AIclinicalresearch/docs/03-业务模块/DC-数据清洗整理/02-技术设计/技术设计文档：工具 B - 病历结构化机器人 (The AI Structurer).md

<EFBFBD><EFBFBD><EFBFBD>航挽霈⊥<EFBFBD>獢<EFBFBD><EFBFBD>撌亙<EFBFBD> B - <20><><EFBFBD>蝏𤘪<E89D8F><F0A498AA>𡝗㦤<F0A19D97>其犖 (The AI Structurer)

<EFBFBD><EFBFBD>﹝蝐餃<EFBFBD>	Technical Design Document (TDD)
撖孵<EFBFBD> PRD	PRD_撌亙<EFBFBD>B_<EFBFBD><EFBFBD><EFBFBD>蝏𤘪<EFBFBD><EFBFBD>𡝗㦤<EFBFBD>其犖_V2.md
<EFBFBD><EFBFBD>𧋦	V2.0 (<28>嗆<EFBFBD><E59786><EFBFBD>漣嚗𡁜<E59A97>璅∪<E79285>鈭文<E988AD>撉諹<E69289>)
*<EFBFBD>嗆<EFBFBD>?	Draft
<EFBFBD>詨<EFBFBD><EFBFBD>格<EFBFBD>	<EFBFBD><EFBFBD>遣銝<EFBFBD>銝芷<EFBFBD><EFBFBD>臭縑摨衣<EFBFBD><EFBFBD>餌<EFBFBD><EFBFBD><EFBFBD>𧋦蝏𤘪<EFBFBD><EFBFBD>硋<EFBFBD><EFBFBD>𠬍<EFBFBD><EFBFBD>朞<EFBFBD>**<2A>峕芋<E5B395>页<EFBFBD>DeepSeek & Qwen嚗匧僎<E58CA7>烐<EFBFBD><E78390>?銝?<EFBFBD>芸𢆡鈭文<EFBFBD>撉諹<EFBFBD>**嚗諹圾<E8ABB9>?AI 撟餉<E6929F><E9A489>桅<EFBFBD><E6A185>?

1. <20>颱<EFBFBD><E9A2B1>嗆<EFBFBD>霈曇恣 (Architecture Overview)

蝟餌<EFBFBD><EFBFBD>嗆<EFBFBD>隞𢛶<EFBFBD>𨅯<EFBFBD>蝥踵<EFBFBD>扳<EFBFBD>瘞渡瑪<EFBFBD>嘥<EFBFBD>蝥找蛹 **<2A>陖<EFBFBD>见僎<E8A781>烐<EFBFBD>瘞渡瑪<E6B8A1>?*<2A><>㺭<EFBFBD>株<EFBFBD><E6A0AA>亙<EFBFBD>嚗<EFBFBD><E59A97><EFBFBD>𤑳<EFBFBD>銝支葵銝滚<E98A9D><E6BB9A>?LLM 璅∪<E79285>撟嗉<E6929F>憭<EFBFBD><E686AD>嚗𣬚<E59A97><F0A3AC9A>𨀣<EFBFBD><F0A880A3>𡁜<EFBFBD><F0A1819C>𨅯<EFBFBD>蝒<EFBFBD><E89D92>瘚见<E7989A><E8A781>𢛶<EFBFBD>肽<EFBFBD>銵峕<E98AB5>撖對<E69296><E5B08D><EFBFBD><EFBFBD>舘<EFBFBD><E88898>箏<EFBFBD>鈭箏極撉諹<E69289>蝵烐聢<E78390>?

*1.1 蝟餌<E89D9F><E9A48C>嗆<EFBFBD><E59786>?

graph TD
Client[React <20>滨垢 (Grid & Drawer UI)]

subgraph API\_Server \[Fastify API <20>滚𦛚\]  
    JobAPI\[隞餃𦛚銝擧芋<E693A7><E88A8B>𦻖<EFBFBD>α]  
    VerifyAPI\[<5B>冽艶蝵烐聢<E78390>亙藁\]  
end  
  
subgraph Async\_Cluster \[<5B>𤾸蝱 Worker <20><>黎\]  
    BullMQ\[BullMQ 隞餃𦛚<E9A483>笔<EFBFBD>\]  
    Orchestrator\[隞餃𦛚蝻𡝗<E89DBB><F0A19D97>沔]  
    PII\_Engine\[<5B>鞟<EFBFBD><E99E9F>望<EFBFBD>撘閙<E69298>\]  
      
    subgraph Dual\_LLM\_Engine \[<5B>𣬚𤩅<F0A3AC9A>𣂼<EFBFBD>撘閙<E69298>\]  
        ClientA\[DeepSeek 摰Ｘ<E691B0>蝡珮]  
        ClientB\[Qwen 摰Ｘ<E691B0>蝡珮]  
    end  
      
