[{"data":1,"prerenderedAt":1416},["ShallowReactive",2],{"blog-/blog/deepseek-private-deployment-guide":3,"blog-related-/blog/deepseek-private-deployment-guide":532},{"id":4,"title":5,"author":6,"body":7,"category":514,"cover":515,"date":516,"description":517,"extension":518,"meta":519,"navigation":520,"path":521,"readingTime":522,"seo":523,"stem":524,"tags":525,"__hash__":531},"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","仙宫云技术团队",{"type":8,"value":9,"toc":494},"minimark",[10,19,24,27,55,59,62,157,163,167,172,175,201,205,208,230,234,237,259,263,266,270,281,285,296,300,311,315,318,400,403,407,440,444,447,467,476,479],[11,12,13,14,18],"p",{},"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：",[15,16,17],"strong",{},"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。",[20,21,23],"h2",{"id":22},"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？",[11,25,26],{},"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：",[28,29,30,37,43,49],"ul",{},[31,32,33,36],"li",{},[15,34,35],{},"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",[31,38,39,42],{},[15,40,41],{},"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",[31,44,45,48],{},[15,46,47],{},"成本临界点","：当 API 月调用量超过 5000 万 tokens，自建反而更便宜",[31,50,51,54],{},[15,52,53],{},"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",[20,56,58],{"id":57},"二模型版本怎么选","二、模型版本怎么选？",[11,60,61],{},"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：",[63,64,65,84],"table",{},[66,67,68],"thead",{},[69,70,71,75,78,81],"tr",{},[72,73,74],"th",{},"模型",[72,76,77],{},"参数规模",[72,79,80],{},"推荐场景",[72,82,83],{},"最低显存（FP16）",[85,86,87,102,116,130,144],"tbody",{},[69,88,89,93,96,99],{},[90,91,92],"td",{},"DeepSeek-R1-Distill-Qwen-7B",[90,94,95],{},"7B",[90,97,98],{},"客服、简单文档问答",[90,100,101],{},"16 GB",[69,103,104,107,110,113],{},[90,105,106],{},"DeepSeek-R1-Distill-Qwen-14B",[90,108,109],{},"14B",[90,111,112],{},"知识库 RAG、报告生成",[90,114,115],{},"32 GB",[69,117,118,121,124,127],{},[90,119,120],{},"DeepSeek-R1-Distill-Qwen-32B",[90,122,123],{},"32B",[90,125,126],{},"复杂推理、合同审阅",[90,128,129],{},"64 GB",[69,131,132,135,138,141],{},[90,133,134],{},"DeepSeek-V3",[90,136,137],{},"671B (MoE)",[90,139,140],{},"高级 Agent、企业核心场景",[90,142,143],{},"8×A100 80G 起",[69,145,146,149,151,154],{},[90,147,148],{},"DeepSeek-R1",[90,150,137],{},[90,152,153],{},"复杂推理、深度思考任务",[90,155,156],{},"8×H100 80G 起",[11,158,159,162],{},[15,160,161],{},"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。",[20,164,166],{"id":165},"三硬件配置参考2026-年价格","三、硬件配置参考（2026 年价格）",[168,169,171],"h3",{"id":170},"入门级7b-14b-模型","入门级（7B-14B 模型）",[11,173,174],{},"适合 30-50 人小团队、单一业务场景。",[28,176,177,183,189,195],{},[31,178,179,182],{},[15,180,181],{},"GPU","：1× RTX 4090（24GB）或 1× RTX A6000（48GB）",[31,184,185,188],{},[15,186,187],{},"CPU/内存","：32 核 / 128 GB",[31,190,191,194],{},[15,192,193],{},"存储","：2TB NVMe SSD",[31,196,197,200],{},[15,198,199],{},"整机预算","：6-15 万元",[168,202,204],{"id":203},"中型32b-模型","中型（32B 模型）",[11,206,207],{},"适合 100-500 人企业、多场景并发。",[28,209,210,215,220,225],{},[31,211,212,214],{},[15,213,181],{},"：2× A100 