app.py

import spaces
import gradio as gr
import logging
import os
import tempfile
import pandas as pd
import requests
from bs4 import BeautifulSoup
import torch
import whisper
import subprocess
from pydub import AudioSegment
import fitz  # PyMuPDF
import docx
import yt_dlp
from functools import lru_cache
import gc
import time
from huggingface_hub import login
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline
import traceback # For detailed error logging

# Configure logging
logging.basicConfig(
    level=logging.INFO,
    format='%(asctime)s - %(levelname)s - %(message)s'
)
logger = logging.getLogger(__name__)

# Login to Hugging Face Hub if token is available
HUGGINGFACE_TOKEN = os.environ.get('HUGGINGFACE_TOKEN')
if HUGGINGFACE_TOKEN:
    try:
        login(token=HUGGINGFACE_TOKEN)
        logger.info("Successfully logged in to Hugging Face Hub.")
    except Exception as e:
        logger.error(f"Failed to login to Hugging Face Hub: {e}")

class ModelManager:
    _instance = None

    def __new__(cls):
        if cls._instance is None:
            cls._instance = super(ModelManager, cls).__new__(cls)
            cls._instance._initialized = False
        return cls._instance

    def __init__(self):
        if not self._initialized:
            self.tokenizer = None
            self.model = None
            self.text_pipeline = None # Renamed for clarity
            self.whisper_model = None
            self._initialized = True
            self.last_used = time.time()
            self.llm_loading = False
            self.whisper_loading = False

    @spaces.GPU(duration=120) # Increased duration for potentially long loads
    def initialize_llm(self):
        """Initialize LLM model with standard transformers"""
        if self.llm_loading:
            logger.info("LLM initialization already in progress.")
            return True # Assume it will succeed or fail elsewhere
        if self.tokenizer and self.model and self.text_pipeline:
            logger.info("LLM already initialized.")
            self.last_used = time.time()
            return True

        self.llm_loading = True
        try:
            # Use small model for ZeroGPU compatibility
            MODEL_NAME = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"

            logger.info("Loading LLM tokenizer...")
            self.tokenizer = AutoTokenizer.from_pretrained(
                MODEL_NAME,
                token=HUGGINGFACE_TOKEN,
                use_fast=True
            )

            if self.tokenizer.pad_token is None:
                self.tokenizer.pad_token = self.tokenizer.eos_token

            # Basic memory settings for ZeroGPU
            logger.info("Loading LLM model...")
            self.model = AutoModelForCausalLM.from_pretrained(
                MODEL_NAME,
                token=HUGGINGFACE_TOKEN,
                device_map="auto",
                torch_dtype=torch.float16,
                low_cpu_mem_usage=True,
                # Optimizations for ZeroGPU
                # max_memory={0: "4GB"}, # Removed for better auto handling initially
                offload_folder="offload",
                offload_state_dict=True
            )

            # Create text generation pipeline
            logger.info("Creating LLM text generation pipeline...")
            self.text_pipeline = pipeline(
                "text-generation",
                model=self.model,
                tokenizer=self.tokenizer,
                torch_dtype=torch.float16,
                device_map="auto",
                max_length=1024 # Default max length
            )

            logger.info("LLM initialized successfully")
            self.last_used = time.time()
            self.llm_loading = False
            return True

        except Exception as e:
            logger.error(f"Error initializing LLM: {str(e)}")
            logger.error(traceback.format_exc()) # Log full traceback
            # Reset partially loaded components
            self.tokenizer = None
            self.model = None
            self.text_pipeline = None
            if torch.cuda.is_available():
                torch.cuda.empty_cache()
            gc.collect()
            self.llm_loading = False
            raise # Re-raise the exception to signal failure

    @spaces.GPU(duration=120) # Increased duration
    def initialize_whisper(self):
        """Initialize Whisper model for audio transcription"""
        if self.whisper_loading:
            logger.info("Whisper initialization already in progress.")
            return True
        if self.whisper_model:
             logger.info("Whisper already initialized.")
             self.last_used = time.time()
             return True

        self.whisper_loading = True
        try:
            logger.info("Loading Whisper model...")
            # Using tiny model for efficiency but can be changed based on needs
            # Specify weights_only=True to address the FutureWarning
            # Note: Whisper's load_model might not directly support weights_only yet.
            # If it errors, remove the weights_only=True. The warning is mainly informative.
            # Let's attempt without weights_only first as whisper might handle it internally
            self.whisper_model = whisper.load_model(
                "tiny", # Consider "base" for better accuracy if "tiny" struggles
                device="cuda" if torch.cuda.is_available() else "cpu",
                download_root="/tmp/whisper" # Use persistent storage if available/needed
            )
            logger.info("Whisper model initialized successfully")
            self.last_used = time.time()
            self.whisper_loading = False
            return True
        except Exception as e:
            logger.error(f"Error initializing Whisper: {str(e)}")
            logger.error(traceback.format_exc())
            self.whisper_model = None
            if torch.cuda.is_available():
                torch.cuda.empty_cache()
            gc.collect()
            self.whisper_loading = False
            raise