    CrossValidator\[鈭文<E988AD>撉諹<E69289>/<2F>脩<EFBFBD>璉<EFBFBD>瘚见膥\]  
end  
  
subgraph Storage \[<5B>唳旿摮睃<E691AE>\]  
    PG\[(PostgreSQL \- 銝𡁜𦛚<F0A1819C>唳旿)\]  
    VectorDB\[(pgvector \- <20>舫<EFBFBD>㚁<EFBFBD><E39A81>其<EFBFBD>霂凋<E99C82>瘥𥪜笆)\]  
    Redis\[(Redis \- <20>笔<EFBFBD>)\]  
end

Client \--1.銝𠹺<E98A9D>&雿𤘪<E99BBF>--\> JobAPI  
JobAPI \--2.<2E>𥕦遣撟嗅<E6929F>隞餃𦛚--\> BullMQ  
BullMQ \--3.瘨<>晶--\> Orchestrator  
Orchestrator \--4.<2E>望<EFBFBD>--\> PII\_Engine  
PII\_Engine \--5.撟嗉<E6929F>靚<EFBFBD>鍂--\> ClientA & ClientB  
ClientA & ClientB \--6.餈𥪜<E9A488>JSON--\> CrossValidator  
CrossValidator \--7.霈∠<E99C88>銝<EFBFBD><E98A9D>湔<EFBFBD>?-\> PG  
Client \--8.<2E>匧<EFBFBD>蝵烐聢<E78390>唳旿--\> VerifyAPI  
VerifyAPI \--9.鈭箏極鋆<E6A5B5><E98B86>--\> PG

2. <20><><EFBFBD>舫<EFBFBD>匧<EFBFBD> (Tech Stack)

撅<EFBFBD>漣	<EFBFBD><EFBFBD><EFBFBD>舐<EFBFBD>隞?	<EFBFBD>匧<EFBFBD><EFBFBD><EFBFBD>眏
<EFBFBD>𡒊垢獢<EFBFBD>沲	Fastify 5.x	擃䀹<EFBFBD>扯<EFBFBD>撘<EFBFBD>郊 I/O嚗屸<E59A97><E5B1B8><EFBFBD>憭<EFBFBD><E686AD>擃睃僎<E79D83>烐芋<E78390>贝<EFBFBD><E8B49D>具<EFBFBD>?
璅∪<EFBFBD><EFBFBD>亙<EFBFBD>	LangChain.js	蝏煺<EFBFBD>撠<EFBFBD><EFBFBD> DeepSeek <20>?Qwen <20><><EFBFBD><EFBFBD>冽𦻖<E586BD><F0A6BB96><EFBFBD>靘蹂<E99D98><E8B982><EFBFBD>揢璅∪<E79285><E288AA>?
隞餃𦛚<EFBFBD>笔<EFBFBD>	BullMQ	<EFBFBD>詨<EFBFBD>蝏<EFBFBD>辣<EFBFBD><EFBFBD>2 <20><>閬<EFBFBD>⏚<EFBFBD>?Flow <20>蠘<EFBFBD><E8A098>𡝗<EFBFBD><F0A19D97>函<EFBFBD><E587BD>埝䔉摰䂿緵<E482BF>𦦵<EFBFBD>敺<EFBFBD>舅銝芣芋<E88AA3>钅<EFBFBD>餈𥪜<E9A488><F0A5AA9C>萘<EFBFBD><E89098>餉<EFBFBD><E9A489>?
*<EFBFBD>脩<EFBFBD>璉<EFBFBD>瘚?	Lodash (<28>箇<EFBFBD>) + Dice Coefficient (餈偦𧫴)	<EFBFBD>其<EFBFBD>瘥𥪜笆銝支葵 JSON 撖寡情<E5AFA1><E68385><EFBFBD>畾萄榆撘<E6A686><E69298><EFBFBD><EFBFBD><EFBFBD>祉㮾隡澆漲<E6BE86>臭蝙<E887AD>函<EFBFBD><E587BD>閧<EFBFBD> Dice 蝟餅㺭<E9A485>?Levenshtein 頝萘氖嚗峕<E59A97>銝漤<E98A9D>閬<EFBFBD><E996AC><EFBFBD>见<EFBFBD><E8A781>誩<EFBFBD><E8AAA9>?
*<EFBFBD>唳旿摨?	PostgreSQL 15	摮睃<EFBFBD> JSONB <20>澆<EFBFBD><E6BE86><EFBFBD><EFBFBD>璅∪<E79285>蝏𤘪<E89D8F><F0A498AA>?
<EFBFBD>滨垢鈭支<EFBFBD>	React + TanStack Table	V2 <20>嫣蛹<E5ABA3>冽艶蝵烐聢嚗峕㺭<E5B395>桅<EFBFBD>憭扳𧒄<E689B3><F0A79284>閬?TanStack Table (Headless) <20>滚<EFBFBD><E6BB9A>𡁏<EFBFBD>皛𡁜𢆡<F0A1819C>?

