app.py

import os
os.environ["TRANSFORMERS_CACHE"] = "/tmp"
os.environ["HF_HOME"] = "/tmp"

from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
import json
import torch
from transformers import MT5ForConditionalGeneration, MT5Tokenizer
from sentence_transformers import SentenceTransformer, util

# Load dataset
with open("data/gpt2_ready_filtered.jsonl", "r", encoding="utf-8") as f:
    data = [json.loads(line) for line in f]
texts = [item["text"] for item in data]

# SomaliQA class
class SomaliQA:
    def __init__(self, dataset_texts):
        self.texts = dataset_texts
        self.embedder = SentenceTransformer("sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2")
        self.embeddings = self.embedder.encode(self.texts, convert_to_tensor=True)
        self.tokenizer = MT5Tokenizer.from_pretrained("nurfarah57/SQ-MT5")
        self.model = MT5ForConditionalGeneration.from_pretrained("nurfarah57/SQ-MT5")
        self.model.eval()

    def extract_qa(self, text):
        parts = text.split("\nJawaab:")
        if len(parts) == 2:
            return parts[0].replace("Su'aal:", "").strip(), parts[1].strip()
        return None, None

    def clean_text(self, text):
        return text.strip().lower().rstrip("?").replace("’", "'").replace("  ", " ")

    def generate_with_mt5(self, question):
        input_text = f"su'aal: {question}"
        inputs = self.tokenizer(input_text, return_tensors="pt", padding=True)
        with torch.no_grad():
            outputs = self.model.generate(**inputs, max_length=128)
        return self.tokenizer.decode(outputs[0], skip_special_tokens=True)

    def answer(self, user_question):
        if not user_question.strip().endswith("?"):
            user_question += "?"
        user_clean = self.clean_text(user_question)

        # Exact match
        for text in self.texts:
            su_aal, jawaab = self.extract_qa(text)
            if su_aal and user_clean == self.clean_text(su_aal):
                return {"answer": jawaab, "source": "exact"}

        # Semantic match
        user_emb = self.embedder.encode(user_clean, convert_to_tensor=True)
        hits = util.semantic_search(user_emb, self.embeddings, top_k=1)
        if hits and len(hits[0]) > 0:
            idx = hits[0][0]['corpus_id']
            su_aal, jawaab = self.extract_qa(self.texts[idx])
            return {"answer": jawaab, "source": "semantic"}

        # Fallback to generation
        return {"answer": self.generate_with_mt5(user_question), "source": "generated"}

# Init model
qa_system = SomaliQA(texts)

# FastAPI
app = FastAPI(
    title="Somali QA API",
    description="Weydii su’aal oo hel jawaab sax ah laga helay dataset ama MT5 generation.",
    version="1.0"
)

class QuestionRequest(BaseModel):
    question: str

@app.get("/")
def root():
    return {"message": "✅ Somali QA API is running!"}

@app.post("/qa")
def get_answer(req: QuestionRequest):
    if not req.question.strip():
        raise HTTPException(status_code=400, detail="Su’aal lama helin.")
    return qa_system.answer(req.question)