Building a China-First Localization Engine: Beyond Simple Translation

graph LR
    A[Korean Source Text] --> B[Stage 1: Viral Draft]
    B --> C[Stage 2: NMPA Reasoning]
    C --> D[Stage 3: Final Polish]
    D --> E[Compliant XHS Copy]
    
    B -.->|Qwen-2.5-72B| F[Fast Generation]
    C -.->|DeepSeek-R1| G[Compliance Check]
    D -.->|Qwen-2.5-72B| H[Refinement]
    
    style C fill:#fee,stroke:#f66
    style E fill:#efe,stroke:#6f6

# backend/src/translation_agent/utils_cn.py
 
XIAOHONGSHU_PERSONA = """
You are a Xiaohongshu (小红书) beauty influencer with 500K+ followers.
Your writing style:
- Enthusiastic and authentic (use "哇", "真的", "绝绝子")
- Emoji-rich (✨, 💕, 🌟)
- Trend-aware (mention "种草", "yyds", "氛围感")
- Personal experience focus ("我用了...", "亲测有效")
"""
 
async def draft_xiaohongshu_copy(source_text: str, category: str) -> str:
    """Stage 1: Generate viral draft"""
    response = await client.chat.completions.create(
        model="Qwen/Qwen2.5-72B-Instruct",
        messages=[
            {"role": "system", "content": XIAOHONGSHU_PERSONA},
            {"role": "user", "content": f"Transform this K-Beauty product description into Xiaohongshu viral copy:\n\n{source_text}"}
        ],
        temperature=0.8,  # Higher creativity for viral content
        max_tokens=500
    )
    return response.choices[0].message.content

# backend/src/translation_agent/rules_cn.py
 
NMPA_BANNED_TERMS = {
    "medical_claims": ["治疗", "祛痘", "抗衰老", "修复", "消炎"],
    "superlatives": ["第一", "最好", "顶级", "冠军", "唯一"],
    "efficacy_promises": ["100%", "立即", "永久", "根治", "彻底"],
    "whitening": ["美白", "淡斑", "祛斑", "白皙"]
}
 
COMPLIANT_ALTERNATIVES = {
    "美白": "提亮肤色",
    "祛痘": "舒缓肌肤",
    "抗衰老": "紧致肌肤",
    "治疗": "护理",
    "第一": "备受欢迎"
}
 
async def check_nmpa_compliance(draft: str) -> dict:
    """Stage 2: NMPA reasoning with DeepSeek-R1"""
    
    # Build compliance prompt
    compliance_prompt = f"""
Analyze this Xiaohongshu beauty copy for NMPA violations:
 
{draft}
 
NMPA Banned Terms:
{json.dumps(NMPA_BANNED_TERMS, ensure_ascii=False, indent=2)}
 
Output JSON:
{{
    "violations": [
        {{"term": "违规词", "category": "类别", "location": "位置"}}
    ],
    "compliance_score": 0-100,
    "suggested_fixes": [
        {{"original": "原文", "replacement": "替换建议", "rationale": "理由"}}
    ]
}}
"""
    
    response = await client.chat.completions.create(
        model="deepseek-ai/DeepSeek-R1-Distill-Llama-70B",
        messages=[
            {"role": "system", "content": "You are an NMPA compliance expert."},
            {"role": "user", "content": compliance_prompt}
        ],
        temperature=0.1,  # Low temperature for consistent rule application
        response_format={"type": "json_object"}
    )
    
    return json.loads(response.choices[0].message.content)

{
  "violations": [
    {
      "term": "祛痘",
      "category": "medical_claims",
      "location": "第2行"
    },
    {
      "term": "100%",
      "category": "efficacy_promises",
      "location": "第4行"
    }
  ],
  "compliance_score": 65,
  "suggested_fixes": [
    {
      "original": "祛痘效果100%",
      "replacement": "舒缓肌肤，改善肤质",
      "rationale": "避免医疗声明和绝对化承诺"
    }
  ]
}

async def polish_with_compliance(draft: str, compliance_result: dict) -> str:
    """Stage 3: Apply fixes and polish"""
    
    fixes_text = "\n".join([
        f"- Replace '{fix['original']}' with '{fix['replacement']}' ({fix['rationale']})"
        for fix in compliance_result['suggested_fixes']
    ])
    
    polish_prompt = f"""
Original Xiaohongshu Draft:
{draft}
 
NMPA Compliance Fixes Required:
{fixes_text}
 
Task: Rewrite the copy by applying all fixes while preserving:
1. Xiaohongshu viral tone (enthusiastic, emoji-rich)
2. Product appeal and benefits
3. Natural flow and readability
 
