当前位置： 首页 > news >正文

多语言编码Agent解决方案（2）-后端服务实现

news 2025/9/15 8:21:27

后端服务实现：支持多语言的vLLM服务

本部分包含后端服务的完整实现，包括vLLM模型服务、多语言支持和API接口。

1. 后端目录结构

backend/
├── server.py          # FastAPI服务入口
├── agent.py           # Agent逻辑实现
├── i18n.py            # 国际化支持
├── locales/           # 语言文件
│   ├── en.json
│   ├── zh-CN.json
│   └── ja.json
└── requirements.txt   # 依赖项

2. 依赖项清单 (requirements.txt)

fastapi==0.104.1
uvicorn==0.24.0
vllm==0.2.7
transformers==4.36.0
torch==2.1.0
pydantic==2.5.0
python-multipart==0.0.6
pyyaml==6.0.1

3. 国际化支持模块 (i18n.py)

"""
国际化（i18n）支持模块
"""
import json
from pathlib import Path
from typing import Dict, Anyclass I18n:"""国际化管理类"""def __init__(self, default_lang: str = 'en'):self.default_lang = default_langself.current_lang = default_langself.translations: Dict[str, Dict] = {}self.load_translations()def load_translations(self):"""加载所有语言文件"""locales_dir = Path(__file__).parent / 'locales'for lang_file in locales_dir.glob('*.json'):lang_code = lang_file.stemtry:with open(lang_file, 'r', encoding='utf-8') as f:self.translations[lang_code] = json.load(f)except Exception as e:print(f"Failed to load language file {lang_file}: {e}")def set_language(self, lang: str):"""设置当前语言"""if lang in self.translations:self.current_lang = langelse:# 尝试匹配语言代码前缀for available_lang in self.translations.keys():if lang.startswith(available_lang.split('-')[0]):self.current_lang = available_langreturnself.current_lang = self.default_langdef get(self, key: str, **kwargs) -> str:"""获取翻译文本"""keys = key.split('.')# 获取当前语言的翻译translation = self.translations.get(self.current_lang, {})# 逐级获取嵌套的值for k in keys:if isinstance(translation, dict):translation = translation.get(k)else:break# 如果找不到，尝试默认语言if translation is None:translation = self.translations.get(self.default_lang, {})for k in keys:if isinstance(translation, dict):translation = translation.get(k)else:break# 如果还是找不到，返回keyif translation is None:return key# 格式化字符串if kwargs:try:return translation.format(**kwargs)except:return translationreturn translationdef get_language_from_header(self, accept_language: str) -> str:"""从Accept-Language请求头获取语言"""if not accept_language:return self.default_lang# 解析Accept-Language头languages = []for lang in accept_language.split(','):parts = lang.strip().split(';')language = parts[0]if len(parts) > 1 and parts[1].startswith('q='):priority = float(parts[1][2:])else:priority = 1.0languages.append((priority, language))# 按优先级排序languages.sort(reverse=True)# 找到第一个支持的语言for _, lang_code in languages:# 精确匹配if lang_code in self.translations:return lang_code# 语言代码匹配（如zh匹配zh-CN）lang_prefix = lang_code.split('-')[0].lower()for available_lang in self.translations.keys():if available_lang.lower().startswith(lang_prefix):return available_langreturn self.default_lang# 全局i18n实例
i18n = I18n()# 便捷函数
def _(key: str, **kwargs) -> str:"""翻译函数的快捷方式"""return i18n.get(key, **kwargs)

4. 语言文件 (locales)