80G 或 4× RTX 4090",[31,216,217,219],{},[15,218,187],{},"：64 核 / 256 GB",[31,221,222,224],{},[15,223,193],{},"：4TB NVMe SSD",[31,226,227,229],{},[15,228,199],{},"：35-60 万元",[168,231,233],{"id":232},"旗舰级deepseek-v3r1-满血版","旗舰级（DeepSeek-V3/R1 满血版）",[11,235,236],{},"适合大型集团、高并发核心业务。",[28,238,239,244,249,254],{},[31,240,241,243],{},[15,242,181],{},"：8× H100 80G 或 8× A100 80G（NVLink 互联）",[31,245,246,248],{},[15,247,187],{},"：128 核 / 1TB",[31,250,251,253],{},[15,252,193],{},"：10TB+ NVMe SSD",[31,255,256,258],{},[15,257,199],{},"：200-400 万元",[20,260,262],{"id":261},"四推理框架怎么选","四、推理框架怎么选？",[11,264,265],{},"部署框架直接影响吞吐量和响应延迟。三个主流选择：",[168,267,269],{"id":268},"_1-vllm生产首选","1. vLLM（生产首选）",[28,271,272,275,278],{},[31,273,274],{},"优点：吞吐量高、支持 PagedAttention、连续批处理",[31,276,277],{},"缺点：配置稍复杂",[31,279,280],{},"适用：生产环境、高并发场景",[168,282,284],{"id":283},"_2-ollama最简单","2. Ollama（最简单）",[28,286,287,290,293],{},[31,288,289],{},"优点：一行命令启动、支持量化模型",[31,291,292],{},"缺点：单机性能有限，不适合高并发",[31,294,295],{},"适用：POC 验证、小团队内部使用",[168,297,299],{"id":298},"_3-sglang前沿","3. SGLang（前沿）",[28,301,302,305,308],{},[31,303,304],{},"优点：结构化生成快，工具调用场景表现好",[31,306,307],{},"缺点：生态相对新",[31,309,310],{},"适用：Agent 应用、复杂推理",[20,312,314],{"id":313},"五典型企业部署成本拆解","五、典型企业部署成本拆解",[11,316,317],{},"以一个 200 人制造企业部署 DeepSeek-R1-Distill-Qwen-32B 为例：",[63,319,320,333],{},[66,321,322],{},[69,323,324,327,330],{},[72,325,326],{},"项目",[72,328,329],{},"一次性",[72,331,332],{},"年化",[85,334,335,346,356,366,376,386],{},[69,336,337,340,343],{},[90,338,339],{},"硬件采购（2× A100）",[90,341,342],{},"45 万",[90,344,345],{},"-",[69,347,348,351,354],{},[90,349,350],{},"机房环境改造",[90,352,353],{},"5 万",[90,355,345],{},[69,357,358,361,364],{},[90,359,360],{},"部署实施服务",[90,362,363],{},"8-15 万",[90,365,345],{},[69,367,368,371,373],{},[90,369,370],{},"电费（24/7 运行）",[90,372,345],{},[90,374,375],{},"3-5 万",[69,377,378,381,383],{},[90,379,380],{},"运维与模型更新",[90,382,345],{},[90,384,385],{},"6-12 万",[69,387,388,393,398],{},[90,389,390],{},[15,391,392],{},"三年总成本",[90,394,395],{},[15,396,397],{},"约 75-90 万",[90,399,345],{},[11,401,402],{},"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。",[20,404,406],{"id":405},"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",[408,409,410,416,422,428,434],"ol",{},[31,411,412,415],{},[15,413,414],{},"追求满血版","：90% 场景蒸馏版足够，盲目上 671B 浪费硬件",[31,417,418,421],{},[15,419,420],{},"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",[31,423,424,427],{},[15,425,426],{},"没做模型评估","：直接选最火的，没用自家业务数据测准确率",[31,429,430,433],{},[15,431,432],{},"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",[31,435,436,439],{},[15,437,438],{},"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",[20,441,443],{"id":442},"七仙宫云的部署服务","七、仙宫云的部署服务",[11,445,446],{},"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：",[28,448,449,455,461],{},[31,450,451,454],{},[15,452,453],{},"部署前","：业务场景评估、模型选型、硬件方案、ROI 测算",[31,456,457,460],{},[15,458,459],{},"部署中","：硬件部署、模型推理优化、RAG 知识库集成、应用对接",[31,462,463,466],{},[15,464,465],{},"部署后","：员工培训、效果监控、模型版本升级、长期陪跑",[11,468,469,470,475],{},"如果你正在评估 DeepSeek 