app.py

from fastapi import FastAPI
from pydantic import BaseModel
from transformers import T5Tokenizer, T5ForConditionalGeneration
import torch
import uvicorn

app = FastAPI()

MODEL_NAME = "obx0x3/empathy-dementia"
tokenizer = T5Tokenizer.from_pretrained(MODEL_NAME)
model = T5ForConditionalGeneration.from_pretrained(MODEL_NAME)

class PromptRequest(BaseModel):
    message: str
    lang: str = None

def detect_language(text: str):
    """Simple French/English detection based on keywords."""
    fr_keywords = ["je", "tu", "c’est", "j’ai", "où", "suis", "pas", "peux"]
    return "fr" if any(word in text.lower() for word in fr_keywords) else "en"

def prefix_message(message: str, lang: str) -> str:
    """Add prefix to help model route context correctly."""
    if lang == "fr":
        return f"émotion: {message}"
    elif any(q in message.lower() for q in ["why", "how", "what", "when", "where", "?"]):
        return f"chat: {message}"
    elif any(e in message.lower() for e in ["feel", "i’m", "i am", "sad", "scared", "lonely", "happy", "forgot"]):
        return f"emotion: {message}"
    else:
        return f"chat: {message}"

@app.post("/generate")
async def generate_response(payload: PromptRequest):
    lang = payload.lang or detect_language(payload.message)
    input_text = prefix_message(payload.message, lang)

    inputs = tokenizer.encode(input_text, return_tensors="pt")
    with torch.no_grad():
        outputs = model.generate(inputs, max_length=128, num_beams=4, early_stopping=True)

    result = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return {"reply": result.strip(), "language": lang}

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