en.json

{"server": {"loading_model": "Loading model: {model}","model_loaded": "Model loaded successfully","agent_not_initialized": "Agent not initialized","unsupported_action": "Unsupported action: {action}","request_failed": "Failed to process request: {error}","server_healthy": "Server is healthy","starting": "Starting Coding Agent Backend..."},"agent": {"actions": {"complete": "Code Completion","generate": "Code Generation","explain": "Code Explanation","refactor": "Code Refactoring","debug": "Debug Analysis","test": "Test Generation"},"prompts": {"complete": "Complete the following {language} code:\n\n{code_before}\n<CURSOR>\n{code_after}\n\nContinue from <CURSOR>:","generate": "You are an expert programmer. Generate {language} code based on the following instruction:\n\nInstruction: {instruction}\nContext: {context}\n\nPlease provide clean, well-commented code:\n```{language}\n","explain": "Explain the following {language} code in detail:\n\n```{language}\n{code}\n```\n\nProvide a clear explanation:","refactor": "Refactor the following {language} code to improve readability, performance, and follow best practices:\n\nOriginal code:\n```{language}\n{code}\n```\n\nRefactored code with improvements:\n```{language}\n","debug": "Debug the following {language} code and identify any issues:\n\nCode:\n```{language}\n{code}\n```\n\nIdentify bugs and provide fixes:","test": "Generate comprehensive unit tests for the following {language} code:\n\nCode to test:\n```{language}\n{code}\n```\n\nTest code:\n```{language}\n"},"suggestions": {"simplified": "Code has been simplified","extracted_function": "Extracted into functions","improved_format": "Improved code formatting"}}
}

zh-CN.json

{"server": {"loading_model": "正在加载模型: {model}","model_loaded": "模型加载成功","agent_not_initialized": "Agent未初始化","unsupported_action": "不支持的操作: {action}","request_failed": "处理请求失败: {error}","server_healthy": "服务器运行正常","starting": "正在启动编码助手后端…"},"agent": {"actions": {"complete": "代码补全","generate": "代码生成","explain": "代码解释","refactor": "代码重构","debug": "调试分析","test": "测试生成"},"prompts": {"complete": "补全以下{language}代码:\n\n{code_before}\n<CURSOR>\n{code_after}\n\n从<CURSOR>继续:","generate": "你是一位专业的程序员。根据以下指令生成{language}代码:\n\n指令: {instruction}\n上下文: {context}\n\n请提供清晰、注释完善的代码:\n```{language}\n","explain": "详细解释以下{language}代码:\n\n```{language}\n{code}\n```\n\n请提供清晰的解释:","refactor": "重构以下{language}代码，提高可读性、性能并遵循最佳实践:\n\n原始代码:\n```{language}\n{code}\n```\n\n改进后的代码:\n```{language}\n","debug": "调试以下{language}代码并识别问题:\n\n代码:\n```{language}\n{code}\n```\n\n识别错误并提供修复方案:","test": "为以下{language}代码生成全面的单元测试:\n\n待测试代码:\n```{language}\n{code}\n```\n\n测试代码:\n```{language}\n"},"suggestions": {"simplified": "代码已简化","extracted_function": "已提取为函数","improved_format": "改进了代码格式"}}
}

ja.json

{"server": {"loading_model": "モデルをロード中: {model}","model_loaded": "モデルのロードが完了しました","agent_not_initialized": "エージェントが初期化されていません","unsupported_action": "サポートされていないアクション: {action}","request_failed": "リクエストの処理に失敗しました: {error}","server_healthy": "サーバーは正常に動作しています","starting": "コーディングエージェントバックエンドを起動中…"},"agent": {"actions": {"complete": "コード補完","generate": "コード生成","explain": "コード説明","refactor": "コードリファクタリング","debug": "デバッグ分析","test": "テスト生成"},"prompts": {"complete": "次の{language}コードを補完してください:\n\n{code_before}\n<CURSOR>\n{code_after}\n\n<CURSOR>から続けてください:","generate": "あなたはエキスパートプログラマーです。次の指示に基づいて{language}コードを生成してください:\n\n指示: {instruction}\nコンテキスト: {context}\n\nクリーンでコメント付きのコードを提供してください:\n```{language}\n","explain": "次の{language}コードを詳しく説明してください:\n\n```{language}\n{code}\n```\n\n明確な説明を提供してください:","refactor": "次の{language}コードをリファクタリングして、可読性、パフォーマンスを向上させ、ベストプラクティスに従ってください:\n\n元のコード:\n```{language}\n{code}\n```\n\n改善されたコード:\n```{language}\n","debug": "次の{language}コードをデバッグして問題を特定してください:\n\nコード:\n```{language}\n{code}\n```\n\nバグを特定して修正を提供してください:","test": "次の{language}コードに対して包括的な単体テストを生成してください:\n\nテスト対象コード:\n```{language}\n{code}\n```\n\nテストコード:\n```{language}\n"},"suggestions": {"simplified": "コードが簡略化されました","extracted_function": "関数として抽出されました","improved_format": "コード形式が改善されました"}}
}