3. <20>詨<EFBFBD>瘚<EFBFBD><E7989A>霈曇恣 (Core Logic)

3.1 <20>箄<EFBFBD>雿𤘪<E99BBF> (Health Check Logic)

• *閫血<EFBFBD><EFBFBD>嗆㦤嚗? <20>冽<EFBFBD><E586BD>典<EFBFBD>蝡舫<E89DA1>㗇𥋘<E39787>𨀣<EFBFBD><F0A880A3>砍<EFBFBD><E7A08D>萘<EFBFBD><E89098>祇𡢿<E7A587>?
• *<EFBFBD>扯<EFBFBD><EFBFBD>餉<EFBFBD>嚗?
  1. <EFBFBD>𡒊垢霂餃<EFBFBD>霂亙<EFBFBD><EFBFBD><EFBFBD><EFBFBD> 100 銵䕘<E98AB5>銝滩粉<E6BBA9>券<EFBFBD>嚗剹<E59A97>?
  2. 霈∠<EFBFBD>蝏蠘恣<EFBFBD><EFBFBD><EFBFBD>嚗?
     • emptyRate: 蝛箏<E89D9B>?/ <20>餉<EFBFBD><E9A489>啜<EFBFBD>?
     • avgLength: <20>䂿征銵𣬚<E98AB5>撟喳<E6929F>摮㛖泵<E39B96>啜<EFBFBD>?
  3. *<EFBFBD>行⏛蝑𣇉裦嚗? <20>?emptyRate > 0.8 <20>?avgLength < 10嚗諹<E59A97><E8ABB9>?status: 'BAD'<27>?
  4. *Token 憸<>摯嚗? totalRows * avgLength * 1.5 (蝎㛖裦隡啁<E99AA1>)<29>?

3.2 <20>𣬚𤩅<F0A3AC9A>𣂼<EFBFBD>銝𦒘漱<F0A69298>厰<EFBFBD>霂?(Double-Blind & Validation)

餈蹱糓 V2 <20><><EFBFBD><EFBFBD>譌<EFBFBD>?

A. <20>鞟內霂滚極蝔?(Prompt Engineering)

銝箔<EFBFBD><EFBFBD>嫣噶瘥𥪜笆嚗<EFBFBD><EFBFBD>憿餃撩<EFBFBD>嗡舅銝芣芋<EFBFBD>贝<EFBFBD><EFBFBD>?*摰<><E691B0>銝<EFBFBD><E98A9D>渡<EFBFBD> JSON 蝏𤘪<E89D8F>**<2A>?

• System Prompt: "You are a medical structural extraction assistant..."
• Constraint: "Output strictly in JSON format. Keys must be: ['tumor_size', 'lymph_node', ...]."
• Temperature: 霈曆蛹 0嚗諹蕭瘙<E895AD><E79899>憭抒＆摰𡁏<E691B0>扼<EFBFBD>?

B. 鈭文<E988AD>撉諹<E69289>蝞埈<E89D9E> (The Judge)

敶?Model A (DeepSeek) <20>?Model B (Qwen) 餈𥪜<E9A488>蝏𤘪<E89D8F><F0A498AA>𠬍<EFBFBD><F0A0AC8D>扯<EFBFBD>瘥𥪜笆嚗? function validate(jsonA, jsonB) {
const conflicts = [];
const keys = Object.keys(jsonA);

for (const key of keys) {
const valA = normalize(jsonA[key]); // 敶雴<E695B6><E99BB4>吔<EFBFBD><E59094>駁膄蝛箸聢<E7AEB8><E881A2>蓮撠誩<E692A0><E8AAA9><EFBFBD><EFBFBD>閫鍦<E996AB>
const valB = normalize(jsonB[key]);

// 1\. 蝎曄＆<E69B84>寥<EFBFBD>  
if (valA \=== valB) continue;  
  
// 2\. <20>啣<EFBFBD>澆<EFBFBD>銝<EFBFBD><E98A9D>硋龪<E7A18B>?(憒?"3cm" vs "3.0cm")  
if (isNumber(valA) && isNumber(valB) && parse(valA) \=== parse(valB)) continue;  
  
