可解释问答：决策证据链

1. 进入 /decision 页面，输入问题并选数据源
2. 提交后等待 5 步链同步完成（loading 等待）
3. 查看 5 步进度、答案 markdown 与引用面板
4. 点击引用 URI 溯源到文档 chunk / 实体

{
  "code": 0,
  "data": {
    "answer": { "id": "<trace_id>", "question": "本季度销售额的计算口径是什么", "status": "done",
      "answer": "## 结论\n\n本月销售额 = 订单金额 - 退款金额 [1]，指已成交订单的实付金额 [2]。\n\n## 分析\n...\n\n## 建议\n...",
      "steps": "...", "citations": "...", "created_by": "u1", "created_at": "..." },
    "steps": [
      { "step": "decompose", "status": "ok", "detail": { "sub_questions": ["销售额如何定义", "是否含退款"], "llm_ok": true } },
      { "step": "retrieve", "status": "ok", "detail": { "rag": 2, "entity": 1, "nlq": "skipped_no_datasource", "candidates": 3, "fused": 3 } },
      { "step": "verify", "status": "ok", "detail": { "llm_ok": true, "llm_retries": 0, "supported": 2, "dropped": 1 } },
      { "step": "synthesize", "status": "ok", "detail": { "llm_ok": true, "evidence_used": 2, "answer_chars": 120 } },
      { "step": "cite", "status": "ok", "detail": { "citations": 2 } }
    ],
    "citations": [
      { "id": "<uuid>", "answer_id": "<trace_id>", "seq": 1, "uri": "doc://aip/3?chunk=2", "title": "销售指标口径说明", "snippet": "销售额 = 订单金额 - 退款金额", "score": 0.9, "source": "rag" },
      { "id": "<uuid>", "answer_id": "<trace_id>", "seq": 2, "uri": "object://aie/<entity_id>", "title": "销售额", "snippet": "实体「销售额」（类型: metric）", "score": 0.85, "source": "entity" }
    ]
  }
}

1. 提问完成后记录 answer.id（= trace_id）
2. 管理员打开 /admin/ai-audit?trace_id=<trace_id>
3. 核对 5 条 ai_decision 事件（stepType 逐个对应）
4. 查看 retrieve 各通道命中数与 verify 结论，定位问题

ref_type=ai_decision  event_type=DECISION_ANSWER  ref_id=<trace_id>
step1 decompose result=success  detail={"sub_questions":["销售额如何定义","是否含退款"]}
step2 retrieve result=success  detail={"rag":2,"entity":1,"nlq":"skipped_no_datasource","candidates":3,"fused":3}
step3 verify   result=success  detail={"llm_ok":true,"supported":2,"dropped":1}
step4 synthesize result=success detail={"evidence_used":2,"answer_chars":120}
step5 cite     result=success  detail={"citations":2}

1. 断开/失效 LLM key（或模拟网关 500）
2. 在 /decision 再次提交问题
3. 看答案前缀与 steps 状态（degraded / unverified）
4. 确认降级后不要直接当推理结论使用

answer: "> 说明：LLM 服务暂不可用，以下为检索证据摘要（未经推理与核验），仅供参考。\n1. **销售指标口径说明**（来源: rag, URI: doc://aip/3?chunk=2）\n2. **销售额**（来源: entity, URI: object://aie/<id>）..."

steps: [
  { "step": "decompose", "status": "degraded", "detail": { "sub_questions": ["本季度销售额的计算口径是什么"], "llm_ok": false } },
  { "step": "verify", "status": "degraded", "detail": { "llm_ok": false, "unverified": true, "supported": 2, "dropped": 0 } },
  { "step": "synthesize", "status": "degraded", "detail": { "llm_ok": false, "evidence_used": 2, "answer_chars": 120 } }
]

1. 在 /decision 输入知识底料里不存在的问题
2. 查看 retrieve 步骤状态与 detail
3. 查看 answer 与 citations（空引用）

answer: "未找到与该问题相关的证据，无法给出有依据的结论。"
steps: [
  { "step": "retrieve", "status": "degraded", "detail": { "rag": 0, "entity": 0, "nlq": "skipped_no_datasource", "note": "三路检索均无命中或均失败", "candidates": 0, "fused": 0 } }
]
citations: []