{
  "datasetId": "rag-fine-tuning-decision-matrix-v1",
  "schemaVersion": "1.0.0",
  "createdAt": "2026-07-20",
  "language": "zh-CN",
  "license": "https://creativecommons.org/licenses/by/4.0/",
  "licenseScope": "The four-option decision structure, eight decision dimensions, Chinese descriptions, verification fields, and blank experiment record created by DragonAI. Third-party source material is excluded and remains under its original terms.",
  "purposeZh": "用于用同一任务、测试集和运行口径比较基础方案、RAG、微调与必要时的混合方案，并保存可复核的选择、停止和回滚证据。",
  "disclosureZh": "DragonAI 运营 AI 产品经理课程。本矩阵是第一方编辑决策工具，不是统一技术标准、厂商性能结论、法律意见、安全审查或项目结果保证。",
  "decisionRuleZh": "先建立基础方案；知识或来源缺口优先验证 RAG，固定任务或行为缺口在样本和权利条件成立时验证微调；只有两类缺口分别被证实且单项方案不足时才验证混合方案。A–D 候选方案在共享开发/验证集上迭代；候选方案、配置和阈值冻结后，才可用锁定的未见最终测试集执行发布门禁。",
  "candidateOptions": [
    {
      "optionId": "baseline",
      "nameZh": "基础模型与提示/规则/工作流",
      "roleZh": "保留最低复杂度基线，证明新增架构的真实增益。"
    },
    {
      "optionId": "rag",
      "nameZh": "RAG",
      "roleZh": "在回答时检索可更新、可追溯且受权限控制的外部知识。"
    },
    {
      "optionId": "fine_tuning",
      "nameZh": "微调",
      "roleZh": "用特定输入输出数据更新模型权重，以验证固定任务、行为、格式或风格增益。"
    },
    {
      "optionId": "hybrid",
      "nameZh": "RAG + 微调",
      "roleZh": "仅在知识缺口与行为缺口分别有证据且单项不足时组合。"
    }
  ],
  "decisionDimensions": [
    {
      "dimensionId": "task-gap",
      "nameZh": "用户任务与缺口",
      "questionZh": "缺的是外部知识、固定行为，还是两者都缺？",
      "requiredEvidenceZh": [
        "用户任务",
        "非 AI 或基础模型基线",
        "可复现错误样例",
        "目标结果"
      ],
      "evidenceIds": [
        "microsoft-rag-fine-tuning-choice-20260720"
      ]
    },
    {
      "dimensionId": "knowledge-change",
      "nameZh": "知识变化与撤回",
      "questionZh": "内容多久变化，纠错或删除多久必须生效？",
      "requiredEvidenceZh": [
        "更新频率",
        "删除时限",
        "索引更新记录",
        "重训触发器",
        "旧版本退役规则"
      ],
      "evidenceIds": [
        "microsoft-rag-fine-tuning-choice-20260720",
        "aws-rag-fine-tuning-comparison-20260720"
      ]
    },
    {
      "dimensionId": "provenance-access",
      "nameZh": "来源与权限",
      "questionZh": "用户是否需要打开依据，不同身份允许看到什么？",
      "requiredEvidenceZh": [
        "来源与版本",
        "权限矩阵",
        "引用打开测试",
        "训练数据权利责任角色确认"
      ],
      "evidenceIds": [
        "aws-rag-fine-tuning-comparison-20260720",
        "nist-ai-rmf-core-20260720"
      ]
    },
    {
      "dimensionId": "data-readiness",
      "nameZh": "数据条件",
      "questionZh": "有哪些文档、标注输入输出、关键切片和质量复核？",
      "requiredEvidenceZh": [
        "语料清单",
        "标注样本清单",
        "训练/开发验证/锁定最终测试划分",
        "去重、相似样例与泄漏检查",
        "最终测试暴露记录",
        "质量抽查"
      ],
      "evidenceIds": [
        "microsoft-rag-fine-tuning-choice-20260720"
      ]
    },
    {
      "dimensionId": "quality-failure",
      "nameZh": "质量与失败",
      "questionZh": "哪些指标、切片和严重错误阻断发布？",
      "requiredEvidenceZh": [
        "固定测试集",
        "逐例结果",
        "失败分类",
        "关键切片",
        "发布门槛"
      ],
      "evidenceIds": [
        "google-cloud-rag-evaluation-20260720"
      ]
    },
    {
      "dimensionId": "latency-cost",
      "nameZh": "延迟与成本",
      "questionZh": "峰值下的端到端延迟、单位成本和一次性投入分别是多少？",
      "requiredEvidenceZh": [
        "负载条件",
        "延迟分位数",
        "单位调用成本",
        "数据制作与训练投入",
        "更新维护成本"
      ],
      "evidenceIds": [
        "microsoft-rag-fine-tuning-choice-20260720",
        "aws-rag-fine-tuning-comparison-20260720"
      ]
    },
    {
      "dimensionId": "operations-rollback",
      "nameZh": "运行与回滚",
      "questionZh": "怎样从结果追踪到数据、检索、提示和模型版本，并恢复到已知版本？",
      "requiredEvidenceZh": [
        "版本清单",
        "日志字段",
        "告警",
        "降级与回滚演练",
        "复审日期"
      ],
      "evidenceIds": [
        "nist-ai-rmf-core-20260720"
      ]
    },
    {
      "dimensionId": "hybrid-necessity",
      "nameZh": "组合必要性",
      "questionZh": "RAG 与微调是否分别解决不同且已证明的缺口？",
      "requiredEvidenceZh": [