// 3\. (<28>舫<EFBFBD>? 霂凋<E99C82><E5878B>訾撮摨血龪<E8A180>? 
// if (similarity(valA, valB) \> 0.95) continue;  
  
conflicts.push(key);  

}

return conflicts.length === 0 ? 'CLEAN' : 'CONFLICT';
}

4. <20>唳旿摨栞挽霈?(Database Schema)

V2 <20><>閬<EFBFBD><E996AC><EFBFBD>其舅隞?AI 蝏𤘪<E89D8F>隞亙<E99A9E><E4BA99>冽<EFBFBD><E586BD><EFBFBD><EFBFBD><EFBFBD>喟<EFBFBD><E5969F>栶<EFBFBD>?

Prisma Schema <20>湔鰵

// 隞餃𦛚銵? model ExtractionJob {
id String @id @default(uuid())
// ...<2E>嗡<EFBFBD>摮埈挾
diseaseType String // <20>曄<EFBFBD>蝐餃<E89D90> (<28>箇<EFBFBD>)
reportType String // <20>亙<EFBFBD>蝐餃<E89D90> (<28><><EFBFBD>)
targetFields Json // <20>格<EFBFBD>摮埈挾摰帋<E691B0> [{name: "<22>輻𠈔憭批<E686AD>", desc: "..."}]
}

// <20>閗<EFBFBD>霈啣<E99C88>銵? model ExtractionItem {
id String @id @default(uuid())
jobId String
originalText String @db.Text

// V2 <20>詨<EFBFBD>摮埈挾
resultA Json? // DeepSeek 蝏𤘪<E89D8F> { "size": "3cm" }
resultB Json? // Qwen 蝏𤘪<E89D8F> { "size": "3.0 cm" }

// <20>脩<EFBFBD>璉<EFBFBD>瘚讠<E7989A><E8AEA0>? status ItemStatus // PENDING, CLEAN, CONFLICT, RESOLVED
conflictFields String[] // ["size"] 霈啣<E99C88><E595A3>芯<EFBFBD>摮埈挾<E59F88>脩<EFBFBD>鈭?

// <20><>蝏<EFBFBD><E89D8F>蝥喟<E89DA5><E5969F>?(<28>冽<EFBFBD>鋆<EFBFBD><E98B86><EFBFBD>𤾸<EFBFBD><F0A4BEB8>伐<EFBFBD><E4BC90>𤥁<EFBFBD><F0A4A581><EFBFBD><EFBFBD>湔𧒄<E6B994>芸𢆡<E88AB8>坔<EFBFBD>)
finalResult Json?
}

5. <20>亙藁霈曇恣 (API Endpoints)

*5.1 璅∠<E79285>銝𡡞<E98A9D>蝵?

• GET /api/templates: <20>瑕<EFBFBD>憸<EFBFBD>挽<EFBFBD><E68CBD>𪆴<EFBFBD><F0AA86B4><EFBFBD><EFBFBD>亙<EFBFBD>璅∠<E79285><E288A0>𡑒”<F0A19192>?
• POST /api/jobs: <20>𥕦遣隞餃𦛚嚗釶ayload 銝剝<E98A9D><E5899D><EFBFBD>鉄 diseaseType <20>?reportType嚗䔶噶鈭𤾸<E988AD>蝡舐<E89DA1>鋆?Prompt<70>?

5.2 蝵烐聢撉諹<E69289> (Grid Verification)

• GET /api/jobs/:id/rows: <20><>△<EFBFBD>瑕<EFBFBD>撉諹<E69289><E8ABB9>唳旿<E594B3>?
  • Response: 餈𥪜<E9A488> originalText, resultA, resultB, conflictFields<64>?
• POST /api/items/:id/resolve: <20>閗<EFBFBD>鋆<EFBFBD><E98B86><EFBFBD>?
  • Payload: { field: "tumor_size", chosenValue: "3cm" }<7D>?
  • Logic: <20>湔鰵 finalResult嚗<74><E59A97><EFBFBD>𡏭砲銵峕<E98AB5><E5B395>匧<EFBFBD>蝒<EFBFBD><E89D92>畾菟<E795BE>撌脰圾<E884B0>喉<EFBFBD>撠?status <20>湔鰵銝?RESOLVED<45>?

