app.py

from fastapi import FastAPI, Form
from fastapi.responses import Response
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
from twilio.twiml.messaging_response import MessagingResponse
import os

# -----------------------------
# Environment-safe cache path
# -----------------------------
HF_CACHE_DIR = os.getenv("HF_HOME", "/tmp/hf_cache")

# -----------------------------
# Load regression model from Hugging Face
# -----------------------------
model_id = "ST-THOMAS-OF-AQUINAS/impersonation-bart"

tokenizer = AutoTokenizer.from_pretrained(model_id, cache_dir=HF_CACHE_DIR)
model = AutoModelForSequenceClassification.from_pretrained(model_id, cache_dir=HF_CACHE_DIR)
model.eval()

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)

# -----------------------------
# Helper function
# -----------------------------
def predict_score(text: str, debug: bool = False):
    inputs = tokenizer(
        text,
        return_tensors="pt",
        truncation=True,
        padding=True,
        max_length=256
    )
    inputs = {k: v.to(device) for k, v in inputs.items()}

    with torch.no_grad():
        outputs = model(**inputs)
        raw_logits = outputs.logits.squeeze()

        # Apply sigmoid to convert raw logits into probability [0,1]
        score = torch.sigmoid(raw_logits).item()

    if debug:
        return {
            "raw_logits": raw_logits.item(),
            "sigmoid_score": score
        }

    return round(score, 3)

# -----------------------------
# FastAPI app
# -----------------------------
app = FastAPI(title="Impersonation Detector st thomas of aquinas")

# Health-check route
@app.get("/")
async def health_check():
    return {"status": "✅ API is running"}

# Simple GET test
@app.get("/predict")
async def get_predict(text: str):
    score = predict_score(text)
    return {"impersonation_score": score}

# Debugging route
@app.get("/debug")
async def debug_predict(text: str):
    debug_data = predict_score(text, debug=True)
    return {"debug_output": debug_data}

# -----------------------------
# Twilio WhatsApp POST
# -----------------------------
@app.post("/whatsapp")
async def whatsapp_reply(Body: str = Form(...)):
    resp = MessagingResponse()

    if Body.strip():
        score = predict_score(Body)
        reply = f"Impersonation Score: {score}\n(0.0 = genuine, 1.0 = impersonation)"
    else:
        reply = "⚠️ No text detected."

    resp.message(reply)
    
    # Return proper TwiML XML
    return Response(content=str(resp), media_type="application/xml")