app.py

from fastapi import FastAPI, Request
from transformers import MarianMTModel, MarianTokenizer
import torch

# import chunking service
from chunking import get_max_word_length, chunk_text

app = FastAPI()

# Map target languages to Hugging Face model IDs
MODEL_MAP = {
    "bg": "Helsinki-NLP/opus-mt-tc-big-en-bg",
    "cs": "Helsinki-NLP/opus-mt-en-cs",
    "da": "Helsinki-NLP/opus-mt-en-da",
    "de": "Helsinki-NLP/opus-mt-en-de",
    "el": "Helsinki-NLP/opus-mt-tc-big-en-el",
    "es": "Helsinki-NLP/opus-mt-tc-big-en-es",
    "et": "Helsinki-NLP/opus-mt-tc-big-en-et",
    "fi": "Helsinki-NLP/opus-mt-tc-big-en-fi",
    "fr": "Helsinki-NLP/opus-mt-en-fr",
    "hr": "facebook/mbart-large-50-many-to-many-mmt",
    "hu": "Helsinki-NLP/opus-mt-tc-big-en-hu",
    "is": "Helsinki-NLP/opus-mt-tc-big-en-gmq",
    "it": "Helsinki-NLP/opus-mt-tc-big-en-it",
    "lt": "Helsinki-NLP/opus-mt-tc-big-en-lt",
    "lv": "facebook/mbart-large-50-many-to-many-mmt",
    "mk": "Helsinki-NLP/opus-mt-en-mk",
    "nb": "facebook/mbart-large-50-many-to-many-mmt", #place holder!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
    "nl": "facebook/mbart-large-50-many-to-many-mmt",
    "no": "facebook/mbart-large-50-many-to-many-mmt", #place holder!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
    "pl": "Helsinki-NLP/opus-mt-en-sla",
    "pt": "facebook/mbart-large-50-many-to-many-mmt",
    "ro": "facebook/mbart-large-50-many-to-many-mmt",
    "sk": "Helsinki-NLP/opus-mt-en-sk",
    "sl": "alirezamsh/small100",
    "sq": "alirezamsh/small100",
    "sv": "Helsinki-NLP/opus-mt-en-sv",
    "tr": "Helsinki-NLP/opus-mt-tc-big-en-tr"
}


MODEL_CACHE = {}

# ✅ Load Hugging Face model (Helsinki or Small100)
def load_model(model_id):
    if model_id not in MODEL_CACHE:
        tokenizer = MarianTokenizer.from_pretrained(model_id)
        model = MarianMTModel.from_pretrained(model_id).to("cpu")
        MODEL_CACHE[model_id] = (tokenizer, model)
    return MODEL_CACHE[model_id]

# ✅ POST /translate
@app.post("/translate")
async def translate(request: Request):
    data = await request.json()
    text = data.get("text")
    target_lang = data.get("target_lang")

    if not text or not target_lang:
        return {"error": "Missing 'text' or 'target_lang'"}

    model_id = MODEL_MAP.get(target_lang)
    if not model_id:
        return {"error": f"No model found for target language '{target_lang}'"}

    # Facebook/mbart placeholder check
    if model_id.startswith("facebook/"):
        return {"translation": f"[{target_lang}] uses model '{model_id}', which is not supported in this Space yet."}

    try:
        # 1. figure out your safe word limit for this language
        safe_limit = get_max_word_length([target_lang])

        # 2. break the input up into chunks
        chunks = chunk_text(text, safe_limit)

        # 3. translate each chunk and collect results
        tokenizer, model = load_model(model_id)
        full_translation = []
        for chunk in chunks:
            inputs = tokenizer(chunk, return_tensors="pt", padding=True, truncation=True).to(model.device)
            outputs = model.generate(**inputs, num_beams=5, length_penalty=1.2, early_stopping=True)
            full_translation.append(tokenizer.decode(outputs[0], skip_special_tokens=True))

        # 4. re-join the translated pieces
        joined = " ".join(full_translation)
        return {"translation": joined}

    except Exception as e:
        return {"error": f"Translation failed: {str(e)}"}

# ✅ GET /languages
@app.get("/languages")
def list_languages():
    return {"supported_languages": list(MODEL_MAP.keys())}

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

# ✅ Uvicorn startup (required by Hugging Face)
import uvicorn
if __name__ == "__main__":
    uvicorn.run("app:app", host="0.0.0.0", port=7860)