6. <20>滨垢霂衣<E99C82>霈曇恣 (Frontend)

6.1 <20>冽艶撉諹<E69289>蝵烐聢 (Verification Grid)

• *蝏<EFBFBD>辣<EFBFBD>匧<EFBFBD>嚗? 靘萘<E99D98><E89098>刻<EFBFBD> TanStack Table (<28>餉<EFBFBD>撅? + *UI 蝏<>辣摨? (皜脫<E79A9C>撅?<3F>?
• <EFBFBD>脩<EFBFBD><EFBFBD>訫<EFBFBD><EFBFBD>潭葡<EFBFBD>橒<EFBFBD>
  • 敶?conflictFields.includes(column.id) <20>塚<EFBFBD><E5A19A>訫<EFBFBD><E8A8AB>潭葡<E6BDAD>㮖蛹撖寞<EFBFBD>璅∪<EFBFBD><EFBFBD>?
  • <EFBFBD>曄內銝支葵撠𤩺<EFBFBD><EFBFBD>殷<EFBFBD>[DS: 3cm] <20>?[QW: 3.0cm]<5D>?
  • <EFBFBD>冽<EFBFBD><EFBFBD>孵稬隞颱<EFBFBD><EFBFBD>厰僼嚗諹圻<EFBFBD>?resolve API嚗<49><E59A97>蝡臭<E89DA1>閫<EFBFBD>凒<EFBFBD>堆<EFBFBD>Optimistic Update嚗劐蛹<E58A90>劐葉<E58A90>嗆<EFBFBD><E59786><EFBFBD>?

6.2 靘扯器<E689AF>誩<EFBFBD><E8AAA9>?(Context Drawer)

• *閫血<EFBFBD>嚗? <20>孵稬銵冽聢銵𣬚<E98AB5>蝛箇蒾憭<E892BE><E686AD><EFBFBD>𨀣䰻<F0A880A3>见<EFBFBD><E8A781><EFBFBD><EFBFBD>嘥㦛<E598A5><E3A69B><EFBFBD>?
• *<EFBFBD>蠘<EFBFBD>嚗? 撅閧內 originalText<78>?
• *擃䀝漁隡睃<EFBFBD>嚗? 蝞<><E89D9E>訫<EFBFBD><E8A8AB>?String.indexOf <20>交𪄳敶枏<E695B6>摮埈挾<E59F88><E68CBE><EFBFBD>澆僎<E6BE86><E5838E><EFBFBD><EFBFBD>?

7. 憌𡡞埯<F0A1A19E>批<EFBFBD>銝擧<E98A9D>扯<EFBFBD>隡睃<E99AA1>

瞏𨅯銁憌𡡞埯	閫<EFBFBD><EFBFBD><EFBFBD>寞<EFBFBD>
<EFBFBD><EFBFBD><EFBFBD>?Token <20>鞉𧋦	1. 暺䁅恕雿輻鍂 DeepSeek (<28><><EFBFBD><EFBFBD>鞉𧋦) + Qwen (雿擧<E99BBF><E693A7>? 蝏<><E89D8F><EFBFBD>?2. <20>兩<EFBFBD>靝<EFBFBD>璉<EFBFBD><E79289>嗪𧫴畾萎艇<E8908E>潭㜃<E6BDAD>芣<EFBFBD><E88AA3><EFBFBD>㺭<EFBFBD>柴<EFBFBD>?
*憭<EFBFBD><EFBFBD><EFBFBD>笔漲<EFBFBD>?	銝支葵璅∪<EFBFBD>敹<EFBFBD>◆ **撟嗅<E6929F>靚<EFBFBD>鍂 (Promise.all)**嚗諹<E59A97>䔶<EFBFBD><E494B6>臭葡銵䎚<E98AB5><E48E9A>㟲雿栞<E99BBF>埈𧒄<E59F88>硋<EFBFBD>鈭擧<E988AD><E693A7>Ｙ<EFBFBD><EFBCB9><EFBFBD>葵璅∪<E79285><E288AA>?
*璅∪<EFBFBD><EFBFBD>澆<EFBFBD>銝滚𨯬霂?	Prompt 銝剖<E98A9D><E58996>?Few-Shot (撠烐甅<E78390>? 蝷箔<E89DB7>嚗峕<E59A97>蝖桀<E89D96>蝷?JSON <20>澆<EFBFBD><E6BE86><EFBFBD><EFBFBD><EFBFBD>?JSON 閫<><E996AB>憭梯揖嚗諹䌊<E8ABB9>券<EFBFBD>霂?1 甈～<E79488>?
<EFBFBD>滨垢蝵烐聢<EFBFBD>⊿▼	憒<EFBFBD><EFBFBD><EFBFBD>唳旿頞<EFBFBD><EFBFBD> 1000 <20>∴<EFBFBD>撘<EFBFBD><E69298>?Virtual Scrolling (<28>𡁏<EFBFBD>皛𡁜𢆡)<29>?