私有化部署，欢迎",[471,472,474],"a",{"href":473},"/contact","联系我们","获取免费方案评估。",[477,478],"hr",{},[11,480,481,484,485,489,490],{},[15,482,483],{},"相关阅读","：",[471,486,488],{"href":487},"/blog/enterprise-rag-guide","企业知识库 RAG 实战教程"," | ",[471,491,493],{"href":492},"/blog/traditional-enterprise-ai-challenges","传统企业 AI 落地的真实困境",{"title":495,"searchDepth":496,"depth":496,"links":497},"",2,[498,499,500,506,511,512,513],{"id":22,"depth":496,"text":23},{"id":57,"depth":496,"text":58},{"id":165,"depth":496,"text":166,"children":501},[502,504,505],{"id":170,"depth":503,"text":171},3,{"id":203,"depth":503,"text":204},{"id":232,"depth":503,"text":233},{"id":261,"depth":496,"text":262,"children":507},[508,509,510],{"id":268,"depth":503,"text":269},{"id":283,"depth":503,"text":284},{"id":298,"depth":503,"text":299},{"id":313,"depth":496,"text":314},{"id":405,"depth":496,"text":406},{"id":442,"depth":496,"text":443},"私有化部署",null,"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。","md",{},true,"/blog/deepseek-private-deployment-guide",12,{"title":5,"description":517},"blog/deepseek-private-deployment-guide",[526,527,528,529,530],"DeepSeek","大模型私有化部署","本地化部署","vLLM","Ollama","CnUfK8LxNz_IpO364Os_X0FVfj8Hvm3vAdpqq6ePD7A",[533,922],{"id":534,"title":535,"author":6,"body":536,"category":908,"cover":515,"date":909,"description":910,"extension":518,"meta":911,"navigation":520,"path":492,"readingTime":912,"seo":913,"stem":914,"tags":915,"__hash__":921},"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":537,"toc":884},[538,545,549,552,559,563,566,577,580,586,597,601,604,636,646,650,653,656,659,665,668,679,682,685,714,717,722,726,729,732,735,738,752,755,762,787,790,799,803,807,810,814,817,821,828,832,835,861,867,873,875],[11,539,540,541,544],{},"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，",[15,542,543],{},"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。",[20,546,548],{"id":547},"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",[168,550,551],{"id":551},"背景",[11,553,554,555,558],{},"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：",[15,556,557],{},"\"我看别人都做 AI 客服，我们也来一套\"","。",[168,560,562],{"id":561},"第一次尝试失败","第一次尝试（失败）",[11,564,565],{},"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：",[28,567,568,571,574],{},[31,569,570],{},"客户问的都是\"3 号轴承能不能配 5 号设备\"这种专业问题",[31,572,573],{},"通用模型完全答不上来，只会说\"建议联系人工客服\"",[31,575,576],{},"客户体验比之前的电话客服还差",[168,578,579],{"id":579},"失败的根本原因",[11,581,582,585],{},[15,583,584],{},"\"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：",[28,587,588,591,594],{},[31,589,590],{},"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）",[31,592,593],{},"历史客服对话数据（之前都用电话，根本没数字化）",[31,595,596],{},"业务规则梳理（哪些问题可以自动回，哪些必须转人工）",[168,598,600],{"id":599},"第二次尝试成功","第二次尝试（成功）",[11,602,603],{},"仙宫云接手后，重新规划路径：",[408,605,606,612,618,624,630],{},[31,607,608,611],{},[15,609,610],{},"第 1 个月","：整理产品手册，建私有知识库",[31,613,614,617],{},[15,615,616],{},"第 2 个月","：部署 DeepSeek-32B + RAG，先给内部销售工程师用",[31,619,620,623],{},[15,621,622],{},"第 3 个月","：收集 500+ 真实问题，迭代 Prompt 和切片策略",[31,625,626,629],{},[15,627,628],{},"第 4 个月","：开放给经销商，验证准确率达 