5. Agent逻辑实现 (agent.py)

from vllm import SamplingParams
import re
from datetime import datetime
from typing import List
from .i18n import i18n, _class CodeRequest:def __init__(self, action, code="", context="", language="python", cursor_position=0, instruction="", max_tokens=512, temperature=0.7):self.action = actionself.code = codeself.context = contextself.language = languageself.cursor_position = cursor_positionself.instruction = instructionself.max_tokens = max_tokensself.temperature = temperatureclass CodingAgent:"""编码助手Agent（多语言支持）"""def __init__(self, engine, tokenizer):self.engine = engineself.tokenizer = tokenizerself.conversation_history = []async def generate_completion(self, request: CodeRequest, lang: str = 'en') -> dict:i18n.set_language(lang)prompt = i18n.get("agent.prompts.complete", language=request.language, code_before=request.code[:request.cursor_position], code_after=request.code[request.cursor_position:])result = await self._generate(prompt, request.max_tokens, request.temperature)code = self._extract_code(result)return {"result": code,"confidence": 0.85,"suggestions": [code],"metadata": {"action": i18n.get("agent.actions.complete"), "timestamp": datetime.now().isoformat()}}async def generate_code(self, request: CodeRequest, lang: str = 'en') -> dict:i18n.set_language(lang)prompt_suffix = "" if lang == 'en' else ("\nPlease comment in " + lang) if lang == 'ja' else "\n请用中文注释。"prompt = i18n.get("agent.prompts.generate", language=request.language, instruction=request.instruction, context=request.context) + prompt_suffixresult = await self._generate(prompt, request.max_tokens, request.temperature)code = self._extract_code(result)return {"result": code,"confidence": 0.9,"metadata": {"action": i18n.get("agent.actions.generate"), "language": request.language}}async def explain_code(self, request: CodeRequest, lang: str = 'en') -> dict:i18n.set_language(lang)prompt = i18n.get("agent.prompts.explain", language=request.language, code=request.code)result = await self._generate(prompt, 512, 0.7)return {"result": result,"confidence": 0.95,"metadata": {"action": i18n.get("agent.actions.explain")}}async def refactor_code(self, request: CodeRequest, lang: str = 'en') -> dict:i18n.set_language(lang)prompt = i18n.get("agent.prompts.refactor", language=request.language, code=request.code)result = await self._generate(prompt, request.max_tokens, 0.5)code = self._extract_code(result)suggestions = self._generate_suggestions(request.code, code, lang)return {"result": code,"confidence": 0.8,"suggestions": suggestions,"metadata": {"action": i18n.get("agent.actions.refactor")}}async def debug_code(self, request: CodeRequest, lang: str = 'en') -> dict:i18n.set_language(lang)prompt = i18n.get("agent.prompts.debug", language=request.language, code=request.code)result = await self._generate(prompt, request.max_tokens, 0.3)return {"result": result,"confidence": 0.85,"metadata": {"action": i18n.get("agent.actions.debug")}}async def generate_tests(self, request: CodeRequest, lang: str = 'en') -> dict:i18n.set_language(lang)prompt = i18n.get("agent.prompts.test", language=request.language, code=request.code)result = await self._generate(prompt, request.max_tokens, 0.5)code = self._extract_code(result)return {"result": code,"confidence": 0.8,"metadata": {"action": i18n.get("agent.actions.test")}}async def _generate(self, prompt: str, max_tokens: int, temperature: float) -> str:sampling_params = SamplingParams(temperature=temperature,max_tokens=max_tokens,top_p=0.95,frequency_penalty=0.1,)request_id = f"req-{datetime.now().timestamp()}"results_generator = self.engine.generate(prompt, sampling_params, request_id)async for request_output in results_generator:if request_output.finished:return request_output.outputs[0].textreturn ""def _extract_code(self, text: str) -> str:code_pattern = r'```[\w]*\n(.*?)```'matches = re.findall(code_pattern, text, re.DOTALL)return matches[0].strip() if matches else text.strip()def _generate_suggestions(self, original: str, refactored: str, lang: str) -> List[str]:i18n.set_language(lang)suggestions = []if len(refactored) < len(original):suggestions.append(i18n.get("agent.suggestions.simplified"))if "def " in refactored and "def " not in original:suggestions.append(i18n.get("agent.suggestions.extracted_function"))if refactored.count("\n") > original.count("\n"):suggestions.append(i18n.get("agent.suggestions.improved_format"))return suggestions

