[{"data":1,"prerenderedAt":1668},["ShallowReactive",2],{"blog-/blog/enterprise-rag-guide":3,"blog-related-/blog/enterprise-rag-guide":530},{"id":4,"title":5,"author":6,"body":7,"category":513,"cover":514,"date":515,"description":516,"extension":517,"meta":518,"navigation":519,"path":520,"readingTime":521,"seo":522,"stem":523,"tags":524,"__hash__":529},"blog/blog/enterprise-rag-guide.md","企业知识库 RAG 实战：从文档到 AI 问答的 5 个关键步骤","仙宫云技术团队",{"type":8,"value":9,"toc":490},"minimark",[10,19,24,30,36,103,109,113,116,121,146,150,153,173,177,180,184,187,191,211,215,261,265,272,293,296,300,303,313,318,329,333,336,340,343,363,367,375,379,382,396,400,432,436,439,465,473,476],[11,12,13,14,18],"p",{},"\"我们公司有几万份 Word/PDF 文档，想做一个 AI 问答助手，新员工有问题直接问就能拿答案，可不可行？\"——这是仙宫云客户最高频的需求之一。答案是肯定的，技术路径就是 ",[15,16,17],"strong",{},"RAG（Retrieval-Augmented Generation，检索增强生成）","。本文拆解从 0 到 1 的 5 个关键步骤。",[20,21,23],"h2",{"id":22},"一rag-是什么为什么不直接微调模型","一、RAG 是什么？为什么不直接微调模型？",[11,25,26,29],{},[15,27,28],{},"RAG 的核心思想","：用户提问 → 先从企业文档库检索最相关的几段内容 → 把这些内容作为上下文交给大模型 → 大模型基于上下文生成回答。",[11,31,32,35],{},[15,33,34],{},"对比微调（Fine-tuning）","，RAG 有三个企业级优势：",[37,38,39,55],"table",{},[40,41,42],"thead",{},[43,44,45,49,52],"tr",{},[46,47,48],"th",{},"维度",[46,50,51],{},"RAG",[46,53,54],{},"微调",[56,57,58,70,81,92],"tbody",{},[43,59,60,64,67],{},[61,62,63],"td",{},"知识更新",[61,65,66],{},"改文档即可",[61,68,69],{},"需要重新训练",[43,71,72,75,78],{},[61,73,74],{},"成本",[61,76,77],{},"低（无需 GPU 训练）",[61,79,80],{},"高（数据 + 算力）",[43,82,83,86,89],{},[61,84,85],{},"可追溯",[61,87,88],{},"答案能引用原文",[61,90,91],{},"黑盒输出",[43,93,94,97,100],{},[61,95,96],{},"数据安全",[61,98,99],{},"文档保留在向量库",[61,101,102],{},"知识被吸收进权重",[11,104,105,108],{},[15,106,107],{},"结论","：90% 的企业知识库场景，用 RAG 比微调更合适。",[20,110,112],{"id":111},"二step-1文档预处理最容易被低估的环节","二、Step 1：文档预处理（最容易被低估的环节）",[11,114,115],{},"垃圾进，垃圾出。RAG 效果上限被这一步决定。",[117,118,120],"h3",{"id":119},"_21-文档收集与格式统一","2.1 文档收集与格式统一",[122,123,124,128,131],"ul",{},[125,126,127],"li",{},"收集来源：Word、PDF、PPT、Markdown、Confluence、邮件归档",[125,129,130],{},"统一转 Markdown 或纯文本，保留标题层级",[125,132,133,134,138,139,138,142,145],{},"工具推荐：",[135,136,137],"code",{},"unstructured","、",[135,140,141],{},"Docling",[135,143,144],{},"MinerU","（中文 PDF 表现好）",[117,147,149],{"id":148},"_22-切片chunking策略","2.2 切片（Chunking）策略",[11,151,152],{},"切片大小直接影响检索精度：",[122,154,155,161,167],{},[125,156,157,160],{},[15,158,159],{},"太大","（>1500 字）：检索粒度粗，无关内容多",[125,162,163,166],{},[15,164,165],{},"太小","（\u003C200 字）：上下文不完整，模型无法理解",[125,168,169,172],{},[15,170,171],{},"推荐","：500-800 字 + 50-100 字重叠（overlap）",[117,174,176],{"id":175},"_23-元数据标注","2.3 元数据标注",[11,178,179],{},"每个切片附加元数据：来源文档、章节、更新日期、部门、权限等级。这些字段在检索阶段可以做过滤，比如\"只查财务部 2025 年之后的制度\"。",[20,181,183],{"id":182},"三step-2向量化与向量数据库","三、Step 2：向量化与向量数据库",[11,185,186],{},"把文本切片转成向量，让\"语义相似度\"可以被计算。",[117,188,190],{"id":189},"_31-中文-embedding-模型推荐","3.1 中文 Embedding 模型推荐",[122,192,193,199,205],{},[125,194,195,198],{},[15,196,197],{},"bge-m3","（智源）：多语言、长文本、目前中文综合最佳",[125,200,201,204],{},[15,202,203],{},"text2vec-base-chinese","：轻量，适合资源有限场景",[125,206,207,210],{},[15,208,209],{},"OpenAI text-embedding-3-large","：闭源但效果稳定（数据出域慎用）",[117,212,214],{"id":213},"_32-向量数据库选型","3.2 向量数据库选型",[37,216,217,227],{},[40,218,219],{},[43,220,221,224],{},[46,222,223],{},"数据库",[46,225,226],{},"适用场景",[56,228,229,237,245,253],{},[43,230,231,234],{},[61,232,233],{},"Milvus",[61,235,236],{},"大规模（千万级以上向量），生产首选",[43,238,239,242],{},[61,240,241],{},"Qdrant",[61,243,244],{},"中小规模，部署简单，过滤能力强",[43,246,247,250],{},[61,248,249],{},"Chroma",[61,251,252],{},"POC 验证、小团队",[43,254,255,258],{},[61,256,257],{},"PostgreSQL + pgvector",[61,259,260],{},"已有 PG 基础设施，向量量级 100 万以内",[20,262,264],{"id":263},"四step-3检索策略决定准确率的关键","四、Step 3：检索策略（决定准确率的关键）",[11,266,267,268,271],{},"只用向量相似度（dense retrieval）远远不够。