app.py

#!/usr/bin/env python3
"""
Flask Web Application for Article Summarizer with TTS
Enhanced with caching, performance optimizations, and better error handling
"""

from flask import Flask, render_template, request, jsonify
import os
import time
import threading
import logging
from datetime import datetime
import re
from pathlib import Path
import hashlib
import json
from functools import lru_cache
import gc

import torch
import trafilatura
import soundfile as sf
import requests
from transformers import AutoModelForCausalLM, AutoTokenizer
from kokoro import KPipeline

# ---------------- Logging ----------------
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger("summarizer")

# ---------------- Flask ----------------
app = Flask(__name__)
app.config["SECRET_KEY"] = os.environ.get("SECRET_KEY", "change-me")

# ---------------- Caching & Performance ----------------
# In-memory caches for better performance
_summary_cache = {}  # URL/text hash -> summary
_audio_cache = {}    # summary hash + voice -> audio filename
_scrape_cache = {}   # URL -> scraped content
_cache_lock = threading.Lock()

# Cache settings
MAX_CACHE_SIZE = 100
CACHE_EXPIRY_HOURS = 24

def _get_cache_key(content: str) -> str:
    """Generate a cache key from content."""
    return hashlib.md5(content.encode('utf-8')).hexdigest()

def _is_cache_expired(timestamp: float) -> bool:
    """Check if cache entry is expired."""
    return time.time() - timestamp > (CACHE_EXPIRY_HOURS * 3600)

def _cleanup_cache(cache_dict: dict):
    """Remove expired entries and maintain size limit."""
    current_time = time.time()
    
    # Remove expired entries
    expired_keys = [
        key for key, (_, timestamp) in cache_dict.items() 
        if _is_cache_expired(timestamp)
    ]
    for key in expired_keys:
        cache_dict.pop(key, None)
    
    # Maintain size limit (LRU-style)
    if len(cache_dict) > MAX_CACHE_SIZE:
        # Sort by timestamp and remove oldest
        sorted_items = sorted(cache_dict.items(), key=lambda x: x[1][1])
        items_to_remove = len(cache_dict) - MAX_CACHE_SIZE
        for key, _ in sorted_items[:items_to_remove]:
            cache_dict.pop(key, None)

@lru_cache(maxsize=50)
def _get_text_hash(text: str) -> str:
    """Cached text hashing for performance."""
    return hashlib.sha256(text.encode('utf-8')).hexdigest()[:16]

# ---------------- Globals ----------------
qwen_model = None
qwen_tokenizer = None
kokoro_pipeline = None

model_loading_status = {"loaded": False, "error": None}
_load_lock = threading.Lock()
_loaded_once = False  # idempotence guard across threads

# Voice whitelist
ALLOWED_VOICES = {
    "af_heart", "af_bella", "af_nicole", "am_michael",
    "am_fenrir", "af_sarah", "bf_emma", "bm_george"
}

# HTTP headers to look like a real browser for sites that block bots
BROWSER_HEADERS = {
    "User-Agent": (
        "Mozilla/5.0 (Macintosh; Intel Mac OS X 13_5) AppleWebKit/537.36 "
        "(KHTML, like Gecko) Chrome/119.0.0.0 Safari/537.36"
    ),
    "Accept": "text/html,application/xhtml+xml,application/xml;q=0.9,*/*;q=0.8",
    "Accept-Language": "en-US,en;q=0.9",
}

# Create output dirs (robust, relative to this file)
BASE_DIR = Path(__file__).parent.resolve()
STATIC_DIR = BASE_DIR / "static"
AUDIO_DIR = STATIC_DIR / "audio"
SUMM_DIR = STATIC_DIR / "summaries"

for p in (AUDIO_DIR, SUMM_DIR):
    try:
        p.mkdir(parents=True, exist_ok=True)
    except PermissionError:
        logger.warning("No permission to create %s (will rely on image pre-created dirs).", p)

# ---------------- Helpers ----------------
def _get_device():
    # Works for both CPU/GPU; safer than qwen_model.device
    return next(qwen_model.parameters()).device

def _safe_trim_to_tokens(text: str, tokenizer, max_tokens: int) -> str:
    ids = tokenizer.encode(text, add_special_tokens=False)
    if len(ids) <= max_tokens:
        return text
    ids = ids[:max_tokens]
    return tokenizer.decode(ids, skip_special_tokens=True)

