app.py

import os
from fastapi import FastAPI
from pydantic import BaseModel
from transformers import pipeline
from typing import Dict, Any
import uvicorn

os.environ['HF_HOME'] = '/app/.cache'
os.environ['TRANSFORMERS_CACHE'] = '/app/.cache'
os.environ['HF_DATASETS_CACHE'] = '/app/.cache'

INTENT_MODEL = "rasouzadev/medgo-intent-classifier"
HATE_MODEL = "unitary/multilingual-toxic-xlm-roberta"

print("Loading pipelines (this may take a minute)...")

intent_pipe = pipeline("text-classification", model=INTENT_MODEL, truncation=True, max_length=512)
hate_pipe = pipeline("text-classification", model=HATE_MODEL, truncation=True, max_length=512, top_k=None)

app = FastAPI(title="MedGo - Intent & Hate Detector API")

class InputText(BaseModel):
    text: str

class PredictionResponse(BaseModel):
    intent: str
    intent_score: float
    hate_label: str | None
    hate_score: float
    note: str | None

@app.get("/")
def root():
    return {
        "message": "MedGo API - Intent & Hate Detector",
        "endpoints": {
            "predict": "/predict",
            "health": "/health",
            "docs": "/docs"
        }
    }

@app.get("/health")
def health():
    return {"status": "ok"}

@app.post("/predict", response_model=PredictionResponse)
def classify(input_data: InputText) -> Dict[str, Any]:
    text = input_data.text
    
    hate_results = hate_pipe(text)[0]  # Retorna lista de todos os scores
    
    toxicity_result = next((r for r in hate_results if r.get("label") == "toxicity"), None)
    
    if toxicity_result:
        hate_score = float(toxicity_result.get("score", 0.0))
        is_toxic = hate_score >= 0.5
        hate_label = "toxic" if is_toxic else "non-toxic"
    else:
        best_result = max(hate_results, key=lambda x: x.get("score", 0.0))
        hate_score = float(best_result.get("score", 0.0))
        hate_label = best_result.get("label", "unknown")
        is_toxic = hate_score >= 0.5 and hate_label in ["toxicity", "severe_toxicity", "obscene", "threat", "insult", "identity_attack"]
    
    if is_toxic:
        return {
            "intent": "HateSpeech",
            "intent_score": hate_score,
            "hate_label": hate_label,
            "hate_score": hate_score,
            "note": "flagged_by_hate_model"
        }
    
    intent_res = intent_pipe(text)
    intent_label = intent_res[0].get("label") if intent_res and isinstance(intent_res, list) else None
    intent_score = float(intent_res[0].get("score", 0.0)) if intent_res and isinstance(intent_res, list) else 0.0
    
    return {
        "intent": intent_label or "Unknown",
        "intent_score": intent_score,
        "hate_label": hate_label,
        "hate_score": hate_score,
        "note": None
    }

if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=7860)