生产级 RAG 一定要做 ",[15,269,270],{},"混合检索","：",[273,274,275,281,287],"ol",{},[125,276,277,280],{},[15,278,279],{},"向量检索","：找语义相似的切片",[125,282,283,286],{},[15,284,285],{},"关键词检索（BM25）","：找精确匹配关键词的切片",[125,288,289,292],{},[15,290,291],{},"重排（Rerank）","：用 bge-reranker 等模型对 Top-20 结果重新打分，取 Top-5",[11,294,295],{},"加上 Rerank 后准确率通常能再提升 15-25%，是性价比最高的优化点。",[20,297,299],{"id":298},"五step-4prompt-设计","五、Step 4：Prompt 设计",[11,301,302],{},"RAG 的 Prompt 模板看似简单，细节决定效果：",[304,305,310],"pre",{"className":306,"code":308,"language":309},[307],"language-text","你是企业知识助手。请严格基于以下\"参考资料\"回答用户问题。\n\n要求：\n1. 答案必须来自参考资料，不要编造\n2. 如果资料中没有相关信息，明确说\"知识库中暂无相关内容\"\n3. 回答末尾标注引用的来源文档\n\n参考资料：\n{retrieved_chunks}\n\n用户问题：{question}\n","text",[135,311,308],{"__ignoreMap":312},"",[11,314,315,271],{},[15,316,317],{},"反幻觉的三个关键约束",[122,319,320,323,326],{},[125,321,322],{},"明确\"必须基于资料\"",[125,324,325],{},"给出\"无答案\"的退出路径",[125,327,328],{},"强制引用来源（用户也能验证）",[20,330,332],{"id":331},"六step-5效果评估与迭代","六、Step 5：效果评估与迭代",[11,334,335],{},"很多企业上线 RAG 后没有评估机制，导致问题积累、用户流失。建议建立三层评估：",[117,337,339],{"id":338},"_61-离线评估","6.1 离线评估",[11,341,342],{},"构建 100-500 条测试问答对，定期跑：",[122,344,345,351,357],{},[125,346,347,350],{},[15,348,349],{},"召回率","：相关切片是否在检索结果 Top-K 中",[125,352,353,356],{},[15,354,355],{},"答案准确率","：人工或大模型评分",[125,358,359,362],{},[15,360,361],{},"拒答率","：无答案问题是否正确拒答",[117,364,366],{"id":365},"_62-在线监控","6.2 在线监控",[122,368,369,372],{},[125,370,371],{},"记录每个问答的：query、检索结果、最终答案、用户反馈（赞/踩）",[125,373,374],{},"重点关注被踩的问答，定位是检索失败还是生成失败",[117,376,378],{"id":377},"_63-持续优化循环","6.3 持续优化循环",[11,380,381],{},"每周/每月迭代一次：",[122,383,384,387,390,393],{},[125,385,386],{},"补充缺失文档",[125,388,389],{},"调整切片策略",[125,391,392],{},"优化 Prompt",[125,394,395],{},"升级 Embedding/Rerank 模型",[20,397,399],{"id":398},"七企业-rag-落地的常见误区","七、企业 RAG 落地的常见误区",[273,401,402,408,414,420,426],{},[125,403,404,407],{},[15,405,406],{},"以为上线就完事","：RAG 是持续运营产品，不是一次性项目",[125,409,410,413],{},[15,411,412],{},"只用单一检索","：纯向量检索准确率上限低",[125,415,416,419],{},[15,417,418],{},"忽略权限控制","：财务文档不能让所有员工查到",[125,421,422,425],{},[15,423,424],{},"没做引用展示","：用户无法验证答案，信任度低",[125,427,428,431],{},[15,429,430],{},"没建反馈闭环","：不知道哪里错、怎么改",[20,433,435],{"id":434},"八仙宫云的企业知识库方案","八、仙宫云的企业知识库方案",[11,437,438],{},"仙宫云提供从大模型私有化部署到 RAG 应用的完整服务：",[122,440,441,447,453,459],{},[125,442,443,446],{},[15,444,445],{},"场景调研","：识别哪些文档值得做、用户高频问题摸底",[125,448,449,452],{},[15,450,451],{},"数据治理","：文档清洗、敏感信息脱敏、权限分级",[125,454,455,458],{},[15,456,457],{},"技术实施","：私有化部署 + Embedding 模型 + 向量库 + 应用界面",[125,460,461,464],{},[15,462,463],{},"持续运营","：评估体系建设、效果迭代、新场景扩展",[11,466,467,472],{},[468,469,471],"a",{"href":470},"/contact","联系我们","获取企业知识库免费方案评估。",[474,475],"hr",{},[11,477,478,271,481,485,486],{},[15,479,480],{},"相关阅读",[468,482,484],{"href":483},"/blog/deepseek-private-deployment-guide","DeepSeek 私有化部署完整指南"," | ",[468,487,489],{"href":488},"/blog/traditional-enterprise-ai-challenges","传统企业 AI 