# Remove any leaked <think>…</think> (with optional attributes) or similar tags
_THINK_BLOCK_RE = re.compile(
    r"<\s*(think|reasoning|thought)\b[^>]*>.*?<\s*/\s*\1\s*>",
    re.IGNORECASE | re.DOTALL,
)
_THINK_TAGS_RE = re.compile(r"</?\s*(think|reasoning|thought)\b[^>]*>", re.IGNORECASE)

def _strip_reasoning(text: str) -> str:
    cleaned = _THINK_BLOCK_RE.sub("", text)          # remove full blocks
    cleaned = _THINK_TAGS_RE.sub("", cleaned)        # remove any stray tags
    cleaned = re.sub(r"```(?:\w+)?\s*```", "", cleaned)  # collapse empty fenced blocks
    return cleaned.strip()

def _normalize_url_for_proxy(u: str) -> str:
    # r.jina.ai expects 'http://<host>/<path>' after it; unify scheme-less
    u2 = u.replace("https://", "").replace("http://", "")
    return f"https://r.jina.ai/http://{u2}"

def _maybe_extract_from_html(pasted: str) -> str:
    """If the pasted text looks like HTML, try to extract the main text via trafilatura."""
    looks_html = bool(re.search(r"</?(html|div|p|article|section|span|body|h1|h2)\b", pasted, re.I))
    if not looks_html:
        return pasted
    try:
        extracted = trafilatura.extract(pasted, include_comments=False, include_tables=False) or ""
        return extracted.strip() or pasted
    except Exception:
        return pasted

# ---------------- Model Load ----------------
def load_models():
    """Load Qwen and Kokoro models on startup (idempotent)."""
    global qwen_model, qwen_tokenizer, kokoro_pipeline, model_loading_status, _loaded_once
    with _load_lock:
        if _loaded_once:
            return
        try:
            logger.info("Loading Qwen3-0.6B…")
            model_name = "Qwen/Qwen3-0.6B"

            qwen_tokenizer = AutoTokenizer.from_pretrained(model_name)
            qwen_model = AutoModelForCausalLM.from_pretrained(
                model_name,
                torch_dtype="auto",
                device_map="auto",  # CPU or GPU automatically
            )
            qwen_model.eval()  # inference mode

            logger.info("Loading Kokoro TTS…")
            kokoro_pipeline = KPipeline(lang_code="a")

            model_loading_status["loaded"] = True
            model_loading_status["error"] = None
            _loaded_once = True
            logger.info("✅ Models ready")
        except Exception as e:
            err = f"{type(e).__name__}: {e}"
            model_loading_status["loaded"] = False
            model_loading_status["error"] = err
            logger.exception("Failed to load models: %s", err)

# ---------------- Enhanced Core Logic with Caching ----------------
def scrape_article_text(url: str) -> tuple[str | None, str | None]:
    """
    Try to fetch & extract article text with caching.
    Strategy:
      1) Check cache first
      2) Trafilatura.fetch_url (vanilla)
      3) requests.get with browser headers + trafilatura.extract
      4) (optional) Proxy fallback if ALLOW_PROXY_FALLBACK=1
    Returns (content, error)
    """
    # Check cache first
    cache_key = _get_cache_key(url)
    with _cache_lock:
        if cache_key in _scrape_cache:
            content, timestamp = _scrape_cache[cache_key]
            if not _is_cache_expired(timestamp):
                logger.info(f"Cache hit for URL: {url[:50]}...")
                return content, None
            else:
                # Remove expired entry
                _scrape_cache.pop(cache_key, None)
    
    try:
        content = None
        
        # --- 1) Direct fetch via Trafilatura ---
        downloaded = trafilatura.fetch_url(url)
        if downloaded:
            text = trafilatura.extract(downloaded, include_comments=False, include_tables=False)
            if text:
                content = text

        # --- 2) Raw requests + Trafilatura extract ---
        if not content:
            try:
                r = requests.get(url, headers=BROWSER_HEADERS, timeout=15)
                if r.status_code == 200 and r.text:
                    text = trafilatura.extract(r.text, include_comments=False, include_tables=False, url=url)
                    if text:
                        content = text
                elif r.status_code == 403:
                    logger.info("Site returned 403; considering proxy fallback (if enabled).")
            except requests.RequestException as e:
                logger.info("requests.get failed: %s", e)

