字段脱敏后再分析、用私有化部署的模型、或只让AI处理汇总后的非敏感数据。涉及上市公司或有严格保密要求的，必须私有化部署。",{"q":1524,"a":1525},"AI做财务分析的准确率可靠吗？","对结构化数据的计算和趋势识别较可靠，但对复杂业务判断和因果分析可能出错（幻觉）。正确用法是AI出初稿和提示，财务专业人员复核确认。AI不能替代财务判断，尤其涉及投资、税务、合规决策时，必须以人为准。",[1527,1528,1529,1530],"AI财务分析","财务报表AI","AI数据分析","智能财务",{},"\u002Fblog\u002Fai-agent\u002Fai-caiwu-baobiao",{"title":1175,"description":1515},{"loc":1532},"blog\u002Fai-agent\u002Fai-caiwu-baobiao",[1537,1538,1539],"财务","数据分析","合规","Pdri7PqftoKRgTfvW1VvvY-liWyZ8b5GUMVIU9UtDsk",{"id":1542,"title":1543,"author":6,"body":1544,"category":403,"cover":404,"date":1876,"description":1877,"draft":407,"extension":408,"faq":1878,"featured":407,"image":404,"keywords":1888,"meta":1893,"navigation":425,"path":1894,"seo":1895,"sitemap":1896,"stem":1897,"tags":1898,"updated":1876,"__hash__":1902},"blog\u002Fblog\u002Fai-agent\u002Fai-canyin.md","餐饮行业怎么用AI做营销和排班",{"type":8,"value":1545,"toc":1862},[1546,1553,1556,1573,1577,1581,1592,1599,1603,1614,1618,1629,1633,1644,1648,1659,1662,1742,1748,1750,1782,1784,1838,1840,1857],[11,1547,1548,1549,1552],{},"餐饮行业人力成本高、客流波动大、点评和营销又特别耗时。",[15,1550,1551],{},"AI帮餐饮解决的不是\"做饭\"，而是那些重复、依赖经验、占用店长精力的环节。"," 这篇讲清餐饮AI的落地场景。",[20,1554,1555],{"id":1555},"餐饮的痛点",[27,1557,1558,1561,1564,1567,1570],{},[30,1559,1560],{},"排班靠店长经验，忙时人不够、闲时人过剩。",[30,1562,1563],{},"客流预测不准，备料和排班都跟着错。",[30,1565,1566],{},"大众点评、外卖平台的评价回复耗时。",[30,1568,1569],{},"朋友圈、社群营销文案写不出来。",[30,1571,1572],{},"会员复购靠拍脑袋，缺少数据支撑。",[20,1574,1576],{"id":1575},"ai能落地的场景","AI能落地的场景",[141,1578,1580],{"id":1579},"_1-客流预测与智能排班","1. 客流预测与智能排班",[27,1582,1583,1586,1589],{},[30,1584,1585],{},"基于历史客流、天气、节假日、周边活动预测各时段客流。",[30,1587,1588],{},"自动生成排班建议，匹配高峰人力需求。",[30,1590,1591],{},"连锁门店可统一调度。",[11,1593,1594,1595,1598],{},"这是餐饮AI",[15,1596,1597],{},"价值最直接","的场景——直接关联人力成本。",[141,1600,1602],{"id":1601},"_2-智能推荐与连带销售","2. 智能推荐与连带销售",[27,1604,1605,1608,1611],{},[30,1606,1607],{},"根据顾客点单、时段、季节推荐搭配。",[30,1609,1610],{},"收银\u002F小程序点餐页的智能推荐位。",[30,1612,1613],{},"提升客单价。",[141,1615,1617],{"id":1616},"_3-点评分析与自动回复","3. 点评分析与自动回复",[27,1619,1620,1623,1626],{},[30,1621,1622],{},"批量分析大众点评\u002F外卖评价，提取菜品、服务、卫生的口碑趋势。",[30,1624,1625],{},"辅助生成评价回复（人工把关后发送）。",[30,1627,1628],{},"发现差评共性问题，改进运营。",[141,1630,1632],{"id":1631},"_4-营销文案生成","4. 营销文案生成",[27,1634,1635,1638,1641],{},[30,1636,1637],{},"朋友圈、社群、公众号文案快速生成。",[30,1639,1640],{},"节日活动、新品上市的推广素材。",[30,1642,1643],{},"多门店内容统一又本地化。",[141,1645,1647],{"id":1646},"_5-会员与私域","5. 会员与私域",[27,1649,1650,1653,1656],{},[30,1651,1652],{},"会员消费数据分析，识别高价值客户和流失风险。",[30,1654,1655],{},"个性化优惠券和召回。",[30,1657,1658],{},"复购预测。",[20,1660,1661],{"id":1661},"落地的现实考量",[61,1663,1664,1679],{},[64,1665,1666],{},[67,1667,1668,1670,1673,1676],{},[70,1669,1060],{},[70,1671,1672],{},"数据依赖",[70,1674,1675],{},"难度",[70,1677,1678],{},"见效",[80,1680,1681,1693,1704,1717,1729],{},[67,1682,1683,1686,1688,1690],{},[85,1684,1685],{},"营销文案生成",[85,1687,90],{},[85,1689,90],{},[85,1691,1692],{},"快",[67,1694,1695,1698,1700,1702],{},[85,1696,1697],{},"点评分析回复",[85,1699,336],{},[85,1701,90],{},[85,1703,1692],{},[67,1705,1706,1709,1712,1714],{},[85,1707,1708],{},"客流预测排班",[85,1710,1711],{},"高（历史数据）",[85,1713,336