落地的真实困境",{"title":312,"searchDepth":491,"depth":491,"links":492},2,[493,494,500,504,505,506,511,512],{"id":22,"depth":491,"text":23},{"id":111,"depth":491,"text":112,"children":495},[496,498,499],{"id":119,"depth":497,"text":120},3,{"id":148,"depth":497,"text":149},{"id":175,"depth":497,"text":176},{"id":182,"depth":491,"text":183,"children":501},[502,503],{"id":189,"depth":497,"text":190},{"id":213,"depth":497,"text":214},{"id":263,"depth":491,"text":264},{"id":298,"depth":491,"text":299},{"id":331,"depth":491,"text":332,"children":507},[508,509,510],{"id":338,"depth":497,"text":339},{"id":365,"depth":497,"text":366},{"id":377,"depth":497,"text":378},{"id":398,"depth":491,"text":399},{"id":434,"depth":491,"text":435},"AI 应用",null,"2026-04-20","深入讲解企业知识库 RAG（检索增强生成）的落地路径，包含文档预处理、向量化、检索策略、Prompt 设计、效果评估全流程。","md",{},true,"/blog/enterprise-rag-guide",10,{"title":5,"description":516},"blog/enterprise-rag-guide",[51,525,526,527,528],"企业知识库","向量数据库","大模型应用","智能问答","Y0G6dTIxg0ImOVUAy2MlgEoJR7edaGj3_G8mTRjzN5U",[531,785,1174],{"id":532,"title":533,"author":534,"body":535,"category":771,"cover":514,"date":772,"description":773,"extension":517,"meta":774,"navigation":519,"path":775,"readingTime":776,"seo":777,"stem":778,"tags":779,"__hash__":784},"blog/blog/enterprise-ai-private-deployment-mainstream-2026.md","60% 的企业已选择 AI 私有化部署，传统企业还在等什么？","仙宫云团队",{"type":8,"value":536,"toc":764},[537,540,543,546,552,556,559,570,573,583,593,599,602,605,608,614,620,626,630,636,641,644,649,656,661,664,689,694,697,701,706,709,712,715,717,722,725,728,731,733,737,740,743,749],[11,538,539],{},"去年，广东某制造集团的 IT 负责人找到我们，说起公司错失的半年时间。",[11,541,542],{},"2025 年初，集团高层已达成共识：要引入大模型，解决技术文档检索效率低的老大难问题。但方案迟迟没落地——有人担心员工把图纸上传到云端 AI，数据泄露怎么办？有人觉得国产模型还不成熟，再等等。等来等去，半年过去了，竞争对手的 AI 知识库已经上线三个月，新人培训周期从两个月缩短到不足一个月。",[11,544,545],{},"这家企业的遭遇并不罕见。",[11,547,548,549],{},"今天，摆在传统企业面前的问题，早已不是\"要不要做 AI\"，而是：",[15,550,551],{},"怎么做，才能安全？",[20,553,555],{"id":554},"行业拐点私有化部署正在成为标配","行业拐点：私有化部署正在成为标配",[11,557,558],{},"2026 年，企业 AI 落地进入了一个新阶段。",[11,560,561,562,565,566,569],{},"根据多份行业研究，目前国内企业级 AI 部署中，",[15,563,564],{},"私有化部署占比已超过 60%","，成为金融、政务、制造等核心领域的首选模式。中国大模型市场规模预计今年突破 ",[15,567,568],{},"700 亿元","，其中相当大比例来自企业私有化部署和定制化服务。",[11,571,572],{},"是什么让私有化部署从\"少数人的选择\"变成主流？三个力量同时在推动：",[11,574,575,578,579,582],{},[15,576,577],{},"监管压力。"," 2026 年 1 月 1 日起施行的《网络安全法》修订版，首次将 AI 安全写入法条。数据跨境管控更严，关键信息基础设施数据须境内存储，违规最高罚款从 50 万元上升至 ",[15,580,581],{},"1000 万元","。对于金融、医疗、政务等企业，把数据传到第三方云端 AI 服务，已经不只是技术风险，更是合规风险。",[11,584,585,588,589,592],{},[15,586,587],{},"硬件成本大幅下降。"," 过去谈大模型私有化，很多企业第一反应是\"买不起 GPU\"。但量化压缩技术的成熟改变了这一局面——通过 INT4/INT8 混合精度量化，",[15,590,591],{},"在普通服务器上即可运行百亿参数级别的大模型","，硬件成本相比两年前降低了约 70%。私有化部署的门槛，已经不是大企业的专属。",[11,594,595,598],{},[15,596,597],{},"开源模型生态成熟。"," DeepSeek-V3、Qwen3、ChatGLM 等国产开源大模型持续迭代，综合能力已可媲美商用模型。企业无需从零训练一个专属大模型，基于成熟开源模型做私有化部署和微调，既省钱又省时间。",[11,600,601],{},"三个条件同时具备，这就是为什么 2026 年私有化部署进入爆发期。",[20,603,604],{"id":604},"企业落地的三道坎",[11,606,607],{},"尽管大环境已经成熟，但很多企业依然卡在起步阶段。根据我们服务超过 50 家企业的经验，落地难通常集中在三个环节：",[11,609,610,613],{},[15,611,612],{},"选型难。"," DeepSeek、Qwen、Llama、ChatGLM……国内外可用的开源大模型超过数十个，每隔几个月就有新版本发布。企业 IT 团队往往花了大量时间研究模型参数，却不知道哪个适合自己的业务场景。制造业和零售业对模型的要求完全不同，同样是\"客服机器人\"，对话轮次、工具调用、多语言支持的侧重也各有差异。没有专业的场景分析，选型很容易走弯路。",[11,615,616,619],{},[15,617,618],{},"部署难。"," 模型选好了，怎么在公司服务器上跑起来？vLLM、Ollama、SGLang 这些推理框架怎么配置？显存不够用了怎么量化？RAG 检索系统怎么和内部文档对接？这些问题，需要算法工程师、运维工程师协作配合，对于大多数传统企业而言，这样的技术团队根本不存在。",[11,621,622,625],{},[15,623,624],{},"用起来难。"," 这是最容易被低估的一环。很多企业花钱买了咨询方案，把模型部署好了，却发现员工不会用，或者系统跟实际业务流程脱节，最终成了摆设。