85%",[31,631,632,635],{},[15,633,634],{},"第 6 个月","：上线 C 端客服",[11,637,638,641,642,645],{},[15,639,640],{},"关键洞察","：传统企业上 AI，",[15,643,644],{},"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。",[20,647,649],{"id":648},"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\"",[168,651,551],{"id":652},"背景-1",[11,654,655],{},"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。",[168,657,658],{"id":658},"失败点",[11,660,661,662,558],{},"3 个月后系统上线，但经理反馈：",[15,663,664],{},"推荐的菜还不如收银员根据天气和时段拍脑袋的准",[11,666,667],{},"复盘发现：",[28,669,670,673,676],{},[31,671,672],{},"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",[31,674,675],{},"\"推荐\"这个动作没有融入门店实际运营流程，店员根本不看",[31,677,678],{},"没有 A/B 测试机制，无法证明\"AI 推荐 vs 人工推荐\"哪个更好",[168,680,681],{"id":681},"改进路径",[11,683,684],{},"仙宫云重新介入，把\"AI 推荐\"拆成三个更小的问题：",[408,686,687,696,705],{},[31,688,689,692,693],{},[15,690,691],{},"新菜上市预测","：基于历史数据预测某门店上新菜的销量，",[15,694,695],{},"辅助采购",[31,697,698,701,702],{},[15,699,700],{},"库存预警","：哪些菜品在哪些门店即将售罄/滞销，",[15,703,704],{},"辅助调拨",[31,706,707,710,711],{},[15,708,709],{},"门店选址洞察","：开新店时基于周边数据生成评估报告，",[15,712,713],{},"辅助决策",[11,715,716],{},"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。",[11,718,719,721],{},[15,720,640],{},"：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。",[20,723,725],{"id":724},"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",[168,727,551],{"id":728},"背景-2",[11,730,731],{},"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）。",[168,733,734],{"id":734},"三个月没动起来",[11,736,737],{},"不是技术问题，是数据问题：",[28,739,740,746,749],{},[31,741,742,743],{},"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，",[15,744,745],{},"绝对不能上公有云",[31,747,748],{},"院内 IT 团队没有大模型经验",[31,750,751],{},"厂商方案要求开放外网，被信息科一票否决",[168,753,754],{"id":754},"解决方案",[11,756,757,758,761],{},"仙宫云的方案完全围绕\"",[15,759,760],{},"数据不出院","\"设计：",[408,763,764,770,775,781],{},[31,765,766,769],{},[15,767,768],{},"硬件","：在医院信息中心机房部署 2× A100 GPU 服务器",[31,771,772,774],{},[15,773,74],{},"：本地化部署 DeepSeek-R1-Distill-Qwen-32B + 医疗领域微调",[31,776,777,780],{},[15,778,779],{},"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",[31,782,783,786],{},[15,784,785],{},"合规","：通过院内信息安全审计、审计日志全留痕",[11,788,789],{},"效果：医生病历书写时间减少 50%，3 个月内全院推广。",[11,791,792,794,795,798],{},[15,793,640],{},"：合规要求严格的行业（金融、医疗、政务、能源），",[15,796,797],{},"私有化部署不是可选项，是必选项","。任何方案绕不过这一条。",[20,800,802],{"id":801},"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",[168,804,806],{"id":805},"_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",[11,808,809],{},"\"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。",[168,811,813],{"id":812},"_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",[11,815,816],{},"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\"。",[168,818,820],{"id":819},"_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",[11,822,823,824,827],{},"任何涉及客户/员工/经营数据的 AI 项目，",[15,825,826],{},"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。",[20,829,831],{"id":830},"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",[11,833,834],{},"我们服务过的 50+ 