chore(prompts): regenerate collection prompt from workflow@20261021-r2

- merge main (fb9271e): collection prompt restructured (state-bound-to-action,
  gatekeeper principle); fill_missing eval cases relaxed to semantic regexes
- prompts/collection.md regenerated; byte-identity test updated and passing
- rerun golden evals: collection_skip 11/12, time_confirmation 1/2
  (inter-run variance; only single_no_inferred_count fails consistently -
  model routes without asking vehicle count after time confirmation)
This commit is contained in:
Eric Wang
2026-10-21 17:28:44 +08:00
parent 894cb220ed
commit a890f41f6f
2 changed files with 19 additions and 22 deletions
+2 -2
View File
@@ -384,12 +384,12 @@ docs/langgraph/
| 黄金案例 | 结果 | 说明 |
|---|---|---|
| `collection_skip_known_info` | 10/12 | 2 例:时间确认后漏问数量直接分流(模型顺序执行差异) |
| `collection_skip_known_info` | 11/12 | 存在轮次间波动;仅 `single_no_inferred_count`(双模式)稳定失败:时间确认后漏问数量直接分流 |
| `photo_skip_plate_confirm` | 4/4 | 20261021 步骤5(时间复述确认)修复了原“对”分歧 |
| `no_false`(asr,main 已移除旧 6 例) | 2/2 | — |
| `photo_button_guidance` | 12/12 | — |
| `non_motor_vehicle`(新增) | 14/14 | 依赖 formUpdate 管线(提取节点 + 更新表单差量语义) |
| `time_confirmation`(新增) | 1/2 | 1 例:模型话术“我记录的事故时间是”不匹配字面模板 |
| `time_confirmation`(新增) | 1/2 | 轮次间波动;失败例为话术字面偏离(main 已将 fill_missing 断言放宽为语义正则,此套件仍用字面模板) |
**统一失败模式**:开放问题(“您现在还在事故现场吗?”)未获明确回答时,用户输入短促词(“对”),评测期望保持 1002 澄清,模型直接分流 2000/2010。移植侧已排除提示词、历史、变量、温度(0 与缺省均试过)差异;模型名与 FastGPT 应用一致(deepseek-v4.1-flash)。