AI 落地不是\"跑通 Demo\"，而是要真正嵌入采购、客服、研发、运营等具体流程，才能产生可衡量的 ROI。",[20,627,629],{"id":628},"仙宫云的解法陪企业拿到结果","仙宫云的解法：陪企业拿到结果",[11,631,632,633],{},"仙宫云成立于 2021 年，从企业级软件起家，2023 年完成 AI 战略转型。我们不卖单纯的\"部署服务\"，而是提供从场景调研到员工培训的端到端交付——",[15,634,635],{},"因为我们知道，模型跑通只是开始。",[11,637,638],{},[15,639,640],{},"第一步：场景调研与选型",[11,642,643],{},"在正式报价之前，我们会先和企业做 2–4 小时的业务调研，梳理哪些场景 AI 价值最高、ROI 最快。然后横向评估 DeepSeek、Qwen、Llama 等主流模型，结合企业硬件现状，给出选型建议和 ROI 测算。企业不需要懂模型，只需要讲清楚业务目标。",[11,645,646],{},[15,647,648],{},"第二步：私有化部署与推理优化",[11,650,651,652,655],{},"在企业本地服务器或私有云完成全套部署，数据 ",[15,653,654],{},"100% 不出企业网络","。我们会做量化、并行推理、缓存等性能优化，确保模型响应速度满足业务需求。同时配合企业完成等保对接、数据加密和权限管理，满足金融、医疗等行业的合规要求。",[11,657,658],{},[15,659,660],{},"第三步：业务应用开发",[11,662,663],{},"部署完成后，我们基于企业实际业务开发应用层：",[122,665,666,671,677,683],{},[125,667,668,670],{},[15,669,525],{},"：打通内部文档、手册、规范，支持自然语言检索，实现\"问就有答\"",[125,672,673,676],{},[15,674,675],{},"智能客服","：多轮对话、工单生成、跨系统数据查询，7×24 小时响应",[125,678,679,682],{},[15,680,681],{},"AI Agent","：业务流程自动化，多工具调用，解放重复性人工操作",[125,684,685,688],{},[15,686,687],{},"文档智能","：合同审阅、报告生成、单据解析，提升文字处理效率",[11,690,691],{},[15,692,693],{},"第四步：培训与持续陪跑",[11,695,696],{},"上线不是终点。我们提供管理层赋能培训、员工 Prompt 技巧培训，以及季度运营复盘。新模型发布时，我们会第一时间评估是否值得升级并给出建议。这个阶段，我们更像企业的 AI 内部顾问，而不是外包供应商。",[20,698,700],{"id":699},"数字说话两个已落地的案例","数字说话：两个已落地的案例",[11,702,703],{},[15,704,705],{},"某大型制造集团（5000+ 员工）",[11,707,708],{},"痛点：技术文档分散，工程师查阅资料费时费力，新员工培训周期长达两个月。",[11,710,711],{},"方案：DeepSeek-R1 本地部署 + 工艺文档 / 设备手册 / 质检规范知识库，数据 100% 私有化存储。",[11,713,714],{},"成效：技术资料检索效率提升 60%，新人培训周期缩短 50%，3 个月完成交付。",[474,716],{},[11,718,719],{},[15,720,721],{},"某金融科技公司",[11,723,724],{},"痛点：信贷审批依赖人工，风险识别准确率不稳定，合规审计耗时长。",[11,726,727],{},"方案：私有化大模型 + 信用评估 / 欺诈检测 / 反洗钱应用，满足等保三级合规要求。",[11,729,730],{},"成效：风险识别准确率达到 95% 以上，审批效率提升 80%，5 个月完成交付并上线。",[474,732],{},[20,734,736],{"id":735},"现在入场正当其时","现在入场，正当其时",[11,738,739],{},"Gartner 预测，2026 年将有 40% 的企业级应用嵌入 AI 智能体——而一年前，这个比例还不到 5%。这不是一个缓慢的趋势，而是一次快速的分水岭。",[11,741,742],{},"硬件成本已经降到足够低，开源模型已经成熟到足够好，监管框架已经明确到可以照着执行。此刻入场，三个条件同时具备，再等下去，等来的只有竞争对手扩大的领先优势。",[11,744,745,748],{},[15,746,747],{},"仙宫云目前开放企业 AI 场景免费诊断服务。"," 无论你的企业还在观望、刚开始选型、还是已经部署遇到困难，欢迎通过邮件或企业微信联系我们，2 小时内响应，给出可落地的分析建议。",[122,750,751,758,761],{},[125,752,753,754],{},"商务邮箱：",[468,755,757],{"href":756},"mailto:info@xiangong.net","info@xiangong.net",[125,759,760],{},"地址：广东省东莞市",[125,762,763],{},"工作时间：周一至周六 9:00–18:00",{"title":312,"searchDepth":491,"depth":491,"links":765},[766,767,768,769,770],{"id":554,"depth":491,"text":555},{"id":604,"depth":491,"text":604},{"id":628,"depth":491,"text":629},{"id":699,"depth":491,"text":700},{"id":735,"depth":491,"text":736},"行业洞察","2026-05-07","2026年大模型市场规模预计突破700亿，私有化部署占比已超60%。本文结合最新行业数据，解析企业AI落地三大难题与仙宫云的端到端解法。",{},"/blog/enterprise-ai-private-deployment-mainstream-2026",5,{"title":533,"description":773},"blog/enterprise-ai-private-deployment-mainstream-2026",[780,781,782,783,525],"企业AI私有化部署","大模型私有化","DeepSeek","AI落地","t02CWcLmcOQK7_BnGYzlVkBcCjPDZQ7C8ql_iUaxhUU",{"id":786,"title":787,"author":6,"body":788,"category":771,"cover":514,"date":1161,"description":1162,"extension":517,"meta":1163,"navigation":519,"path":488,"readingTime":1164,"seo":1165,"stem":1166,"tags":1167,"__hash__":1173},"blog/blog/traditional-enterprise-ai-challenges.md","传统企业 AI 落地为什么这么难？三个真实案例分析",{"type":8,"value":789,"toc":1137},[790,797,801,804,811,815,818,829,832,838,849,853,856,888,898,902,905,908,911,917,920,931,934,937,966,969,974,978,981,984,987,990,1004,1007,1014,1040,1043,1052,1056,1060,1063,1067,1070,1074,1081,1085,1088,1114,1120,1126,1128],[11,791,792,793,796],{},"DeepSeek 火了之后，老板们都在问\"我们怎么用 AI\"。