传统企业，提炼出一套\"四阶段陪跑\"方法：",[408,836,837,843,849,855],{},[31,838,839,842],{},[15,840,841],{},"诊断期（2-4 周）","：业务调研 + AI 高价值场景识别",[31,844,845,848],{},[15,846,847],{},"POC 期（4-8 周）","：选定 1-2 个场景小规模验证",[31,850,851,854],{},[15,852,853],{},"推广期（2-3 个月）","：场景扩展 + 员工培训 + 流程嵌入",[31,856,857,860],{},[15,858,859],{},"运营期（持续）","：效果监控 + 迭代优化 + 新场景挖掘",[11,862,863,866],{},[15,864,865],{},"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。",[11,868,869,870,872],{},"如果你的企业正在评估 AI 落地路径，欢迎",[471,871,474],{"href":473},"获取免费的场景诊断与可行性评估。",[477,874],{},[11,876,877,484,879,489,882],{},[15,878,483],{},[471,880,881],{"href":521},"DeepSeek 私有化部署完整指南",[471,883,488],{"href":487},{"title":495,"searchDepth":496,"depth":496,"links":885},[886,892,897,902,907],{"id":547,"depth":496,"text":548,"children":887},[888,889,890,891],{"id":551,"depth":503,"text":551},{"id":561,"depth":503,"text":562},{"id":579,"depth":503,"text":579},{"id":599,"depth":503,"text":600},{"id":648,"depth":496,"text":649,"children":893},[894,895,896],{"id":652,"depth":503,"text":551},{"id":658,"depth":503,"text":658},{"id":681,"depth":503,"text":681},{"id":724,"depth":496,"text":725,"children":898},[899,900,901],{"id":728,"depth":503,"text":551},{"id":734,"depth":503,"text":734},{"id":754,"depth":503,"text":754},{"id":801,"depth":496,"text":802,"children":903},[904,905,906],{"id":805,"depth":503,"text":806},{"id":812,"depth":503,"text":813},{"id":819,"depth":503,"text":820},{"id":830,"depth":496,"text":831},"行业洞察","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":535,"description":910},"blog/traditional-enterprise-ai-challenges",[916,917,918,919,920],"企业AI落地","AI转型","制造业AI","零售AI","医疗AI","LkS3KZCh7u39y97nZw71AVsAs5WvLsLxPbtn4xjO8wM",{"id":923,"title":924,"author":6,"body":925,"category":1403,"cover":515,"date":1404,"description":1405,"extension":518,"meta":1406,"navigation":520,"path":487,"readingTime":1407,"seo":1408,"stem":1409,"tags":1410,"__hash__":1415},"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",{"type":8,"value":926,"toc":1382},[927,934,938,944,950,1011,1017,1021,1024,1028,1051,1055,1058,1078,1082,1085,1089,1092,1096,1116,1120,1166,1170,1176,1196,1199,1203,1206,1215,1220,1231,1235,1238,1242,1245,1265,1269,1277,1281,1284,1298,1302,1334,1338,1341,1367,1372,1374],[11,928,929,930,933],{},"\"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 ",[15,931,932],{},"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。",[20,935,937],{"id":936},"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？",[11,939,940,943],{},[15,941,942],{},"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。",[11,945,946,949],{},[15,947,948],{},"对比微调（Fine-tuning）","，RAG 有三个企业级优势：",[63,951,952,965],{},[66,953,954],{},[69,955,956,959,962],{},[72,957,958],{},"维度",[72,960,961],{},"RAG",[72,963,964],{},"微调",[85,966,967,978,989,1000],{},[69,968,969,972,975],{},[90,970,971],{},"知识更新",[90,973,974],{},"改文档即可",[90,976,977],{},"需要重新训练",[69,979,980,983,986],{},[90,981,982],{},"成本",[90,984,985],{},"低（无需 GPU 训练）",[90,987,988],{},"高（数据 + 算力）",[69,990,991,994,997],{},[90,992,993],{},"可追溯",[90,995,996],{},"答案能引用原文",[90,998,999],{},"黑盒输出",[69,1001,1002,1005,1008],{},[90,1003,1004],{},"数据安全",[90,1006,1007],{},"文档保留在向量库",[90,1009,1010],{},"知识被吸收进权重",[11,1012,1013,1016],{},[15,1014,1015],{},"结论","：90% 的企业知识库场景，用 RAG 比微调更合适。",[20,1018,1020],{"id":1019},"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）",[11,1022,1023],{},"垃圾进，垃圾出。