feat(script): 话术三档架构(稳健/标准/深度)+ 目录重组 + input_snapshot 结构化
目录重组:shared/(脊柱周边 + 病种名/安全单一源 + 厚事实块)+ tiers/{stable,standard,deep}/。
脊柱(AiCallRunner)恒定,策略可插(每档一套 prompt/schema/format,深度为 3 步 pipeline)。
- input_snapshot 全字段结构化:reasons[].toothPositions(FDI)/medicalRecord(SOAP 全字段,per-reason);
recentTreatments/pendingTreatments/lastVisit 对象化;personaHighlights/guardian 补原始 key;
聚焦单一源 reasons[0](orchestrator 单一排序,prompt/fallback 不再各自 sort)
- 标准档:去模板 4 段自由编排(标题由 LLM 起)+ 厚输入(病历/其他reason/近期治疗),病种只给名
- 深度档:3 步(规划→写多段→独立对抗校验)+ repair(≤1)+ 兜底;多段不定输出(ScriptSectionDto.id→string)
- user prompt 补 {牙位}/本次目标(plan.goal)/医生医嘱·建议·治疗计划;≤18 禁拍片 belt
- 儿童复查段修:对齐本次病种 + 用{复查时长},删写死"3个月常规涂氟检查"
- 安全单一源 safety-rules.ts(禁词/承诺/加粗时间 + machineSafetyScan);composer tier-aware
- runner generateObject 解析失败重试一次再兜底(flash 多段 JSON 偶发脆)
- controller +tier query;实时教练后端上下文改读结构化字段
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Showing
This diff is collapsed.
Click to expand it.
This diff is collapsed.
Click to expand it.
This diff is collapsed.
Click to expand it.
Please
register
or
sign in
to comment