        "单项实验",
        "组合实验",
        "增量对照",
        "交叉故障",
        "简化条件"
      ],
      "evidenceIds": [
        "aws-rag-fine-tuning-comparison-20260720"
      ]
    }
  ],
  "experimentOrder": [
    {
      "experimentId": "A",
      "optionId": "baseline",
      "changeZh": "固定基础模型，优化提示、示例、规则或工具工作流。"
    },
    {
      "experimentId": "B",
      "optionId": "rag",
      "changeZh": "在同一生成基线上加入固定版本的检索上下文。"
    },
    {
      "experimentId": "C",
      "optionId": "fine_tuning",
      "changeZh": "用冻结训练集训练候选模型，并在共享开发/验证集比较。"
    },
    {
      "experimentId": "D",
      "optionId": "hybrid",
      "changeZh": "仅在 B、C 分别证明不同增益后组合。"
    }
  ],
  "finalGateRuleZh": "候选方案、模型和数据版本、推理参数、聚合方式与发布阈值全部冻结后，才使用锁定且未见的最终测试集。若查看最终结果后继续调整，则该集合必须重新标记为验证集，增加暴露次数，并更换新的未见最终测试集。",
  "decisionValues": [
    "not_evaluated",
    "continue",
    "stop",
    "adopt",
    "rollback"
  ],
  "blankExperimentRecord": {
    "experimentRecordId": "",
    "taskZh": "",
    "gapHypothesisZh": "",
    "optionId": "baseline",
    "baseModelVersion": "",
    "promptOrWorkflowVersion": "",
    "corpusVersion": "",
    "trainingSetVersion": "",
    "validationSetVersion": "",
    "finalTestSetVersion": "",
    "finalTestLockedAt": "",
    "finalTestExposureCount": 0,
    "inferenceParameters": {
      "temperature": null,
      "topP": null,
      "seed": null,
      "other": {}
    },
    "runCount": 1,
    "runAggregationMethodZh": "",
    "uncertaintySummaryZh": "",
    "loadProfile": "",
    "taskOutcome": null,
    "groundedOrSourceSupported": null,
    "behaviorOrFormatCompliant": null,
    "criticalFailures": [],
    "sliceResults": [],
    "latencyP50Ms": null,
    "latencyP95Ms": null,
    "unitCost": null,
    "oneTimeCost": null,
    "updateCost": null,
    "securityPrivacyFindings": [],
    "rollbackEvidenceUrls": [],
    "decision": "not_evaluated",
    "decisionReasonZh": "",
    "stopConditionsZh": [],
    "owner": "",
    "reviewedAt": "",
    "reviewAfter": ""
  },
  "sources": [
    {
      "evidenceId": "microsoft-rag-fine-tuning-choice-20260720",
      "title": "Augment large language models with retrieval-augmented generation or fine-tuning",
      "publisher": "Microsoft Learn",
      "sourceType": "first_party_technical_guidance",
      "publishedAt": "2026-01-30",
      "accessedAt": "2026-07-20",
      "url": "https://learn.microsoft.com/en-us/azure/developer/ai/augment-llm-rag-fine-tuning"
    },
    {
      "evidenceId": "aws-rag-fine-tuning-comparison-20260720",
      "title": "Comparing Retrieval Augmented Generation and fine-tuning",
      "publisher": "AWS Prescriptive Guidance",
      "sourceType": "first_party_technical_guidance",
      "accessedAt": "2026-07-20",
      "url": "https://docs.aws.amazon.com/prescriptive-guidance/latest/retrieval-augmented-generation-options/rag-vs-fine-tuning.html"
    },
    {
      "evidenceId": "google-cloud-rag-evaluation-20260720",
      "title": "Optimizing RAG retrieval: Test, tune, succeed",
      "publisher": "Google Cloud",
      "sourceType": "first_party_technical_guidance",
      "publishedAt": "2024-12-18",
      "accessedAt": "2026-07-20",
      "url": "https://cloud.google.com/blog/products/ai-machine-learning/optimizing-rag-retrieval?hl=en"
    },
    {
      "evidenceId": "nist-ai-rmf-core-20260720",
      "title": "AI Risk Management Framework Core",
      "publisher": "National Institute of Standards and Technology",
      "sourceType": "government_primary_guidance",
      "publishedAt": "2023-01-26",
      "accessedAt": "2026-07-20",
      "url": "https://airc.nist.gov/airmf-resources/airmf/5-sec-core/"
    }
  ],
  "title": "RAG、微调与混合方案决策矩阵 v1",
  "description": "用问题类型、更新频率、来源追溯、数据条件、评测、成本延迟、运维和组合必要性八项字段比较基础方案、RAG、微调与混合方案。",
  "url": "https://course.dragonai.tech/datasets/rag-fine-tuning-decision-matrix-v1.json",
  "creator": "烛龙智元内容研究组"
}