但真正动手做的传统企业里，",[15,794,795],{},"有 70% 的项目在 6 个月内被搁置","。原因不是 AI 不行，而是落地路径选错了。本文用三个真实案例（已脱敏），讲清楚传统企业 AI 落地的难点和正确姿势。",[20,798,800],{"id":799},"案例一某制造集团的ai-客服折戟","案例一：某制造集团的\"AI 客服\"折戟",[117,802,803],{"id":803},"背景",[11,805,806,807,810],{},"一家年产值 50 亿的工业设备制造商，2024 年初决定上 AI。老板的诉求很明确：",[15,808,809],{},"\"我看别人都做 AI 客服，我们也来一套\"","。",[117,812,814],{"id":813},"第一次尝试失败","第一次尝试（失败）",[11,816,817],{},"公司 IT 部门花 30 万买了某 SaaS 厂商的\"通用 AI 客服\"。3 个月后下线，原因：",[122,819,820,823,826],{},[125,821,822],{},"客户问的都是\"3 号轴承能不能配 5 号设备\"这种专业问题",[125,824,825],{},"通用模型完全答不上来，只会说\"建议联系人工客服\"",[125,827,828],{},"客户体验比之前的电话客服还差",[117,830,831],{"id":831},"失败的根本原因",[11,833,834,837],{},[15,835,836],{},"\"AI 客服\"是结果，不是起点。"," 老板只看到别人有 AI 客服，没看到背后需要：",[122,839,840,843,846],{},[125,841,842],{},"完整的产品知识库（这家公司的产品手册散落在 5 个部门的硬盘里）",[125,844,845],{},"历史客服对话数据（之前都用电话，根本没数字化）",[125,847,848],{},"业务规则梳理（哪些问题可以自动回，哪些必须转人工）",[117,850,852],{"id":851},"第二次尝试成功","第二次尝试（成功）",[11,854,855],{},"仙宫云接手后，重新规划路径：",[273,857,858,864,870,876,882],{},[125,859,860,863],{},[15,861,862],{},"第 1 个月","：整理产品手册，建私有知识库",[125,865,866,869],{},[15,867,868],{},"第 2 个月","：部署 DeepSeek-32B + RAG，先给内部销售工程师用",[125,871,872,875],{},[15,873,874],{},"第 3 个月","：收集 500+ 真实问题，迭代 Prompt 和切片策略",[125,877,878,881],{},[15,879,880],{},"第 4 个月","：开放给经销商，验证准确率达 85%",[125,883,884,887],{},[15,885,886],{},"第 6 个月","：上线 C 端客服",[11,889,890,893,894,897],{},[15,891,892],{},"关键洞察","：传统企业上 AI，",[15,895,896],{},"先做内部工具，再做对外应用","。内部用户容忍度高，是 AI 应用最好的 POC 场景。",[20,899,901],{"id":900},"案例二某连锁零售的ai-推荐失灵","案例二：某连锁零售的\"AI 推荐失灵\"",[117,903,803],{"id":904},"背景-1",[11,906,907],{},"300+ 门店连锁餐饮品牌，想做\"AI 个性化菜品推荐\"。第三方乙方报价 80 万，承诺三个月上线。",[117,909,910],{"id":910},"失败点",[11,912,913,914,810],{},"3 个月后系统上线，但经理反馈：",[15,915,916],{},"推荐的菜还不如收银员根据天气和时段拍脑袋的准",[11,918,919],{},"复盘发现：",[122,921,922,925,928],{},[125,923,924],{},"训练数据只有近 6 个月销售记录，没有节假日、天气、促销变量",[125,926,927],{},"\"推荐\"这个动作没有融入门店实际运营流程，店员根本不看",[125,929,930],{},"没有 A/B 测试机制，无法证明\"AI 推荐 vs 人工推荐\"哪个更好",[117,932,933],{"id":933},"改进路径",[11,935,936],{},"仙宫云重新介入，把\"AI 推荐\"拆成三个更小的问题：",[273,938,939,948,957],{},[125,940,941,944,945],{},[15,942,943],{},"新菜上市预测","：基于历史数据预测某门店上新菜的销量，",[15,946,947],{},"辅助采购",[125,949,950,953,954],{},[15,951,952],{},"库存预警","：哪些菜品在哪些门店即将售罄/滞销，",[15,955,956],{},"辅助调拨",[125,958,959,962,963],{},[15,960,961],{},"门店选址洞察","：开新店时基于周边数据生成评估报告，",[15,964,965],{},"辅助决策",[11,967,968],{},"这三个场景都是\"AI 给建议，人做决策\"，门店运营效率提升 18%，年化收益约 2400 万。",[11,970,971,973],{},[15,972,892],{},"：传统行业不要追求\"AI 替代人\"，先做\"AI 辅助决策\"。决策权留给业务人员，反而推广得更顺。",[20,975,977],{"id":976},"案例三某三甲医院的合规死局","案例三：某三甲医院的合规死局",[117,979,803],{"id":980},"背景-2",[11,982,983],{},"某省级三甲医院想做 AI 病历助手，提升医生写病历效率（医生抱怨写病历占 30% 工作时间）。",[117,985,986],{"id":986},"三个月没动起来",[11,988,989],{},"不是技术问题，是数据问题：",[122,991,992,998,1001],{},[125,993,994,995],{},"病历是核心医疗数据，根据《医疗机构病历管理规定》和等保三级要求，",[15,996,997],{},"绝对不能上公有云",[125,999,1000],{},"院内 IT 团队没有大模型经验",[125,1002,1003],{},"厂商方案要求开放外网，被信息科一票否决",[117,1005,1006],{"id":1006},"解决方案",[11,1008,1009,1010,1013],{},"仙宫云的方案完全围绕\"",[15,1011,1012],{},"数据不出院","\"设计：",[273,1015,1016,1022,1028,1034],{},[125,1017,1018,1021],{},[15,1019,1020],{},"硬件","：在医院信息中心机房部署 2× A100 GPU 服务器",[125,1023,1024,1027],{},[15,1025,1026],{},"模型","：本地化部署 DeepSeek-R1-Distill-Qwen-32B + 医疗领域微调",[125,1029,1030,1033],{},[15,1031,1032],{},"应用","：与院内 HIS/EMR 系统对接，医生在熟悉的系统里使用 AI 辅助",[125,1035,1036,1039],{},[15,1037,1038],{},"合规","：通过院内信息安全审计、审计日志全留痕",[11,1041,1042],{},"效果：医生病历书写时间减少 50%，3 个月内全院推广。",[11,1044,1045,1047,1048,1051],{},[15,1046,892],{},"：合规要求严格的行业（金融、医疗、政务、能源），",[15,1049,1050],{},"私有化部署不是可选项，是必选项","。