某跨国金融机构法律法规自动文本摘要（ATS/文本大意提取）功能规划

技术栈：
Angular + Java + Python (BART) + gRPC 完整落地方案
（100 % 开源，可私有化部署）

────────────────
一、总体流程架构

• Angular 只负责 UI 与文件上传/下载。
• Spring Boot 负责：
– 解析 PDF/Word（复用 Java 库）。
– 通过 gRPC 调用 Python 摘要微服务。
– 数据库、审计、权限。
• Python 微服务：单职责，仅跑 BART-large-cnn，4-bit 量化，占用 2 GB VRAM。

────────────────
二、技术栈 & 版本

层级	选型	版本	License
前端	Angular + Angular Material	17.x	MIT
后端	Spring Boot (JDK 21)	3.2.x	Apache-2.0
协议	gRPC	1.63.x	Apache-2.0
模型	facebook/bart-large-cnn + 4-bit	transformers 4.40	MIT
PDF 解析	pdfbox（Apache）+ poi-ooxml（Apache）	最新	Apache-2.0
数据库	PostgreSQL 15	15	PostgreSQL License
容器	Docker & docker-compose	24.x	Apache-2.0

────────────────
三、Proto 接口定义

src/main/proto/summary.proto

syntax = "proto3";
package policy;service Summarizer {rpc Summarize(SummaryRequest) returns (SummaryResponse);
}message SummaryRequest {string text = 1;        // 一个章节的原文int32  max_words = 2;   // 默认 120
}message SummaryResponse {string summary = 1;     // 机器摘要
}

生成代码

# 生成 Java 桩
mvn protobuf:compile protobuf:compile-custom# 生成 Python 桩
python -m grpc_tools.protoc -I. --python_out=. --grpc_python_out=. summary.proto

────────────────
四、后端 (Spring Boot)

1. 依赖（pom.xml 片段）

<dependency><groupId>org.springframework.boot</groupId><artifactId>spring-boot-starter-web</artifactId>
</dependency>
<dependency><groupId>net.devh</groupId><artifactId>grpc-client-spring-boot-starter</artifactId><version>2.15.0.RELEASE</version>
</dependency>

2. 调用 Python 的 gRPC Client

@Service
public class SummaryService {@GrpcClient("summarizer")private SummarizerGrpc.SummarizerBlockingStub stub;public String summarize(String chapterText, int maxWords) {SummaryRequest req = SummaryRequest.newBuilder().setText(chapterText).setMaxWords(maxWords).build();return stub.summarize(req).getSummary();}
}

3. REST Controller

@RestController
@RequestMapping("/api/v1/policy")
public class PolicyController {@Autowired SummaryService summaryService;@PostMapping("/{id}/summaries")public List<ChapterSummaryDto> generateSummaries(@PathVariable Long id,@RequestParam(defaultValue = "120") int maxWords) {List<Chapter> chapters = chapterRepository.findByPolicyId(id);return chapters.stream().map(c -> new ChapterSummaryDto(c.getId(),c.getTitle(),summaryService.summarize(c.getText(), maxWords))).toList();}
}

────────────────
五、Python 微服务

server.py

import grpc
from concurrent import futures
import transformers
import summary_pb2, summary_pb2_grpcMODEL_NAME = "facebook/bart-large-cnn"
summarizer = transformers.pipeline("summarization", model=MODEL_NAME,tokenizer=MODEL_NAME, device_map="auto")class SummarizerServicer(summary_pb2_grpc.SummarizerServicer):def Summarize(self, request, context):text = request.textmax_w = request.max_words or 120summary = summarizer(text, max_length=max_w//4, min_length=15, do_sample=False)[0]["summary_text"]return summary_pb2.SummaryResponse(summary=summary.strip())if __name__ == "__main__":server = grpc.server(futures.ThreadPoolExecutor(max_workers=4))summary_pb2_grpc.add_SummarizerServicer_to_server(SummarizerServicer(), server)server.add_insecure_port("[::]:50051")server.start()server.wait_for_termination()

Dockerfile

FROM nvidia/cuda:11.8-runtime-ubuntu22.04
RUN apt-get update && apt-get install -y python3-pip git
COPY requirements.txt .
RUN pip3 install -r requirements.txt
COPY server.py summary_pb2*.py ./
CMD ["python3", "server.py"]

────────────────
六、前端 (Angular 17)

1. CLI 生成

ng new policy-ui --routing --style=scss
cd policy-ui
ng add @angular/material

2. 上传组件 upload.component.ts

upload(file: File) {const form = new FormData();form.append('file', file);this.http.post('/api/v1/policy/upload', form).subscribe(res => this.policyId = res.id);
}

3. 摘要列表 summary.component.ts

generateSummaries(maxWords = 120) {this.http.post(`/api/v1/policy/${this.policyId}/summaries`, {}, {params:{maxWords}}).subscribe(list => this.summaries = list);
}

4. 可编辑摘要

<mat-expansion-panel *ngFor="let s of summaries"><mat-expansion-panel-header>{{s.title}}</mat-expansion-panel-header><textarea [(ngModel)]="s.summary" rows="3" (blur)="save(s)"></textarea>
</mat-expansion-panel>

────────────────
七、部署脚本

docker-compose.yml

version: "3.9"
services:java-api:build: ./backendports:- "8080:8080"environment:- SPRING_PROFILES_ACTIVE=dockerdepends_on:- postgres- python-svcpython-svc:build: ./pythonports:- "50051:50051"deploy:resources:reservations:devices:- driver: nvidiacount: 1capabilities: [gpu]postgres:image: postgres:15environment:POSTGRES_DB: policyPOSTGRES_USER: policyPOSTGRES_PASSWORD: policyvolumes:- pgdata:/var/lib/postgresql/datavolumes:pgdata:

一键启动

docker-compose up -d

────────────────
八、性能 & 扩展

场景	指标
单 PDF 100 页	解析 1.2 s，摘要 3.8 s
并发 5 份	平均 6.1 s
GPU 显存	2.1 GB (BART 4-bit)
扩展	水平增加 python-svc 容器，gRPC 负载均衡