],{},[85,1715,1716],{},"中期",[67,1718,1719,1722,1725,1727],{},[85,1720,1721],{},"智能推荐",[85,1723,1724],{},"中（点单数据）",[85,1726,336],{},[85,1728,1716],{},[67,1730,1731,1734,1736,1739],{},[85,1732,1733],{},"会员复购预测",[85,1735,243],{},[85,1737,1738],{},"中高",[85,1740,1741],{},"中长期",[11,1743,1744,1747],{},[15,1745,1746],{},"先从低门槛、见效快的场景切入","（文案、点评），再逐步做需要数据的场景（排班、推荐）。",[20,1749,274],{"id":274},[27,1751,1752,1758,1764,1770,1776],{},[30,1753,1754,1757],{},[15,1755,1756],{},"数据质量差就上预测","：历史客流数据不准，预测也不准。",[30,1759,1760,1763],{},[15,1761,1762],{},"自动回复不经审核","：AI回复可能不当，引发公关问题。",[30,1765,1766,1769],{},[15,1767,1768],{},"忽视门店执行","：AI给排班建议，门店不执行等于零。",[30,1771,1772,1775],{},[15,1773,1774],{},"一上来搞大而全","：建议单点突破，验证后再扩展。",[30,1777,1778,1781],{},[15,1779,1780],{},"把AI当万能","：餐饮核心仍是产品和体验，AI是提效工具。",[20,1783,309],{"id":309},[61,1785,1786,1796],{},[64,1787,1788],{},[67,1789,1790,1792,1794],{},[70,1791,72],{},[70,1793,320],{},[70,1795,323],{},[80,1797,1798,1808,1818,1827],{},[67,1799,1800,1803,1806],{},[85,1801,1802],{},"文案\u002F点评辅助",[85,1804,1805],{},"SaaS 或 API，按量计费",[85,1807,90],{},[67,1809,1810,1813,1816],{},[85,1811,1812],{},"客流排班模块",[85,1814,1815],{},"和门店系统打通，定制",[85,1817,336],{},[67,1819,1820,1822,1825],{},[85,1821,1721],{},[85,1823,1824],{},"点餐\u002F收银集成",[85,1826,336],{},[67,1828,1829,1832,1835],{},[85,1830,1831],{},"连锁智能运营中台",[85,1833,1834],{},"多门店 + 预测 + 推荐 + 私域",[85,1836,1837],{},"高，定制",[20,1839,360],{"id":360},[362,1841,1842,1845,1848,1851,1854],{},[30,1843,1844],{},"找出最耗时的环节（排班？点评？文案？）。",[30,1846,1847],{},"从低成本场景试点（文案、点评辅助）。",[30,1849,1850],{},"积累门店数据，再做客流预测排班。",[30,1852,1853],{},"逐步扩展到推荐、会员。",[30,1855,1856],{},"始终让人把关关键输出。",[380,1858,1859],{},[11,1860,1861],{},"广州市汉诺雷斯（HNREIS）帮餐饮企业落地AI应用，从客流预测排班、点评分析到营销文案和会员私域，和小程序、收银系统打通。把你的门店运营痛点告诉我们，我们给出务实的AI落地方案。",{"title":386,"searchDepth":387,"depth":387,"links":1863},[1864,1865,1872,1873,1874,1875],{"id":1555,"depth":387,"text":1555},{"id":1575,"depth":387,"text":1576,"children":1866},[1867,1868,1869,1870,1871],{"id":1579,"depth":395,"text":1580},{"id":1601,"depth":395,"text":1602},{"id":1616,"depth":395,"text":1617},{"id":1631,"depth":395,"text":1632},{"id":1646,"depth":395,"text":1647},{"id":1661,"depth":387,"text":1661},{"id":274,"depth":387,"text":274},{"id":309,"depth":387,"text":309},{"id":360,"depth":387,"text":360},"2024-06-16","餐饮门店客流波动大、排班靠经验、点评回复耗时。AI能帮餐饮做客流预测排班、智能推荐、点评分析和营销文案，本文讲清落地场景与成本。",[1879,1882,1885],{"q":1880,"a":1881},"餐饮店有必要上AI吗？","看规模和痛点。单店小店，痛点不突出可以先不上；连锁或多门店，客流预测排班、点评批量回复、营销文案生成这类重复工作，AI能明显提效。建议从最耗时的环节切入，比如用AI生成朋友圈\u002F社群文案、辅助排班，先看到效果再扩展。",{"q":1883,"a":1884},"AI排班靠谱吗？","AI排班是基于历史客流、天气、节假日等数据预测各时段客流，再匹配人力需求，比纯靠店长经验更稳定。但它依赖历史数据质量和门店执行，适合作为排班参考，最终仍需店长结合实际情况调整。新店数据不足时预测效果有限。",{"q":1886,"a":1887},"餐饮用AI要花多少钱？","轻量场景（文案生成、点评回复辅助）用现有SaaS或大模型API，成本很低；客流预测排班、智能推荐这类要和门店系统打通的，属于定制开发，几万到十几万不等，看门店数量和集成深度。建议先从低成本场景验证价值。",[1889,1890,1891,1892],"餐饮AI","餐饮排班AI","AI营销文案","餐饮数字化",{},"\u002Fblog\u002Fai-agent\u002Fai-canyin",{"title":1543,"description":1877},{"loc":1894},"blog\u002Fai-agent\u002Fai-canyin",[1899,1900,1901],"餐饮","营销","排班","wwPJmcDR-PabsSXv_ZcpsIAe0cgabBUgBGYdFdSCdPc",1781688906715]