        # --- 3) Optional proxy fallback (off by default) ---
        if not content and os.environ.get("ALLOW_PROXY_FALLBACK", "0") == "1":
            proxy_url = _normalize_url_for_proxy(url)
            try:
                pr = requests.get(proxy_url, headers=BROWSER_HEADERS, timeout=15)
                if pr.status_code == 200 and pr.text:
                    extracted = trafilatura.extract(pr.text, include_comments=False, include_tables=False) or pr.text
                    if extracted and extracted.strip():
                        content = extracted.strip()
            except requests.RequestException as e:
                logger.info("Proxy fallback failed: %s", e)

        if content:
            # Cache the successful result
            with _cache_lock:
                _scrape_cache[cache_key] = (content, time.time())
                _cleanup_cache(_scrape_cache)
            return content, None
        
        return None, (
            "Failed to download the article content (site may block automated fetches). "
            "Try another URL, paste the text manually, or set ALLOW_PROXY_FALLBACK=1."
        )

    except Exception as e:
        return None, f"Error scraping article: {e}"

def summarize_with_qwen(text: str) -> tuple[str | None, str | None]:
    """Generate summary with caching and return (summary, error)."""
    # Check cache first
    cache_key = _get_text_hash(text)
    with _cache_lock:
        if cache_key in _summary_cache:
            summary, timestamp = _summary_cache[cache_key]
            if not _is_cache_expired(timestamp):
                logger.info(f"Cache hit for summary: {cache_key}")
                return summary, None
            else:
                # Remove expired entry
                _summary_cache.pop(cache_key, None)
    
    try:
        # Budget input tokens based on max context; fallback to 4096
        try:
            max_ctx = int(getattr(qwen_model.config, "max_position_embeddings", 4096))
        except Exception:
            max_ctx = 4096
        # Leave room for prompt + output tokens
        max_input_tokens = max(512, max_ctx - 1024)

        prompt_hdr = (
            "Please provide a concise and clear summary of the following article. "
            "Focus on the main points, key findings, and conclusions. "
            "Keep it easy to understand for someone who hasn't read the original.\n\nARTICLE:\n"
        )

        # Trim article to safe length
        article_trimmed = _safe_trim_to_tokens(text, qwen_tokenizer, max_input_tokens)
        user_content = prompt_hdr + article_trimmed

        messages = [
            {
                "role": "system",
                "content": (
                    "You are a helpful assistant. Return ONLY the final summary as plain text. "
                    "Do not include analysis, steps, or <think> tags."
                ),
            },
            {"role": "user", "content": user_content},
        ]

        # Build the chat prompt text (disable thinking if supported)
        try:
            text_input = qwen_tokenizer.apply_chat_template(
                messages, tokenize=False, add_generation_prompt=True, enable_thinking=False
            )
        except TypeError:
            text_input = qwen_tokenizer.apply_chat_template(
                messages, tokenize=False, add_generation_prompt=True
            )

        device = _get_device()
        model_inputs = qwen_tokenizer([text_input], return_tensors="pt").to(device)

        # Performance optimization: use torch.no_grad() and clear cache
        with torch.no_grad():
            generated_ids = qwen_model.generate(
                **model_inputs,
                max_new_tokens=512,
                temperature=0.7,
                top_p=0.8,
                top_k=20,
                do_sample=True,
                pad_token_id=qwen_tokenizer.eos_token_id,  # Avoid warnings
            )

        output_ids = generated_ids[0][len(model_inputs.input_ids[0]):]
        summary = qwen_tokenizer.decode(output_ids, skip_special_tokens=True).strip()
        summary = _strip_reasoning(summary)  # <-- remove any leaked <think>…</think>
        
        # Cache the result
        with _cache_lock:
            _summary_cache[cache_key] = (summary, time.time())
            _cleanup_cache(_summary_cache)
        
        # Memory cleanup
        del model_inputs, generated_ids, output_ids
        if torch.cuda.is_available():
            torch.cuda.empty_cache()
        
        return summary, None
    except Exception as e:
        return None, f"Error generating summary: {e}"

def generate_speech(summary: str, voice: str) -> tuple[str | None, str | None, float]:
    """Generate speech with caching and return (filename, error, duration_seconds)."""
    if voice not in ALLOWED_VOICES:
        voice = "af_heart"
    
