app.py

from fastapi import FastAPI
from pydantic import BaseModel
from transformers import pipeline
import math

app = FastAPI(title="MODERATION_API_V2")

# 初始化模型
# Model A: 多语言通用毒性检测
pipe_a = pipeline("text-classification", model="textdetox/bert-multilingual-toxicity-classifier")
# Model B: 中文专门化攻击性检测
pipe_b = pipeline("text-classification", model="thu-coai/roberta-base-cold")

class CheckRequest(BaseModel):
    text: str

@app.post("/analyze")
async def analyze(request: CheckRequest):
    text = request.text
    if not text.strip():
        return {"STATUS": "ERROR", "REASON": "EMPTY_TEXT"}

    # 推理并提取风险概率 (0.0 - 1.0)
    # 针对 textdetox: LABEL_1 为有害
    res_a = pipe_a(text)[0]
    risk_a = res_a['score'] if res_a['label'] == 'LABEL_1' else 1 - res_a['score']

    # 针对 thu-coai: LABEL_1 为攻击性
    res_b = pipe_b(text)[0]
    risk_b = res_b['score'] if res_b['label'] == 'LABEL_1' else 1 - res_b['score']

    # 综合风险评分逻辑 (加权计算)
    # 权重分配：Max_Score(70%) + Avg_Score(30%)
    combined_risk = (max(risk_a, risk_b) * 0.7) + (((risk_a + risk_b) / 2) * 0.3)

    # 计算数字风险等级 (1-5 级)
    # Level 1: [0.0-0.2] SAFE
    # Level 5: [0.8-1.0] BLOCKED
    risk_level = math.ceil(combined_risk * 5)
    risk_level = max(1, min(5, risk_level))

    # 状态映射
    if risk_level >= 4:
        status = "BLOCKED"
    elif risk_level == 3:
        status = "REVIEW"
    else:
        status = "PASSED"

    return {
        "TEXT": text,
        "STATUS": status,
        "RISK_LEVEL": risk_level,
        "CONFIDENCE_SCORE": round(combined_risk, 4),
        "RAW_DATA": {
            "GENERAL_MODEL": round(risk_a, 4),
            "SPECIALIZED_MODEL": round(risk_b, 4)
        }
    }

@app.get("/health")
async def health():
    return {"STATUS": "UP"}