6. FastAPI服务入口 (server.py)

from fastapi import FastAPI, HTTPException, Header, Request
from fastapi.middleware.cors import CORSMiddleware
from pydantic import BaseModel
from typing import Optional, List, Dict, Any
import uvicorn
import asyncio
from vllm import AsyncLLMEngine, SamplingParams, AsyncEngineArgs
from transformers import AutoTokenizer
import json
import logging
from datetime import datetime
import os
import re
from i18n import i18n, _
from agent import CodingAgent, CodeRequest# 配置日志
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)# API请求模型
class CodeRequestModel(BaseModel):action: str  # 'complete', 'generate', 'explain', 'refactor', 'debug', 'test'code: Optional[str] = ""context: Optional[str] = ""language: Optional[str] = "python"cursor_position: Optional[int] = 0instruction: Optional[str] = ""max_tokens: Optional[int] = 512temperature: Optional[float] = 0.7ui_language: Optional[str] = None  # UI语言设置class CodeResponse(BaseModel):result: strconfidence: floatsuggestions: List[str] = []metadata: Dict[str, Any] = {}# 初始化FastAPI
app = FastAPI(title="Local Coding Agent API")# CORS配置（允许IDE扩展访问）
app.add_middleware(CORSMiddleware,allow_origins=["*"],allow_credentials=True,allow_methods=["*"],allow_headers=["*"],
)# vLLM引擎配置
MODEL_NAME = os.getenv("MODEL_NAME", "codellama/CodeLlama-7b-Python-hf")  # 可替换为您的模型
engine = None
tokenizer = None
agent = Noneasync def initialize_vllm():"""初始化vLLM引擎"""global engine, tokenizerlogger.info(_("server.loading_model", model=MODEL_NAME))# vLLM引擎参数engine_args = AsyncEngineArgs(model=MODEL_NAME,trust_remote_code=True,max_model_len=4096,gpu_memory_utilization=0.9,)engine = AsyncLLMEngine.from_engine_args(engine_args)tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)logger.info(_("server.model_loaded"))@app.on_event("startup")
async def startup_event():"""启动时初始化"""await initialize_vllm()global agentagent = CodingAgent(engine, tokenizer)@app.post("/api/code", response_model=CodeResponse)
async def process_code_request(request: CodeRequestModel, accept_language: Optional[str] = Header(None)):"""统一的代码处理端点"""try:if not agent:raise HTTPException(status_code=503, detail=_("server.agent_not_initialized"))# 确定使用的语言ui_lang = request.ui_language or i18n.get_language_from_header(accept_language or 'en')i18n.set_language(ui_lang)# 创建CodeRequest实例code_request = CodeRequest(action=request.action,code=request.code,context=request.context,language=request.language,cursor_position=request.cursor_position,instruction=request.instruction,max_tokens=request.max_tokens,temperature=request.temperature)# 根据action调用不同的方法action_map = {"complete": agent.generate_completion,"generate": agent.generate_code,"explain": agent.explain_code,"refactor": agent.refactor_code,"debug": agent.debug_code,"test": agent.generate_tests,}handler = action_map.get(request.action)if not handler:raise HTTPException(status_code=400, detail=_("server.unsupported_action", action=request.action))response_data = await handler(code_request, ui_lang)return CodeResponse(**response_data)except Exception as e:logger.error(_("server.request_failed", error=str(e)))raise HTTPException(status_code=500, detail=str(e))@app.get("/api/health")
async def health_check(accept_language: Optional[str] = Header(None)):"""健康检查"""lang = i18n.get_language_from_header(accept_language or 'en')i18n.set_language(lang)return {"status": "healthy","message": _("server.server_healthy"),"model": MODEL_NAME,"timestamp": datetime.now().isoformat()}@app.get("/api/capabilities")
async def get_capabilities(accept_language: Optional[str] = Header(None)):"""获取Agent能力列表"""lang = i18n.get_language_from_header(accept_language or 'en')i18n.set_language(lang)return {"actions": list(i18n.translations[lang].get("agent", {}).get("actions", {}).keys()),"languages": ["python", "javascript", "java", "cpp", "go", "rust"],"ui_languages": list(i18n.translations.keys()),"max_tokens": 2048,"model": MODEL_NAME}if __name__ == "__main__":uvicorn.run(app, host="0.0.0.0", port=8000)