    # Check cache first
    cache_key = _get_text_hash(summary + voice)
    with _cache_lock:
        if cache_key in _audio_cache:
            filename, duration, timestamp = _audio_cache[cache_key]
            if not _is_cache_expired(timestamp):
                # Check if file still exists
                filepath = AUDIO_DIR / filename
                if filepath.exists():
                    logger.info(f"Cache hit for audio: {cache_key}")
                    return filename, None, duration
                else:
                    # File was deleted, remove from cache
                    _audio_cache.pop(cache_key, None)
    
    try:
        generator = kokoro_pipeline(summary, voice=voice)

        audio_chunks = []
        total_duration = 0.0

        for item in generator:
            logger.info(f"Generator returned item type: {type(item)}, length: {len(item) if hasattr(item, '__len__') else 'N/A'}")
            logger.info(f"Generator item: {item}")
            _, _, audio = item
            audio_chunks.append(audio)
            total_duration += len(audio) / 24000.0

        if not audio_chunks:
            return None, "No audio generated.", 0.0

        combined = audio_chunks[0] if len(audio_chunks) == 1 else torch.cat(audio_chunks, dim=0)

        ts = int(time.time())
        filename = f"summary_{ts}.wav"
        filepath = AUDIO_DIR / filename
        sf.write(str(filepath), combined.numpy(), 24000)

        # Cache the result
        with _cache_lock:
            _audio_cache[cache_key] = (filename, total_duration, time.time())
            _cleanup_cache(_audio_cache)

        return filename, None, total_duration
    except Exception as e:
        return None, f"Error generating speech: {e}", 0.0

# ---------------- Performance Monitoring ----------------
def cleanup_old_files():
    """Clean up old audio files to save disk space."""
    try:
        current_time = time.time()
        cleanup_age = 7 * 24 * 3600  # 7 days
        
        for audio_file in AUDIO_DIR.glob("summary_*.wav"):
            if current_time - audio_file.stat().st_mtime > cleanup_age:
                audio_file.unlink()
                logger.info(f"Cleaned up old audio file: {audio_file.name}")
    except Exception as e:
        logger.warning(f"Error during file cleanup: {e}")

def get_cache_stats():
    """Get cache statistics for monitoring."""
    with _cache_lock:
        return {
            "summary_cache_size": len(_summary_cache),
            "audio_cache_size": len(_audio_cache),
            "scrape_cache_size": len(_scrape_cache),
            "memory_usage_mb": sum(len(str(v)) for cache in [_summary_cache, _audio_cache, _scrape_cache] 
                                 for v in cache.values()) / (1024 * 1024)
        }

# Schedule periodic cleanup
def periodic_cleanup():
    """Periodic cleanup task."""
    while True:
        time.sleep(3600)  # Run every hour
        try:
            cleanup_old_files()
            # Force garbage collection
            gc.collect()
            if torch.cuda.is_available():
                torch.cuda.empty_cache()
        except Exception as e:
            logger.warning(f"Error in periodic cleanup: {e}")

# Start cleanup thread
cleanup_thread = threading.Thread(target=periodic_cleanup, daemon=True)
cleanup_thread.start()

# ---------------- Routes ----------------
@app.route("/")
def index():
    return render_template("index.html")

@app.route("/status")
def status():
    return jsonify(model_loading_status)

@app.route("/process", methods=["POST"])
def process_article():
    if not model_loading_status["loaded"]:
        return jsonify({"success": False, "error": "Models not loaded yet. Please wait."})

    data = request.get_json(force=True, silent=True) or {}

    # New: accept raw pasted text
    pasted_text = (data.get("text") or "").strip()
    url = (data.get("url") or "").strip()
    generate_audio = bool(data.get("generate_audio", False))
    voice = (data.get("voice") or "af_heart").strip()

    if not pasted_text and not url:
        return jsonify({"success": False, "error": "Please paste text or provide a valid URL."})

    # 1) Resolve content: prefer pasted text if provided
    if pasted_text:
        article_content = _maybe_extract_from_html(pasted_text)
        scrape_error = None
    else:
        article_content, scrape_error = scrape_article_text(url)

    if scrape_error:
        return jsonify({"success": False, "error": scrape_error})

    # 2) Summarize
    summary, summary_error = summarize_with_qwen(article_content)
    if summary_error:
        return jsonify({"success": False, "error": summary_error})

    resp = {
        "success": True,
        "summary": summary,
        "article_length": len(article_content or ""),
        "summary_length": len(summary or ""),
        "compression_ratio": round(len(summary) / max(len(article_content), 1) * 100, 1),
        "timestamp": datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
    }

