app.py

"""
Bilingual Arabic-English Text Summarization Suite
Production Hugging Face Space (Gradio 5.x + API Endpoint).
"""

import os
import sys
import time
from typing import Optional

PROJECT_ROOT = os.path.abspath(os.path.dirname(__file__))
if PROJECT_ROOT not in sys.path:
    sys.path.insert(0, PROJECT_ROOT)

# ZeroGPU decorator support
try:
    import spaces
    has_spaces = True
except ImportError:
    has_spaces = False

def gpu_decorator(func):
    if has_spaces and hasattr(spaces, "GPU"):
        return spaces.GPU(func)
    return func

import gradio as gr
from nlp_core.language_detector import LanguageDetector
from nlp_core.tokenizer import BilingualTokenizer
from models.extractive.textrank import TextRankSummarizer
from models.extractive.lsa import LSASummarizer
from models.extractive.hybrid_scorer import HybridSummarizer
from models.abstractive.seq2seq_model import Seq2SeqSummarizer
from evaluation.metrics_manager import MetricsManager

lang_detector = LanguageDetector()
tokenizer = BilingualTokenizer()
metrics_mgr = MetricsManager()

CKPT_AR = os.path.join(PROJECT_ROOT, "checkpoints", "seq2seq_arabic.pt")
CKPT_EN = os.path.join(PROJECT_ROOT, "checkpoints", "seq2seq_english.pt")

seq_ar = Seq2SeqSummarizer.load_checkpoint(CKPT_AR, device="cpu") if os.path.exists(CKPT_AR) else None
seq_en = Seq2SeqSummarizer.load_checkpoint(CKPT_EN, device="cpu") if os.path.exists(CKPT_EN) else None


@gpu_decorator
def summarize_text(text, algorithm="Seq2Seq with Bahdanau Attention (Abstractive 20-Ep)", sentences=3, beam_width=3):
    """Core summarization engine exposed to both Gradio UI and REST API."""
    if not text or not str(text).strip():
        return "Please enter text to summarize.", "N/A", "0%", "0 ms"
    
    text = str(text).strip()
    t0 = time.time()
    detected_lang = lang_detector.detect_language(text)
    
    if "lsa" in str(algorithm).lower():
        res = LSASummarizer().summarize(text, num_sentences=int(sentences), lang=detected_lang)
        summary = res["summary"]
    elif "hybrid" in str(algorithm).lower():
        res = HybridSummarizer().summarize(text, num_sentences=int(sentences), lang=detected_lang)
        summary = res["summary"]
    elif "textrank" in str(algorithm).lower():
        res = TextRankSummarizer().summarize(text, num_sentences=int(sentences), lang=detected_lang)
        summary = res["summary"]
    else:  # Seq2Seq Abstractive
        model = seq_ar if detected_lang == "ar" else seq_en
        if model:
            toks = tokenizer.tokenize_words(text, lang=detected_lang)
            gen_toks = model.summarize_beam(toks, beam_width=int(beam_width), max_len=60)
            summary = " ".join(gen_toks)
        else:
            res = HybridSummarizer().summarize(text, num_sentences=int(sentences), lang=detected_lang)
            summary = res["summary"] + " (Used Hybrid fallback)"
            
    latency = f"{(time.time() - t0)*1000.0:.1f} ms"
    orig_w = len(text.split())
    sum_w = len(summary.split())
    reduction = f"{(1 - (sum_w/orig_w))*100:.1f}%" if orig_w > 0 else "0%"
    
    return summary, detected_lang.upper(), reduction, latency


# Build Gradio UI with named API endpoints
with gr.Blocks(title="Bilingual Arabic-English Text Summarization AI", theme=gr.themes.Soft()) as demo:
    gr.Markdown("# 📖 Bilingual Arabic-English Text Summarization AI System")
    gr.Markdown("""
    **Production NLP Platform** supporting Extractive (TextRank, LSA, Hybrid) and Abstractive (Seq2Seq + Bahdanau Attention) summarization.
    """)
    
    with gr.Row():
        with gr.Column():
            input_text = gr.Textbox(
                label="Source Document Text | النص الأصلي",
                placeholder="Paste Arabic or English article text here...",
                lines=10
            )
            algorithm = gr.Dropdown(
                label="Summarization Algorithm | خوارزمية التلخيص",
                choices=[
                    "Seq2Seq with Bahdanau Attention (Abstractive 20-Ep)",
                    "Hybrid Scorer (Extractive)",
                    "TextRank (Extractive)",
                    "LSA (Extractive)"
                ],
                value="Seq2Seq with Bahdanau Attention (Abstractive 20-Ep)"
            )
            with gr.Row():
                sentences = gr.Slider(minimum=1, maximum=10, value=3, step=1, label="Sentences (Extractive)")
                beam_width = gr.Slider(minimum=1, maximum=6, value=3, step=1, label="Beam Width (Seq2Seq)")
            
            btn = gr.Button("✨ Generate Summary | توليد الملخص", variant="primary")
            
        with gr.Column():
            output_summary = gr.Textbox(label="Generated Summary | الملخص الناتج", lines=8)
            with gr.Row():
                out_lang = gr.Textbox(label="Detected Language", lines=1)
                out_reduc = gr.Textbox(label="Word Reduction", lines=1)
                out_lat = gr.Textbox(label="Latency (ms)", lines=1)
                
    btn.click(
        fn=summarize_text,
        inputs=[input_text, algorithm, sentences, beam_width],
        outputs=[output_summary, out_lang, out_reduc, out_lat],
        api_name="summarize"
    )

# Launch Gradio app directly (required for Hugging Face Space supervisor)
demo.queue()
demo.launch(server_name="0.0.0.0", server_port=7860, ssr_mode=False)