{
  "metadata": {
    "name": "GCPI RAG分层评测样本",
    "version": "1.0.0",
    "date": "2026-08-10",
    "license": "CC BY 4.0",
    "license_url": "https://creativecommons.org/licenses/by/4.0/",
    "attribution": "GCPI RAG分层评测样本v1.0.0，绿色碳汇规划院（北京）有限公司，2026-08-10",
    "origin": "由绿色碳汇规划院（北京）有限公司依据公开RAG方法独立合成",
    "privacy": "不含客户数据、真实个人信息、未公开业务资料或付费语料",
    "purpose": "用于展示RAG召回、版本、引用、拒答、权限、安全与运营评测的结构；不是客户基准或产品准确率声明"
  },
  "failure_reviews": [
    {
      "id": "F01",
      "stage": "知识发布",
      "symptom": "草案被用作正式答案",
      "root_cause": "索引未按publication_status过滤",
      "remediation": "只对published状态建生产索引，检索前再校验",
      "regression": "未发布草案召回率必须为0"
    },
    {
      "id": "F02",
      "stage": "查询理解",
      "symptom": "标准号和常规词混淆",
      "root_cause": "对稀有识别符只做语义向量检索",
      "remediation": "增加字段化精确检索与识别符路由",
      "regression": "预标注编号集的Recall@5达到质量门槛"
    },
    {
      "id": "F03",
      "stage": "分块",
      "symptom": "回答丢失例外与适用条件",
      "root_cause": "条款与紧随其后的例外被分到无父子关系的块",
      "remediation": "按文档结构分块并保留父标题、交叉引用和邻接条款",
      "regression": "否定、例外和附件用例的完整证据覆盖达标"
    },
    {
      "id": "F04",
      "stage": "权限",
      "symptom": "页面隐藏了无权结果但模型上下文仍包含",
      "root_cause": "权限仅在展示层执行",
      "remediation": "在检索、重排、上下文、缓存、显示和导出层做完整调解",
      "regression": "跨租户测试的检索与生成泄漏均为0"
    },
    {
      "id": "F05",
      "stage": "重排",
      "symptom": "过期版本排在现行版前",
      "root_cause": "排序只看语义相似度，未纳入效力、时间和地域",
      "remediation": "将过滤前置并把效力与版本作为硬约束，不只是加分项",
      "regression": "版本冲突集中所有结果都保留正确效力和适用域"
    },
    {
      "id": "F06",
      "stage": "引用生成",
      "symptom": "来源相关但不支持答案中的关键主张",
      "root_cause": "引用按段落绑定，没有主张级支持校验",
      "remediation": "先生成主张—证据对并校验直接支持，不支持则删除或降级",
      "regression": "高风险主张的引用支持率必须达到门槛且无虚构定位"
    },
    {
      "id": "F07",
      "stage": "结果状态",
      "symptom": "证据空缺时仍输出确定final",
      "root_cause": "答案只有文本字段，无证据充足性和拒答契约",
      "remediation": "引入final、provisional、refused及缺失条件、风险和人工转交字段",
      "regression": "拒答集和条件补问集不得输出无依据final"
    },
    {
      "id": "F08",
      "stage": "安全",
      "symptom": "检索文档中的指令触发了额外工具调用",
      "root_cause": "没有将检索内容当作不可信数据，工具权限过宽",
      "remediation": "隔离系统指令与文档内容，收窄工具与数据出域，增加注入检测和安全停止",
      "regression": "间接注入集中越权调用和数据外发均为0"
    },
    {
      "id": "F09",
      "stage": "索引发布",
      "symptom": "新索引上线后罕见术语召回突降",
      "root_cause": "直接覆盖唯一生产索引，无影子评测和回滚",
      "remediation": "新建版本化影子索引，跑分层评测、灰度切换并保留旧版",
      "regression": "发布前所有黄金子集达门槛，回滚演练成功"
    },
    {
      "id": "F10",
      "stage": "运营反馈",
      "symptom": "用户多次纠正同类错误但系统持续重复",
      "root_cause": "反馈只保存文本，未关联trace_id、知识版本和失败阶段",
      "remediation": "将反馈归因到知识、召回、排序、引用或生成，建立修复负责人和回归用例",
      "regression": "每个已确认故障都有状态、修复版本和至少一条回归用例"
    }
  ]
}
