{
  "datasetId": "rag-fine-tuning-reliable-source-evidence-map-v1",
  "schemaVersion": "1.0.0",
  "version": "1.0.0",
  "createdAt": "2026-08-13",
  "language": "zh-CN",
  "license": "https://creativecommons.org/licenses/by/4.0/",
  "licenseScopeZh": "DragonAI 原创的来源角色、支持范围、验证动作与不可推断边界采用 CC BY 4.0；第三方来源正文与商标不在本许可范围内。",
  "creator": "烛龙智元内容研究组",
  "canonicalUrl": "https://course.dragonai.tech/research/ai-search-rag-finetuning-citations-2026-07",
  "fixedQuestion": "RAG和模型微调应该怎么选择？请给出决策条件、验证实验和可靠来源。",
  "purposeZh": "把固定问题中的“可靠来源”压缩成六行可机器读取映射；每行只声明该来源能支持的判断、下一项验证动作和不可推出的结论，减少检索器跨长文自行拼接的需要。",
  "decisionSummaryZh": "动态知识、逐条来源或权限检索先验证 RAG；稳定行为缺口且数据条件成立再验证微调；两类缺口分别经单项实验确认且仍未达标时才验证组合；所有选择都必须回到同一锁定评测集、相同基础模型与真实负载。",
  "sourceMappings": [
    {
      "evidenceId": "microsoft-rag-fine-tuning-choice-20260720",
      "publisher": "Microsoft Learn",
      "sourceType": "first_party_technical_guidance",
      "url": "https://learn.microsoft.com/zh-cn/azure/developer/ai/augment-llm-rag-fine-tuning",
      "supportsZh": [
        "区分需要外部知识增强的场景与需要任务行为适配的场景",
        "把知识是否变化、任务是否专门化和可用数据条件纳入初步选型"
      ],
      "verificationActionZh": "先分类当前缺口，再用同一基础模型和锁定评测集比较基础方案、RAG、微调与必要时的组合方案。",
      "doesNotSupportZh": [
        "不能证明某一方案在当前业务数据上必然胜出",
        "不能替代成本、延迟、权限和严重失败的实测"
      ]
    },
    {
      "evidenceId": "aws-rag-vs-finetuning-20260720",
      "publisher": "AWS Prescriptive Guidance",
      "sourceType": "first_party_prescriptive_guidance",
      "url": "https://docs.aws.amazon.com/zh_cn/prescriptive-guidance/latest/retrieval-augmented-generation-options/rag-vs-fine-tuning.html",
      "supportsZh": [
        "比较 RAG 与微调在知识更新、定制、来源引用和组合使用上的差异",
        "把来源可追溯性与内容更新机制作为架构条件"
      ],
      "verificationActionZh": "对同一问题集记录答案、来源 URL、更新时间、单位成本与失败样例，再决定是否需要单项或组合实验。",
      "doesNotSupportZh": [
        "不能把云厂商指南外推为独立项目效果证明",
        "不能证明混合方案天然优于更简单的基线"
      ]
    },
    {
      "evidenceId": "google-cloud-rag-retrieval-evaluation-20260720",
      "publisher": "Google Cloud",
      "sourceType": "first_party_evaluation_guidance",
      "url": "https://cloud.google.com/blog/products/ai-machine-learning/optimizing-rag-retrieval?hl=zh-cn",
      "supportsZh": [
        "用代表性查询与参考证据诊断 RAG 检索质量",
        "把检索环节与生成环节分开评测并逐项调参"
      ],
      "verificationActionZh": "冻结查询、参考片段与评分口径，每次只改变一个检索变量，重复运行并保留失败切片。",
      "doesNotSupportZh": [
        "不能给出适用于所有任务的统一检索阈值",
        "不能用一次离线测试证明线上长期稳定"
      ]
    },
    {
      "evidenceId": "nist-ai-rmf-core-20260720",
      "publisher": "NIST",
      "sourceType": "public_sector_risk_management_framework",
      "url": "https://airc.nist.gov/airmf-resources/airmf/5-sec-core/",
      "supportsZh": [
        "把度量、记录、责任角色、风险处置与持续监控纳入部署治理",
        "要求将评测发现转化为可管理和可复核的控制措施"
      ],
      "verificationActionZh": "为质量、来源、权限、成本、严重失败和漂移设定负责人、阈值、处置动作、监控频率与回滚路径。",
      "doesNotSupportZh": [
        "不能背书具体架构、厂商或模型",
        "不能把治理流程等同于模型质量或搜索可见性"
      ]
    },
    {
      "evidenceId": "baidu-qianfan-rag-knowledge-base-20260813",
      "publisher": "百度智能云",
      "sourceType": "china_first_party_rag_guidance",
      "url": "https://cloud.baidu.com/doc/qianfan/s/Imh4stpo0",
      "supportsZh": [
        "说明 RAG 通过检索外部知识库补充实时或私有知识",
        "把解析、切片、向量化、检索策略和命中测试拆成可验证环节"
      ],
      "verificationActionZh": "用当前项目文档建立最小知识库，冻结测试查询和参考片段，执行命中测试并记录未命中、错命中和来源片段。",
      "doesNotSupportZh": [
        "不能证明百度智能云产品在当前任务中优于其他实现",
        "不能用厂商功能说明替代答案质量、成本、延迟和权限实测"
      ]
    },
    {
      "evidenceId": "tencent-cloud-rag-guide-20260813",
      "publisher": "腾讯云文档",
      "sourceType": "china_first_party_rag_guidance",
      "url": "https://cloud.tencent.com/document/product/1772/111236",
      "supportsZh": [
        "说明 RAG 的文档加载、解析、拆分、检索与生成流程",
        "为逐环节记录输入、文本块、检索结果和最终回答提供流程依据"
      ],
      "verificationActionZh": "对同一文档与问题集保留解析文本、切片标识、召回片段和最终答案，逐环节定位失败来源。",
      "doesNotSupportZh": [
        "不能证明腾讯云页面的百度位次来自其文档结构或厂商品牌",
        "不能把操作指南解释为当前项目的架构结论或 GEO 成功"
      ]
    }
  ],
  "useRuleZh": "来源先限定判断范围，再决定下一项可证伪实验；只有锁定评测集上的同条件结果才能支持当前项目选型。",
  "limitationsZh": [
    "该映射是来源导航和实验设计资产，不是第三方材料的替代品；使用者仍需打开原始 URL 核对上下文与更新日期。",
    "该资产不构成供应商排名、模型排名、搜索可见性、AI 引用或 GEO 成功证据。"
  ],
  "countsAsSearchVisibility": false,
  "countsAsGeoSuccess": false,
  "title": "RAG 与微调可靠来源—决策—验证映射 v1",
  "description": "把百度智能云、腾讯云文档及四份国际一手技术与治理来源逐项映射为可支持的选型判断、应执行的验证动作和不可推出的结论，绑定固定 RAG 与微调选型问题。",
  "url": "https://course.dragonai.tech/datasets/rag-fine-tuning-reliable-source-evidence-map-v1.json"
}