问题
文档负担是临床医生倦怠最大的单一驱动因素。现有抄录工具能生成文字——但不是临床医生可签署的结构化笔记。指向源话语片段的引用基本不存在。
我们的做法
我们搭了一条抄录管线,带两条不可妥协的不变量:
- 每一条主张都有来源。 临床医生可以把鼠标悬停在任何一句话上,看到产生它的环境转录 span。
- 模板归临床团队所有。 笔记结构(SOAP、叙事式等)按专科在 CMS 里定义;智能体填充槽位。
我们交付的第一版有 12% 的引用失败率。第三版降到 0.8%。evals 抓住了每一次回归。— 工程负责人, 美国医疗科技
架构
02 · Reference architecture
What we actually ship.
Every system we build follows this shape. Client at the edge, tools in a sandbox, traces everywhere, evaluators gating output. No black boxes, no "it works on my machine."
EDGE
Client
web · mobile · API
Gateway
auth · rate-limit · PII redact
ORCHESTRATION
Orchestrator
planner · router · memory
AGENTS & TOOLS
Retrieval
hybrid · rerank · cite
Tool calls
sandboxed · timeboxed
Evaluator
gates · rubrics · LLM-judge
DATA & TRUST
Vector + BM25
tenant-isolated
Traces / logs
OTel · replayable
Signed output
auditable · rollback
一次在该系统上的运行
04 · Run
Watch an agent do the job.
Three real production scenarios, replayed at observed latency. Every box is a span; every span has tokens, cost, and an eval gate. This is what shows up in your traces, not a marketing animation.
POST/api/v1/agent/run7-minute ambient consult → clinician-ready note with source-linked citations.
trace · clinicalspan_id 7c1f…
orchestrator.run
PLANorchestrator.plan
Plan rationale
Scribe pipeline: ASR with medical lexicon, retrieve patient context + template, draft per-SOAP-section, evaluator gates clinical-safety claims.
Subtasks
tool.asrretrieval.contextreasonerevaluator
0ms/3.07s
LATENCY0msbudget 3.00s
TOKENS0in + out
COST$0.0000budget $0.025
EVAL GATE—deterministic + LLM-judge
READY · clinician review
SOAP note drafted. 14 citations resolved. 2 low-confidence claims surfaced for clinician.
latency3.12scost$0.022tokens3,104evals9/9
成果
0.0h
每位医生每周节省时间
0.0%
引用失败率
0%
医生签署率