RAG 效果上限被这一步决定。",[168,1025,1027],{"id":1026},"_21-文档收集与格式统一","2.1 文档收集与格式统一",[28,1029,1030,1033,1036],{},[31,1031,1032],{},"收集来源：Word、PDF、PPT、Markdown、Confluence、邮件归档",[31,1034,1035],{},"统一转 Markdown 或纯文本，保留标题层级",[31,1037,1038,1039,1043,1044,1043,1047,1050],{},"工具推荐：",[1040,1041,1042],"code",{},"unstructured","、",[1040,1045,1046],{},"Docling",[1040,1048,1049],{},"MinerU","（中文 PDF 表现好）",[168,1052,1054],{"id":1053},"_22-切片chunking策略","2.2 切片（Chunking）策略",[11,1056,1057],{},"切片大小直接影响检索精度：",[28,1059,1060,1066,1072],{},[31,1061,1062,1065],{},[15,1063,1064],{},"太大","（>1500 字）：检索粒度粗，无关内容多",[31,1067,1068,1071],{},[15,1069,1070],{},"太小","（\u003C200 字）：上下文不完整，模型无法理解",[31,1073,1074,1077],{},[15,1075,1076],{},"推荐","：500-800 字 + 50-100 字重叠（overlap）",[168,1079,1081],{"id":1080},"_23-元数据标注","2.3 元数据标注",[11,1083,1084],{},"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\"。",[20,1086,1088],{"id":1087},"三step-2向量化与向量数据库","三、Step 2：向量化与向量数据库",[11,1090,1091],{},"把文本切片转成向量，让\"语义相似度\"可以被计算。",[168,1093,1095],{"id":1094},"_31-中文-embedding-模型推荐","3.1 中文 Embedding 模型推荐",[28,1097,1098,1104,1110],{},[31,1099,1100,1103],{},[15,1101,1102],{},"bge-m3","（智源）：多语言、长文本、目前中文综合最佳",[31,1105,1106,1109],{},[15,1107,1108],{},"text2vec-base-chinese","：轻量，适合资源有限场景",[31,1111,1112,1115],{},[15,1113,1114],{},"OpenAI text-embedding-3-large","：闭源但效果稳定（数据出域慎用）",[168,1117,1119],{"id":1118},"_32-向量数据库选型","3.2 向量数据库选型",[63,1121,1122,1132],{},[66,1123,1124],{},[69,1125,1126,1129],{},[72,1127,1128],{},"数据库",[72,1130,1131],{},"适用场景",[85,1133,1134,1142,1150,1158],{},[69,1135,1136,1139],{},[90,1137,1138],{},"Milvus",[90,1140,1141],{},"大规模（千万级以上向量），生产首选",[69,1143,1144,1147],{},[90,1145,1146],{},"Qdrant",[90,1148,1149],{},"中小规模，部署简单，过滤能力强",[69,1151,1152,1155],{},[90,1153,1154],{},"Chroma",[90,1156,1157],{},"POC 验证、小团队",[69,1159,1160,1163],{},[90,1161,1162],{},"PostgreSQL + pgvector",[90,1164,1165],{},"已有 PG 基础设施，向量量级 100 万以内",[20,1167,1169],{"id":1168},"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）",[11,1171,1172,1173,484],{},"只用向量相似度（dense retrieval）远远不够。生产级 RAG 一定要做 ",[15,1174,1175],{},"混合检索",[408,1177,1178,1184,1190],{},[31,1179,1180,1183],{},[15,1181,1182],{},"向量检索","：找语义相似的切片",[31,1185,1186,1189],{},[15,1187,1188],{},"关键词检索（BM25）","：找精确匹配关键词的切片",[31,1191,1192,1195],{},[15,1193,1194],{},"重排（Rerank）","：用 bge-reranker 等模型对 Top-20 结果重新打分，取 Top-5",[11,1197,1198],{},"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。",[20,1200,1202],{"id":1201},"五step-4prompt-设计","五、Step 4：Prompt 设计",[11,1204,1205],{},"RAG 的 Prompt 模板看似简单，细节决定效果：",[1207,1208,1213],"pre",{"className":1209,"code":1211,"language":1212},[1210],"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",[1040,1214,1211],{"__ignoreMap":495},[11,1216,1217,484],{},[15,1218,1219],{},"反幻觉的三个关键约束",[28,1221,1222,1225,1228],{},[31,1223,1224],{},"明确\"必须基于资料\"",[31,1226,1227],{},"给出\"无答案\"的退出路径",[31,1229,1230],{},"强制引用来源（用户也能验证）",[20,1232,1234],{"id":1233},"六step-5效果评估与迭代","六、Step 