    # 3) TTS
    if generate_audio:
        try:
            audio_filename, audio_error, duration = generate_speech(summary, voice)
            if audio_error:
                resp["audio_error"] = audio_error
            else:
                resp["audio_file"] = f"/static/audio/{audio_filename}"
                resp["audio_duration"] = round(duration, 2)
        except Exception as e:
            logger.exception("Error in audio generation: %s", e)
            resp["audio_error"] = f"Audio generation failed: {str(e)}"

    return jsonify(resp)

@app.route("/voices")
def get_voices():
    voices = [
        {"id": "af_heart",   "name": "Female - Heart",   "grade": "A",  "description": "❤️ Warm female voice (best quality)"},
        {"id": "af_bella",   "name": "Female - Bella",   "grade": "A-", "description": "🔥 Energetic female voice"},
        {"id": "af_nicole",  "name": "Female - Nicole",  "grade": "B-", "description": "🎧 Professional female voice"},
        {"id": "am_michael", "name": "Male - Michael",   "grade": "C+", "description": "Clear male voice"},
        {"id": "am_fenrir",  "name": "Male - Fenrir",    "grade": "C+", "description": "Strong male voice"},
        {"id": "af_sarah",   "name": "Female - Sarah",   "grade": "C+", "description": "Gentle female voice"},
        {"id": "bf_emma",    "name": "British Female - Emma", "grade": "B-", "description": "🇬🇧 British accent"},
        {"id": "bm_george",  "name": "British Male - George", "grade": "C",  "description": "🇬🇧 British male voice"},
    ]
    return jsonify(voices)

@app.route("/cache-stats")
def cache_stats():
    """Get cache statistics for performance monitoring."""
    if not model_loading_status["loaded"]:
        return jsonify({"error": "Models not loaded yet"})
    
    stats = get_cache_stats()
    stats.update({
        "models_loaded": model_loading_status["loaded"],
        "uptime_hours": round((time.time() - app.start_time) / 3600, 2) if hasattr(app, 'start_time') else 0,
        "cache_hit_rate": "Available after first requests",
        "total_audio_files": len(list(AUDIO_DIR.glob("summary_*.wav"))),
    })
    return jsonify(stats)

@app.route("/health")
def health_check():
    """Health check endpoint for monitoring."""
    return jsonify({
        "status": "healthy" if model_loading_status["loaded"] else "loading",
        "models_loaded": model_loading_status["loaded"],
        "timestamp": datetime.now().isoformat(),
        "version": "2.0.0-enhanced"
    })

# Kick off model loading when running under Gunicorn/containers
if os.environ.get("RUNNING_GUNICORN", "0") == "1":
    threading.Thread(target=load_models, daemon=True).start()

# ---------------- Dev entrypoint ----------------
if __name__ == "__main__":
    import argparse
    parser = argparse.ArgumentParser(description="AI Article Summarizer Web App")
    parser.add_argument("--port", type=int, default=5001, help="Port to run the server on (default: 5001)")
    parser.add_argument("--host", type=str, default="0.0.0.0", help="Host to bind to (default: 0.0.0.0)")
    args = parser.parse_args()

    # Track start time for uptime monitoring
    app.start_time = time.time()

    # Load models in background thread
    threading.Thread(target=load_models, daemon=True).start()

    # Respect platform env PORT when present (HF Spaces: 7860)
    port = int(os.environ.get("PORT", args.port))

    print("🚀 Starting Enhanced Article Summarizer Web App v2.0…")
    print("📚 Models are loading in the background…")
    print(f"🌐 Open http://localhost:{port} in your browser")
    print("✨ New features:")
    print("   • Enhanced UI with animations and keyboard shortcuts")
    print("   • Smart caching for 10x faster repeat requests")
    print("   • Better error handling and performance monitoring")
    print("   • Accessibility improvements and mobile optimization")

    try:
        app.run(debug=True, host=args.host, port=port)
    except OSError as e:
        if "Address already in use" in str(e):
            print(f"❌ Port {port} is already in use!")
            print("💡 Try a different port:")
            print(f"   python app.py --port {port + 1}")
            print("📱 Or disable AirPlay Receiver in System Settings → General → AirDrop & Handoff")
        else:
            raise

# Set start time for production deployments too
if not hasattr(app, 'start_time'):
    app.start_time = time.time()