任何方案绕不过这一条。",[20,1053,1055],{"id":1054},"总结传统企业-ai-落地的三条铁律","总结：传统企业 AI 落地的三条铁律",[117,1057,1059],{"id":1058},"_1-不要从我要做-x开始要从我想解决-y开始","1. 不要从\"我要做 X\"开始，要从\"我想解决 Y\"开始",[11,1061,1062],{},"\"做 AI 客服\"是结果，\"客户咨询响应慢导致流失\"是问题。从问题出发才能选对路径。",[117,1064,1066],{"id":1065},"_2-先内部再外部先辅助再替代","2. 先内部，再外部；先辅助，再替代",[11,1068,1069],{},"内部工具是最好的 AI 试验田，员工反馈快、容错高。\"AI 替代人\"的项目失败率远高于\"AI 辅助人\"。",[117,1071,1073],{"id":1072},"_3-合规与数据安全是前置条件不是可选项","3. 合规与数据安全是前置条件，不是可选项",[11,1075,1076,1077,1080],{},"任何涉及客户/员工/经营数据的 AI 项目，",[15,1078,1079],{},"先想清楚数据怎么走、合规怎么过","。私有化部署是大多数传统企业的唯一答案。",[20,1082,1084],{"id":1083},"仙宫云的传统企业-ai-落地方法论","仙宫云的传统企业 AI 落地方法论",[11,1086,1087],{},"我们服务过的 50+ 传统企业，提炼出一套\"四阶段陪跑\"方法：",[273,1089,1090,1096,1102,1108],{},[125,1091,1092,1095],{},[15,1093,1094],{},"诊断期（2-4 周）","：业务调研 + AI 高价值场景识别",[125,1097,1098,1101],{},[15,1099,1100],{},"POC 期（4-8 周）","：选定 1-2 个场景小规模验证",[125,1103,1104,1107],{},[15,1105,1106],{},"推广期（2-3 个月）","：场景扩展 + 员工培训 + 流程嵌入",[125,1109,1110,1113],{},[15,1111,1112],{},"运营期（持续）","：效果监控 + 迭代优化 + 新场景挖掘",[11,1115,1116,1119],{},[15,1117,1118],{},"真正难的不是部署模型，而是把 AI 嵌入业务流程并让员工用起来。"," 这是仙宫云区别于纯技术乙方的核心价值。",[11,1121,1122,1123,1125],{},"如果你的企业正在评估 AI 落地路径，欢迎",[468,1124,471],{"href":470},"获取免费的场景诊断与可行性评估。",[474,1127],{},[11,1129,1130,271,1132,485,1134],{},[15,1131,480],{},[468,1133,484],{"href":483},[468,1135,1136],{"href":520},"企业知识库 RAG 实战教程",{"title":312,"searchDepth":491,"depth":491,"links":1138},[1139,1145,1150,1155,1160],{"id":799,"depth":491,"text":800,"children":1140},[1141,1142,1143,1144],{"id":803,"depth":497,"text":803},{"id":813,"depth":497,"text":814},{"id":831,"depth":497,"text":831},{"id":851,"depth":497,"text":852},{"id":900,"depth":491,"text":901,"children":1146},[1147,1148,1149],{"id":904,"depth":497,"text":803},{"id":910,"depth":497,"text":910},{"id":933,"depth":497,"text":933},{"id":976,"depth":491,"text":977,"children":1151},[1152,1153,1154],{"id":980,"depth":497,"text":803},{"id":986,"depth":497,"text":986},{"id":1006,"depth":497,"text":1006},{"id":1054,"depth":491,"text":1055,"children":1156},[1157,1158,1159],{"id":1058,"depth":497,"text":1059},{"id":1065,"depth":497,"text":1066},{"id":1072,"depth":497,"text":1073},{"id":1083,"depth":491,"text":1084},"2026-04-28","从制造、零售、医疗三个真实行业案例，分析传统企业 AI 落地的核心挑战与可行路径，帮助决策者避坑。",{},9,{"title":787,"description":1162},"blog/traditional-enterprise-ai-challenges",[1168,1169,1170,1171,1172],"企业AI落地","AI转型","制造业AI","零售AI","医疗AI","LkS3KZCh7u39y97nZw71AVsAs5WvLsLxPbtn4xjO8wM",{"id":1175,"title":1176,"author":6,"body":1177,"category":1655,"cover":514,"date":1656,"description":1657,"extension":517,"meta":1658,"navigation":519,"path":483,"readingTime":1659,"seo":1660,"stem":1661,"tags":1662,"__hash__":1667},"blog/blog/deepseek-private-deployment-guide.md","DeepSeek 大模型私有化部署完整指南：硬件、成本与避坑要点",{"type":8,"value":1178,"toc":1638},[1179,1186,1190,1193,1219,1223,1226,1314,1320,1324,1328,1331,1357,1361,1364,1386,1390,1393,1415,1419,1422,1426,1437,1441,1452,1456,1467,1471,1474,1556,1559,1563,1595,1599,1602,1622,1628,1630],[11,1180,1181,1182,1185],{},"DeepSeek 在 2024-2025 年成为国内企业大模型私有化部署的首选之一。它开源、中文能力强、推理性能稳定，但真正落地时，企业最常问的三个问题是：",[15,1183,1184],{},"要什么硬件？