    def check_llm_initialized(self):
        """Check if LLM is initialized and initialize if needed"""
        if self.tokenizer is None or self.model is None or self.text_pipeline is None:
            logger.info("LLM not initialized, initializing...")
            if not self.llm_loading: # Prevent re-entry if already loading
                 self.initialize_llm()
            else:
                 logger.info("LLM initialization is already in progress by another request.")
                 # Optional: Wait a bit for the other process to finish
                 time.sleep(5)
                 if self.tokenizer is None or self.model is None or self.text_pipeline is None:
                     raise RuntimeError("LLM initialization timed out or failed.")
        self.last_used = time.time()

    def check_whisper_initialized(self):
        """Check if Whisper model is initialized and initialize if needed"""
        if self.whisper_model is None:
            logger.info("Whisper model not initialized, initializing...")
            if not self.whisper_loading: # Prevent re-entry
                self.initialize_whisper()
            else:
                logger.info("Whisper initialization is already in progress by another request.")
                time.sleep(5)
                if self.whisper_model is None:
                    raise RuntimeError("Whisper initialization timed out or failed.")
        self.last_used = time.time()

    def reset_models(self, force=False):
        """Reset models to free memory if they haven't been used recently"""
        current_time = time.time()
        # Only reset if forced or models haven't been used for 10 minutes (600 seconds)
        if force or (current_time - self.last_used > 600):
            try:
                logger.info("Resetting models to free memory...")

                # Check and delete attributes safely
                if hasattr(self, 'model') and self.model is not None:
                    del self.model
                    self.model = None
                    logger.info("LLM model deleted.")

                if hasattr(self, 'tokenizer') and self.tokenizer is not None:
                    del self.tokenizer
                    self.tokenizer = None
                    logger.info("LLM tokenizer deleted.")

                if hasattr(self, 'text_pipeline') and self.text_pipeline is not None:
                    del self.text_pipeline
                    self.text_pipeline = None
                    logger.info("LLM pipeline deleted.")

                if hasattr(self, 'whisper_model') and self.whisper_model is not None:
                    del self.whisper_model
                    self.whisper_model = None
                    logger.info("Whisper model deleted.")

                # Explicitly clear CUDA cache and collect garbage
                if torch.cuda.is_available():
                    torch.cuda.empty_cache()
                    # torch.cuda.synchronize() # May not be needed and can slow down
                    logger.info("CUDA cache cleared.")

                gc.collect()
                logger.info("Garbage collected. Models reset successfully.")
                self._initialized = False # Mark as uninitialized so they reload on next use

            except Exception as e:
                logger.error(f"Error resetting models: {str(e)}")
                logger.error(traceback.format_exc())

# Create global model manager instance
model_manager = ModelManager()

@lru_cache(maxsize=16) # Reduced cache size slightly
def download_social_media_video(url):
    """Download audio from a social media video URL."""
    temp_dir = tempfile.mkdtemp()
    output_template = os.path.join(temp_dir, '%(id)s.%(ext)s')

    ydl_opts = {
        'format': 'bestaudio/best',
        'postprocessors': [{
            'key': 'FFmpegExtractAudio',
            'preferredcodec': 'mp3',
            'preferredquality': '192', # Standard quality
        }],
        'outtmpl': output_template,
        'quiet': True,
        'no_warnings': True,
        'nocheckcertificate': True, # Sometimes needed for tricky sites
        'retries': 3, # Add retries
        'socket_timeout': 15, # Timeout
    }
    try:
        logger.info(f"Attempting to download audio from: {url}")
        with yt_dlp.YoutubeDL(ydl_opts) as ydl:
            info_dict = ydl.extract_info(url, download=True)
            # Construct the expected final filename after postprocessing
            audio_file = os.path.join(temp_dir, f"{info_dict['id']}.mp3")
            if not os.path.exists(audio_file):
                 # Fallback if filename doesn't match exactly (e.g., webm -> mp3)
                 found_files = [f for f in os.listdir(temp_dir) if f.endswith('.mp3')]
                 if found_files:
                     audio_file = os.path.join(temp_dir, found_files[0])
                 else:
                     raise FileNotFoundError(f"Could not find downloaded MP3 in {temp_dir}")

        logger.info(f"Audio downloaded successfully: {audio_file}")
        # Read the file content to return, as the temp dir might be cleaned up
        with open(audio_file, 'rb') as f:
            audio_content = f.read()

        # Clean up the temporary directory and file
        try:
            os.remove(audio_file)
            os.rmdir(temp_dir)
        except OSError as e:
            logger.warning(f"Could not completely clean up temp download files: {e}")

