business-correctness-validator

核心功能: 本技能提供内容后/内容发布前/价格设定后/风控检查、无校验需求等能力。

business-correctness-validator — 业务正确性验证层

LLM输出后独立校验业务规则,作为LLM与业务落地之间的"最后一道防线"。校验失败→拒绝输出+告警;校验通过→放行。

使用场景

1. LLM生成营销文案/商品描述后,发布前校验内容合规性
2. LLM生成定价建议后,校验价格是否在品类风控区间内
3. 定时Cron任务校验租户风控指标(退款率/投诉率/虚假发货率)是否超阈值
4. 违规历史追溯查询,分析租户风控趋势
5. 多租户场景下,各租户独立风控阈值校验

工作流

1. 接收校验请求

   • 输入: content(内容)/price(价格)/metrics(风控指标), platform(平台), tenant_id(租户ID)
   • 验证: 参数非空+类型正确+平台有效
   • 异常: 参数无效→返回{success:false, error:"...", code:"INVALID_INPUT"}

2. 加载业务规则矩阵

   • 配置路径: config/business_rules_matrix.yaml
   • 降级: 配置文件不存在→使用内置默认规则(来源:01/02手册)
   • 来源: 02手册§二28平台矩阵 + 01手册§十风控阈值

3. 执行校验(按类型)

   • 内容校验: 长度+格式+品类禁令+绝对化用语(严格度分级)
   • 价格校验: 全局区间+品类区间+倾销/欺诈检测
   • 风控校验: 退款率+投诉率+虚假发货率+响应时长(多租户阈值)
   • 来源: 02手册§八8.2平台严格度 + 01手册§七7.3定价风控 + 01手册§十风控

4. 判定结果

   • pass: 无违规→放行
   • warning: 中低风险违规→告警但放行
   • blocked: 高危/严重违规→拒绝输出+记录违规
   • 风险等级: SAFE/LOW/MEDIUM/HIGH/CRITICAL

5. 记录违规(校验失败时)

   • 路径: data/business_violations/{tenant_id}_{date}.jsonl
   • 内容: 时间戳+租户ID+违规类型+详情
   • 用途: 违规历史追溯+风控趋势分析

6. 返回校验结果JSON

   • 结构: {success, data:{result, risk_level, violations, ...}, error, code}

输入格式

内容校验

{"content": "AI代写文案,专业润色", "platform": "xianyu", "tenant_id": "default"}

价格校验

{"price": 99.9, "category": "virtual", "platform": "xianyu"}

风控校验

{"metrics": "{\"refund_rate\":0.08,\"complaint_rate\":0.02}", "tenant_id": "default"}

输出格式

{
  "success": true,
  "data": {
    "result": "pass|warning|blocked",
    "risk_level": "SAFE|LOW|MEDIUM|HIGH|CRITICAL",
    "platform": "xianyu",
    "tenant_id": "default",
    "violation_count": 0,
    "violations": [],
    "validated_at": "2026-07-07T10:00:00",
    "rule_source": "02手册§二+§八8.1+§八8.2"
  },
  "error": null,
  "code": null
}

异常处理

业务规则来源

示例

示例1: 闲鱼内容校验通过

输入: {"content": "AI代写文案,专业润色服务", "platform": "xianyu", "tenant_id": "default"} 输出:

{
  "success": true,
  "data": {
    "result": "pass",
    "risk_level": "SAFE",
    "platform": "xianyu",
    "tenant_id": "default",
    "platform_strictness": 4,
    "content_length": 14,
    "violation_count": 0,
    "violations": [],
    "rule_source": "02手册§二+§八8.1+§八8.2"
  }
}

示例2: 小红书内容含绝对化用语被拦截

输入: {"content": "这是全网最好的产品,100%有效", "platform": "xiaohongshu", "tenant_id": "default"} 输出:

{
  "success": true,
  "data": {
    "result": "blocked",
    "risk_level": "HIGH",
    "platform": "xiaohongshu",
    "platform_strictness": 5,
    "violation_count": 2,
    "violations": [
      {"type": "absolute_word_violation", "severity": "high", "message": "内容包含广告法违禁绝对化用语: 最好", "keyword": "最好"},
      {"type": "absolute_word_violation", "severity": "high", "message": "内容包含广告法违禁绝对化用语: 100%", "keyword": "100%"}
    ]
  }
}

示例3: 价格倾销检测

输入: {"price": 0.05, "category": "service", "platform": "xianyu"} 输出:

{
  "success": true,
  "data": {
    "result": "blocked",
    "risk_level": "CRITICAL",
    "price": 0.05,
    "category": "service",
    "violation_count": 2,
    "violations": [
      {"type": "price_below_category_min", "severity": "medium", "message": "价格0.05低于service品类最低价1.0"},
      {"type": "suspicious_dumping_price", "severity": "critical", "message": "价格0.05低于service品类最低价50%,疑似倾销"}
    ]
  }
}

示例4: 风控阈值超标告警

输入: {"metrics": "{\"refund_rate\":0.08,\"complaint_rate\":0.02,\"false_shipment_rate\":0.01}", "tenant_id": "default"} 输出:

{
  "success": true,
  "data": {
    "result": "blocked",
    "risk_level": "HIGH",
    "tenant_id": "default",
    "violation_count": 1,
    "violations": [
      {"type": "refund_rate_exceeded_warn", "severity": "high", "message": "退款率8.00%超过告警阈值5.00%"}
    ],
    "rule_source": "01手册§十风控阈值"
  }
}

历史记录