花多少钱？怎么避坑？"," 本文给出 2026 年最新的实操答案。",[20,1187,1189],{"id":1188},"一为什么企业要做-deepseek-私有化部署","一、为什么企业要做 DeepSeek 私有化部署？",[11,1191,1192],{},"调用 API 当然便宜，但当业务涉及以下任一情况，私有化部署几乎是唯一选择：",[122,1194,1195,1201,1207,1213],{},[125,1196,1197,1200],{},[15,1198,1199],{},"数据敏感","：客户合同、医疗记录、财务凭证、研发资料这类数据不能出企业内网",[125,1202,1203,1206],{},[15,1204,1205],{},"合规要求","：等保三级、金融监管、医疗行业合规，明确要求数据本地化",[125,1208,1209,1212],{},[15,1210,1211],{},"成本临界点","：当 API 月调用量超过 5000 万 tokens，自建反而更便宜",[125,1214,1215,1218],{},[15,1216,1217],{},"稳定性要求","：业务系统强依赖 AI，不能因为外部 API 限流或宕机而中断",[20,1220,1222],{"id":1221},"二模型版本怎么选","二、模型版本怎么选？",[11,1224,1225],{},"DeepSeek 官方目前主要开源以下几个版本，企业可根据预算和场景选择：",[37,1227,1228,1243],{},[40,1229,1230],{},[43,1231,1232,1234,1237,1240],{},[46,1233,1026],{},[46,1235,1236],{},"参数规模",[46,1238,1239],{},"推荐场景",[46,1241,1242],{},"最低显存（FP16）",[56,1244,1245,1259,1273,1287,1301],{},[43,1246,1247,1250,1253,1256],{},[61,1248,1249],{},"DeepSeek-R1-Distill-Qwen-7B",[61,1251,1252],{},"7B",[61,1254,1255],{},"客服、简单文档问答",[61,1257,1258],{},"16 GB",[43,1260,1261,1264,1267,1270],{},[61,1262,1263],{},"DeepSeek-R1-Distill-Qwen-14B",[61,1265,1266],{},"14B",[61,1268,1269],{},"知识库 RAG、报告生成",[61,1271,1272],{},"32 GB",[43,1274,1275,1278,1281,1284],{},[61,1276,1277],{},"DeepSeek-R1-Distill-Qwen-32B",[61,1279,1280],{},"32B",[61,1282,1283],{},"复杂推理、合同审阅",[61,1285,1286],{},"64 GB",[43,1288,1289,1292,1295,1298],{},[61,1290,1291],{},"DeepSeek-V3",[61,1293,1294],{},"671B (MoE)",[61,1296,1297],{},"高级 Agent、企业核心场景",[61,1299,1300],{},"8×A100 80G 起",[43,1302,1303,1306,1308,1311],{},[61,1304,1305],{},"DeepSeek-R1",[61,1307,1294],{},[61,1309,1310],{},"复杂推理、深度思考任务",[61,1312,1313],{},"8×H100 80G 起",[11,1315,1316,1319],{},[15,1317,1318],{},"经验法则","：90% 的企业内部场景（客服、知识库、文档处理）用 14B-32B 蒸馏版就够了，不要一上来就追 671B 满血版，硬件成本会翻 10 倍以上。",[20,1321,1323],{"id":1322},"三硬件配置参考2026-年价格","三、硬件配置参考（2026 年价格）",[117,1325,1327],{"id":1326},"入门级7b-14b-模型","入门级（7B-14B 模型）",[11,1329,1330],{},"适合 30-50 人小团队、单一业务场景。",[122,1332,1333,1339,1345,1351],{},[125,1334,1335,1338],{},[15,1336,1337],{},"GPU","：1× RTX 4090（24GB）或 1× RTX A6000（48GB）",[125,1340,1341,1344],{},[15,1342,1343],{},"CPU/内存","：32 核 / 128 GB",[125,1346,1347,1350],{},[15,1348,1349],{},"存储","：2TB NVMe SSD",[125,1352,1353,1356],{},[15,1354,1355],{},"整机预算","：6-15 万元",[117,1358,1360],{"id":1359},"中型32b-模型","中型（32B 模型）",[11,1362,1363],{},"适合 100-500 人企业、多场景并发。",[122,1365,1366,1371,1376,1381],{},[125,1367,1368,1370],{},[15,1369,1337],{},"：2× A100 80G 或 4× RTX 4090",[125,1372,1373,1375],{},[15,1374,1343],{},"：64 核 / 256 GB",[125,1377,1378,1380],{},[15,1379,1349],{},"：4TB NVMe SSD",[125,1382,1383,1385],{},[15,1384,1355],{},"：35-60 万元",[117,1387,1389],{"id":1388},"旗舰级deepseek-v3r1-满血版","旗舰级（DeepSeek-V3/R1 满血版）",[11,1391,1392],{},"适合大型集团、高并发核心业务。",[122,1394,1395,1400,1405,1410],{},[125,1396,1397,1399],{},[15,1398,1337],{},"：8× H100 80G 或 8× A100 80G（NVLink 互联）",[125,1401,1402,1404],{},[15,1403,1343],{},"：128 核 / 1TB",[125,1406,1407,1409],{},[15,1408,1349],{},"：10TB+ NVMe SSD",[125,1411,1412,1414],{},[15,1413,1355],{},"：200-400 万元",[20,1416,1418],{"id":1417},"四推理框架怎么选","四、推理框架怎么选？",[11,1420,1421],{},"部署框架直接影响吞吐量和响应延迟。