app.py

import gradio as gr
import json
import torch
from transformers import GPT2LMHeadModel, GPT2Tokenizer
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 = GPT2Tokenizer.from_pretrained("zakihassan04/gpt2-finetuned-somali")
        self.model = GPT2LMHeadModel.from_pretrained("zakihassan04/gpt2-finetuned-somali")
        self.tokenizer.pad_token = self.tokenizer.eos_token

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

        user_clean = self.clean_text(user_question)

        # Step 1: 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 jawaab  # ✅ Return exact dataset answer

        # Step 2: 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 jawaab  # ✅ Return semantically matched answer

        return "Ma helin jawaab ku habboon su’aashaada."

# Init model
qa_system = SomaliQA(texts)

# Gradio UI
def qa_interface(question):
    return qa_system.answer(question)

# Gradio interface
gr.Interface(
    fn=qa_interface,
    inputs="text",
    outputs="text",
    title="Somali GPT-2 QA System (Dataset-based)",
    description="Weydii su’aal ku saabsan beeraha — waxaad helaysaa jawaab sax ah oo laga soo qaaday dataset-kaaga.",
    theme="compact"
).launch()