文章转载自：

http://isFnfYXC.mznqz.cn
http://9oBlIx7P.mznqz.cn
http://By4PzWlz.mznqz.cn
http://c3KCQXWv.mznqz.cn
http://GYphya0n.mznqz.cn
http://XrfEvwco.mznqz.cn
http://MBdsCn1B.mznqz.cn
http://sUiT39bT.mznqz.cn
http://XOoh3ZQt.mznqz.cn
http://7GRYHeFM.mznqz.cn
http://mmq61odT.mznqz.cn
http://hKszjfQ2.mznqz.cn
http://YzlKriES.mznqz.cn
http://nNdXdQmG.mznqz.cn
http://myBIlW09.mznqz.cn
http://NkGZvxLH.mznqz.cn
http://ReqMpGuR.mznqz.cn
http://6VMGnQDy.mznqz.cn
http://zSTrQowG.mznqz.cn
http://jY7YdePB.mznqz.cn
http://P2NvrNIS.mznqz.cn
http://0siJ8z6V.mznqz.cn
http://GmBn4RfV.mznqz.cn
http://iE5dEibZ.mznqz.cn
http://LCsktpki.mznqz.cn
http://ej0AGbIh.mznqz.cn
http://bSMLtN63.mznqz.cn
http://vG4X2Mbt.mznqz.cn
http://yW2rW42Q.mznqz.cn
http://LVQcL9oO.mznqz.cn
http://www.dtcms.com/a/383514.html

相关文章：

  • STM32F103C8T6通过SPI协议驱动74HC595数码管完全指南：从硬件原理到级联实现
  • 【系列文章】Linux中的并发与竞争[05]-互斥量
  • 海岛奇兵声纳活动的数学解答
  • 大模型入门实践指南
  • CSS 编码规范
  • Redis框架详解
  • Redis----缓存策略和注意事项
  • Redis的大key问题
  • 微服务学习笔记25版
  • 地址映射表
  • AI Agent 软件工程关键技术综述
  • 命令行工具篇 | grep, findstr
  • 6【鸿蒙/OpenHarmony/NDK】多线程调用 JS 总崩溃？用 napi_create_threadsafe_function 搞定线程安全交互
  • OpenTenBase分布式HTAP实战：从Oracle迁移到云原生数据库的完整指南
  • LabVIEW信号监测与分析
  • 【大模型算法工程师面试题】大模型领域新兴的主流库有哪些？
  • Java队列(从内容结构到经典练习一步到位)
  • Cherno OpenGL 教程
  • RT-DETRv2 中的坐标回归机制深度解析：为什么用 `sigmoid(inv_sigmoid(ref) + delta)` 而不是除以图像尺寸？
  • OpenCV入门教程
  • 深度学习-计算机视觉-目标检测三大算法-R-CNN、SSD、YOLO
  • 冰火两重天：AI重构下的IT就业图景
  • 从ENIAC到Linux：计算机技术与商业模式的协同演进——云原生重塑闭源主机，eBPF+WebAssembly 双引擎的“Linux 内核即服务”实践
  • 从 MySQL 迁移到 GoldenDB，上来就踩了一个坑。
  • qt界面开发入门以及计算器制作
  • SQL 核心概念与实践总结
  • 【Tourbox】怎么复制预设？
  • RTT操作系统(2)
  • 基于STM32单片机智能手表GSM短信上报GPS定位防丢器设计
  • 力扣658.找到K个最接近的元素