三个主流选择：",[117,1423,1425],{"id":1424},"_1-vllm生产首选","1. vLLM（生产首选）",[122,1427,1428,1431,1434],{},[125,1429,1430],{},"优点：吞吐量高、支持 PagedAttention、连续批处理",[125,1432,1433],{},"缺点：配置稍复杂",[125,1435,1436],{},"适用：生产环境、高并发场景",[117,1438,1440],{"id":1439},"_2-ollama最简单","2. Ollama（最简单）",[122,1442,1443,1446,1449],{},[125,1444,1445],{},"优点：一行命令启动、支持量化模型",[125,1447,1448],{},"缺点：单机性能有限，不适合高并发",[125,1450,1451],{},"适用：POC 验证、小团队内部使用",[117,1453,1455],{"id":1454},"_3-sglang前沿","3. SGLang（前沿）",[122,1457,1458,1461,1464],{},[125,1459,1460],{},"优点：结构化生成快，工具调用场景表现好",[125,1462,1463],{},"缺点：生态相对新",[125,1465,1466],{},"适用：Agent 应用、复杂推理",[20,1468,1470],{"id":1469},"五典型企业部署成本拆解","五、典型企业部署成本拆解",[11,1472,1473],{},"以一个 200 人制造企业部署 DeepSeek-R1-Distill-Qwen-32B 为例：",[37,1475,1476,1489],{},[40,1477,1478],{},[43,1479,1480,1483,1486],{},[46,1481,1482],{},"项目",[46,1484,1485],{},"一次性",[46,1487,1488],{},"年化",[56,1490,1491,1502,1512,1522,1532,1542],{},[43,1492,1493,1496,1499],{},[61,1494,1495],{},"硬件采购（2× A100）",[61,1497,1498],{},"45 万",[61,1500,1501],{},"-",[43,1503,1504,1507,1510],{},[61,1505,1506],{},"机房环境改造",[61,1508,1509],{},"5 万",[61,1511,1501],{},[43,1513,1514,1517,1520],{},[61,1515,1516],{},"部署实施服务",[61,1518,1519],{},"8-15 万",[61,1521,1501],{},[43,1523,1524,1527,1529],{},[61,1525,1526],{},"电费（24/7 运行）",[61,1528,1501],{},[61,1530,1531],{},"3-5 万",[43,1533,1534,1537,1539],{},[61,1535,1536],{},"运维与模型更新",[61,1538,1501],{},[61,1540,1541],{},"6-12 万",[43,1543,1544,1549,1554],{},[61,1545,1546],{},[15,1547,1548],{},"三年总成本",[61,1550,1551],{},[15,1552,1553],{},"约 75-90 万",[61,1555,1501],{},[11,1557,1558],{},"对照 API 方案：同样规模业务调用，按 0.001 元/千 tokens 估算，三年通常在 30-150 万之间——但数据出域、不可控、长期议价权弱。",[20,1560,1562],{"id":1561},"六企业落地最容易踩的-5-个坑","六、企业落地最容易踩的 5 个坑",[273,1564,1565,1571,1577,1583,1589],{},[125,1566,1567,1570],{},[15,1568,1569],{},"追求满血版","：90% 场景蒸馏版足够，盲目上 671B 浪费硬件",[125,1572,1573,1576],{},[15,1574,1575],{},"忽视吞吐量测试","：部署完才发现并发 10 人就卡，前期没做压测",[125,1578,1579,1582],{},[15,1580,1581],{},"没做模型评估","：直接选最火的，没用自家业务数据测准确率",[125,1584,1585,1588],{},[15,1586,1587],{},"忽略 RAG 配套","：模型部署完没接知识库，用户体验和直接调 API 没区别",[125,1590,1591,1594],{},[15,1592,1593],{},"缺乏运维计划","：模型发版迭代、显卡故障处理、效果回归没人管",[20,1596,1598],{"id":1597},"七仙宫云的部署服务","七、仙宫云的部署服务",[11,1600,1601],{},"仙宫云已为多家制造、零售、医疗、金融企业完成 DeepSeek 私有化部署，提供：",[122,1603,1604,1610,1616],{},[125,1605,1606,1609],{},[15,1607,1608],{},"部署前","：业务场景评估、模型选型、硬件方案、ROI 测算",[125,1611,1612,1615],{},[15,1613,1614],{},"部署中","：硬件部署、模型推理优化、RAG 知识库集成、应用对接",[125,1617,1618,1621],{},[15,1619,1620],{},"部署后","：员工培训、效果监控、模型版本升级、长期陪跑",[11,1623,1624,1625,1627],{},"如果你正在评估 DeepSeek 私有化部署，欢迎",[468,1626,471],{"href":470},"获取免费方案评估。",[474,1629],{},[11,1631,1632,271,1634,485,1636],{},[15,1633,480],{},[468,1635,1136],{"href":520},[468,1637,489],{"href":488},{"title":312,"searchDepth":491,"depth":491,"links":1639},[1640,1641,1642,1647,1652,1653,1654],{"id":1188,"depth":491,"text":1189},{"id":1221,"depth":491,"text":1222},{"id":1322,"depth":491,"text":1323,"children":1643},[1644,1645,1646],{"id":1326,"depth":497,"text":1327},{"id":1359,"depth":497,"text":1360},{"id":1388,"depth":497,"text":1389},{"id":1417,"depth":491,"text":1418,"children":1648},[1649,1650,1651],{"id":1424,"depth":497,"text":1425},{"id":1439,"depth":497,"text":1440},{"id":1454,"depth":497,"text":1455},{"id":1469,"depth":491,"text":1470},{"id":1561,"depth":491,"text":1562},{"id":1597,"depth":491,"text":1598},"私有化部署","2026-04-12","一篇文章看懂 DeepSeek-R1/V3 私有化部署所需的硬件、显存、推理框架选择、典型成本区间与企业落地常见坑，2026 年最新版。",{},12,{"title":1176,"description":1657},"blog/deepseek-private-deployment-guide",[782,1663,1664,1665,1666],"大模型私有化部署","本地化部署","vLLM","Ollama","CnUfK8LxNz_IpO364Os_X0FVfj8Hvm3vAdpqq6ePD7A",1778126192184]