Output the final polished copy.
"""
    
    response = await client.chat.completions.create(
        model="Qwen/Qwen2.5-72B-Instruct",
        messages=[
            {"role": "system", "content": XIAOHONGSHU_PERSONA},
            {"role": "user", "content": polish_prompt}
        ],
        temperature=0.7
    )
    
    return response.choices[0].message.content

# ❌ Blocking approach
draft = await draft_xiaohongshu_copy(source_text, category)  # 7s
compliance = await check_nmpa_compliance(draft)              # 10s
final = await polish_with_compliance(draft, compliance)      # 8s
# Total: 25-40 seconds (user sees nothing until complete)

# backend/main.py
 
from fastapi.responses import StreamingResponse
import asyncio
 
@app.post("/cn/translate")
async def translate_china_market(request: LocalizationRequest):
    """Stream transcreation progress via SSE"""
    
    async def event_generator():
        try:
            # Stage 1: Draft
            yield f"data: {json.dumps({'status': 'drafting'})}\n\n"
            draft = await draft_xiaohongshu_copy(request.sourceText, request.categoryId)
            yield f"data: {json.dumps({'status': 'draft_done', 'preview': draft[:100]})}\n\n"
            
            # Stage 2: Reasoning
            yield f"data: {json.dumps({'status': 'reasoning'})}\n\n"
            compliance = await check_nmpa_compliance(draft)
            
            # Stream reasoning thoughts (if available)
            if 'reasoning_steps' in compliance:
                for step in compliance['reasoning_steps']:
                    yield f"data: {json.dumps({'status': 'thought', 'thought': step})}\n\n"
            
            # Stage 3: Polish
            yield f"data: {json.dumps({'status': 'polishing'})}\n\n"
            final = await polish_with_compliance(draft, compliance)
            
            # Complete
            result = {
                'status': 'complete',
                'result': {
                    'targetText': final,
                    'headline': extract_headline(final),
                    'summary': f"Compliance Score: {compliance['compliance_score']}/100",
                    'strategyPoints': [f"Fixed: {fix['original']} → {fix['replacement']}" for fix in compliance['suggested_fixes']]
                }
            }
            yield f"data: {json.dumps(result)}\n\n"
            
        except Exception as e:
            yield f"data: {json.dumps({'status': 'error', 'message': str(e)})}\n\n"
    
    return StreamingResponse(event_generator(), media_type="text/event-stream")

// frontend/components/LocalizationToolCn.tsx
 
const handleTranscreate = async () => {
    setLoading(true);
    setStatusMessage("Initializing AI Agent...");
    
    const response = await fetch('/api/localization-cn', {
        method: 'POST',
        headers: { 'Content-Type': 'application/json' },
        body: JSON.stringify({ categoryId: category.id, sourceText }),
    });
    
    const reader = response.body.getReader();
    const decoder = new TextDecoder();
    let buffer = '';
    
    while (true) {
        const { done, value } = await reader.read();
        if (done) break;
        
        buffer += decoder.decode(value, { stream: true });
        const lines = buffer.split('\n\n');
        buffer = lines.pop() || '';
        
        for (const line of lines) {
            if (line.startsWith('data: ')) {
                const data = JSON.parse(line.replace('data: ', ''));
                
                if (data.status === 'drafting') {
                    setStatusMessage("Step 1: Creating Viral Draft (Qwen-2.5)...");
                } else if (data.status === 'reasoning') {
                    setStatusMessage("Step 2: NMPA Compliance Check (DeepSeek-R1)...");
                } else if (data.status === 'thought') {
                    setReasoningThought(prev => prev + data.thought);
                } else if (data.status === 'polishing') {
                    setStatusMessage("Step 3: Final Polish (Qwen-2.5)...");
                } else if (data.status === 'complete') {
                    setResult(data.result);
                    setLoading(false);
                }
            }
        }
    }
};

# backend/src/translation_agent/utils_cn.py
 
import os
from openai import AsyncOpenAI
 
# Environment-based API configuration
DEEPINFRA_API_KEY = os.getenv("DEEPINFRA_API_KEY")
DEEPINFRA_BASE_URL = "https://api.deepinfra.com/v1/openai"
 
# Model selection
MODEL_DRAFT = "Qwen/Qwen2.5-72B-Instruct"      # Fast, high-quality Chinese
MODEL_REASONING = "deepseek-ai/DeepSeek-R1-Distill-Llama-70B"  # Compliance reasoning
MODEL_POLISH = "Qwen/Qwen2.5-72B-Instruct"     # Refinement
 
client = AsyncOpenAI(
    api_key=DEEPINFRA_API_KEY,
    base_url=DEEPINFRA_BASE_URL
)