lesson-audit-sampling-stratified-sampling-for-k.json

{
  "task_id": "lesson-audit-sampling-stratified-sampling-for-k",
  "title": "巡检题库分层抽样策略",
  "domain": "rag",
  "tags": [
    "rag",
    "audit",
    "sampling",
    "quality",
    "test-bank"
  ],
  "problem": "RAG 知识库质量巡检需要每天抽取少量题目进行自动测试。如果随机抽样，容易连续抽到同一类型的问题（如连续三天都是报警代码题），漏掉其他分类的质量退化。",
  "solution": "实现分层抽样策略：\n\n```\n题库按 level + tag 分层：\n  L2（跨文档）: 弧焊调试 / 型号对比 / IO通信 / 安全配置\n  L3（高难度）: 弧焊故障 / 选型 / IO故障 / 负载惯量 / TCP设定\n\n抽样规则：\n  1. 从 L2 随机抽 2 题\n  2. 从 L3 的每个 tag 中各取至多 1 题，优先覆盖不同 tag\n  3. 如果还不够 5 题 L3，从剩余补齐\n  4. 最终 7 题混排（不按难度顺序）\n```\n\n核心代码：\n\n```python\ndef sample_questions(bank, l2_count=2, l3_count=5):\n    l2 = [q for q in bank if q[\"level\"] == \"L2\"]\n    l3 = [q for q in bank if q[\"level\"] != \"L2\"]\n    l3_by_tag = group_by_key(l3, \"tag\")\n\n    selected = random.sample(l2, min(l2_count, len(l2)))\n    for tag in random.shuffle(list(l3_by_tag.keys())):\n        if len(selected) >= l2_count + l3_count:\n            break\n        selected.append(random.choice(l3_by_tag[tag]))\n\n    remaining = [q for q in l3 if q not in selected]\n    selected += remaining[:l2_count + l3_count - len(selected)]\n    random.shuffle(selected)\n    return selected\n```",
  "source": "lessons/contrib/audit-sampling-stratified-sampling-for-kb-inspection.md",
  "test_cmd": ""
}