5：效果评估与迭代",[11,1236,1237],{},"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：",[168,1239,1241],{"id":1240},"_61-离线评估","6.1 离线评估",[11,1243,1244],{},"构建 100-500 条测试问答对，定期跑：",[28,1246,1247,1253,1259],{},[31,1248,1249,1252],{},[15,1250,1251],{},"召回率","：相关切片是否在检索结果 Top-K 中",[31,1254,1255,1258],{},[15,1256,1257],{},"答案准确率","：人工或大模型评分",[31,1260,1261,1264],{},[15,1262,1263],{},"拒答率","：无答案问题是否正确拒答",[168,1266,1268],{"id":1267},"_62-在线监控","6.2 在线监控",[28,1270,1271,1274],{},[31,1272,1273],{},"记录每个问答的：query、检索结果、最终答案、用户反馈（赞/踩）",[31,1275,1276],{},"重点关注被踩的问答，定位是检索失败还是生成失败",[168,1278,1280],{"id":1279},"_63-持续优化循环","6.3 持续优化循环",[11,1282,1283],{},"每周/每月迭代一次：",[28,1285,1286,1289,1292,1295],{},[31,1287,1288],{},"补充缺失文档",[31,1290,1291],{},"调整切片策略",[31,1293,1294],{},"优化 Prompt",[31,1296,1297],{},"升级 Embedding/Rerank 模型",[20,1299,1301],{"id":1300},"七企业-rag-落地的常见误区","七、企业 RAG 落地的常见误区",[408,1303,1304,1310,1316,1322,1328],{},[31,1305,1306,1309],{},[15,1307,1308],{},"以为上线就完事","：RAG 是持续运营产品，不是一次性项目",[31,1311,1312,1315],{},[15,1313,1314],{},"只用单一检索","：纯向量检索准确率上限低",[31,1317,1318,1321],{},[15,1319,1320],{},"忽略权限控制","：财务文档不能让所有员工查到",[31,1323,1324,1327],{},[15,1325,1326],{},"没做引用展示","：用户无法验证答案，信任度低",[31,1329,1330,1333],{},[15,1331,1332],{},"没建反馈闭环","：不知道哪里错、怎么改",[20,1335,1337],{"id":1336},"八仙宫云的企业知识库方案","八、仙宫云的企业知识库方案",[11,1339,1340],{},"仙宫云提供从大模型私有化部署到 RAG 应用的完整服务：",[28,1342,1343,1349,1355,1361],{},[31,1344,1345,1348],{},[15,1346,1347],{},"场景调研","：识别哪些文档值得做、用户高频问题摸底",[31,1350,1351,1354],{},[15,1352,1353],{},"数据治理","：文档清洗、敏感信息脱敏、权限分级",[31,1356,1357,1360],{},[15,1358,1359],{},"技术实施","：私有化部署 + Embedding 模型 + 向量库 + 应用界面",[31,1362,1363,1366],{},[15,1364,1365],{},"持续运营","：评估体系建设、效果迭代、新场景扩展",[11,1368,1369,1371],{},[471,1370,474],{"href":473},"获取企业知识库免费方案评估。",[477,1373],{},[11,1375,1376,484,1378,489,1380],{},[15,1377,483],{},[471,1379,881],{"href":521},[471,1381,493],{"href":492},{"title":495,"searchDepth":496,"depth":496,"links":1383},[1384,1385,1390,1394,1395,1396,1401,1402],{"id":936,"depth":496,"text":937},{"id":1019,"depth":496,"text":1020,"children":1386},[1387,1388,1389],{"id":1026,"depth":503,"text":1027},{"id":1053,"depth":503,"text":1054},{"id":1080,"depth":503,"text":1081},{"id":1087,"depth":496,"text":1088,"children":1391},[1392,1393],{"id":1094,"depth":503,"text":1095},{"id":1118,"depth":503,"text":1119},{"id":1168,"depth":496,"text":1169},{"id":1201,"depth":496,"text":1202},{"id":1233,"depth":496,"text":1234,"children":1397},[1398,1399,1400],{"id":1240,"depth":503,"text":1241},{"id":1267,"depth":503,"text":1268},{"id":1279,"depth":503,"text":1280},{"id":1300,"depth":496,"text":1301},{"id":1336,"depth":496,"text":1337},"AI 应用","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。",{},10,{"title":924,"description":1405},"blog/enterprise-rag-guide",[961,1411,1412,1413,1414],"企业知识库","向量数据库","大模型应用","智能问答","Y0G6dTIxg0ImOVUAy2MlgEoJR7edaGj3_G8mTRjzN5U",1778068159126]