        # Save the content to a new temporary file that Gradio can handle
        with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as temp_output_file:
            temp_output_file.write(audio_content)
            final_path = temp_output_file.name
        logger.info(f"Audio saved to temporary file: {final_path}")
        return final_path

    except yt_dlp.utils.DownloadError as e:
        logger.error(f"yt-dlp download error for {url}: {str(e)}")
        # Clean up temp dir on error
        try:
            if os.path.exists(temp_dir):
                import shutil
                shutil.rmtree(temp_dir)
        except Exception as cleanup_e:
            logger.warning(f"Error during cleanup after download failure: {cleanup_e}")
        return None # Return None to indicate failure
    except Exception as e:
        logger.error(f"Unexpected error downloading video from {url}: {str(e)}")
        logger.error(traceback.format_exc())
        # Clean up temp dir on error
        try:
             if os.path.exists(temp_dir):
                import shutil
                shutil.rmtree(temp_dir)
        except Exception as cleanup_e:
            logger.warning(f"Error during cleanup after download failure: {cleanup_e}")
        return None # Return None

def convert_video_to_audio(video_file_path):
    """Convert a video file to audio using ffmpeg directly."""
    try:
        # Create a temporary file path for the output MP3
        with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as temp_file:
            output_file_path = temp_file.name

        logger.info(f"Converting video '{video_file_path}' to audio '{output_file_path}'")

        # Use ffmpeg directly via subprocess
        command = [
            "ffmpeg",
            "-i", video_file_path,
            "-vn",          # No video
            "-acodec", "libmp3lame", # Specify MP3 codec
            "-ab", "192k",   # Audio bitrate
            "-ar", "44100",  # Audio sample rate
            "-ac", "2",      # Stereo audio
            output_file_path,
            "-y",           # Overwrite output file if it exists
            "-loglevel", "error" # Suppress verbose ffmpeg output
        ]

        process = subprocess.run(command, check=True, capture_output=True, text=True)
        logger.info(f"ffmpeg conversion successful for {video_file_path}.")
        logger.debug(f"ffmpeg stdout: {process.stdout}")
        logger.debug(f"ffmpeg stderr: {process.stderr}")


        # Verify output file exists and has size
        if not os.path.exists(output_file_path) or os.path.getsize(output_file_path) == 0:
            raise RuntimeError(f"ffmpeg conversion failed: Output file '{output_file_path}' not created or is empty.")

        logger.info(f"Video converted to audio: {output_file_path}")
        return output_file_path
    except subprocess.CalledProcessError as e:
         logger.error(f"ffmpeg command failed with exit code {e.returncode}")
         logger.error(f"ffmpeg stderr: {e.stderr}")
         logger.error(f"ffmpeg stdout: {e.stdout}")
         # Clean up potentially empty output file
         if os.path.exists(output_file_path):
             os.remove(output_file_path)
         raise RuntimeError(f"ffmpeg conversion failed: {e.stderr}") from e
    except Exception as e:
        logger.error(f"Error converting video '{video_file_path}': {str(e)}")
        logger.error(traceback.format_exc())
        # Clean up potentially created output file
        if 'output_file_path' in locals() and os.path.exists(output_file_path):
            os.remove(output_file_path)
        raise # Re-raise the exception

def preprocess_audio(input_audio_path):
    """Preprocess the audio file (e.g., normalize volume)."""
    try:
        logger.info(f"Preprocessing audio file: {input_audio_path}")
        audio = AudioSegment.from_file(input_audio_path)

        # Apply normalization (optional, adjust target dBFS as needed)
        # Target loudness: -20 dBFS. Adjust gain based on current loudness.
        # change_in_dBFS = -20.0 - audio.dBFS
        # audio = audio.apply_gain(change_in_dBFS)

        # Export to a new temporary file
        with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as temp_file:
            output_path = temp_file.name
            audio.export(output_path, format="mp3")

        logger.info(f"Audio preprocessed and saved to: {output_path}")
        return output_path
    except Exception as e:
        logger.error(f"Error preprocessing audio '{input_audio_path}': {str(e)}")
        logger.error(traceback.format_exc())
        # Return original path if preprocessing fails? Or raise error?
        # Let's raise the error to signal failure clearly.
        raise

@spaces.GPU(duration=300) # Allow more time for transcription
def transcribe_audio_or_video(file_input):
    """Transcribe an audio or video file (local path or Gradio File object)."""
    audio_file_to_transcribe = None
    original_input_path = None
    temp_files_to_clean = []

    try:
        model_manager.check_whisper_initialized()

        if file_input is None:
            logger.info("No file input provided for transcription.")
            return "" # Return empty string for None input

        # Determine input type and get file path
        if isinstance(file_input, str): # Input is a path
            original_input_path = file_input
            logger.info(f"Processing path input: {original_input_path}")
            if not os.path.exists(original_input_path):
                 logger.error(f"Input file path does not exist: {original_input_path}")
                 raise FileNotFoundError(f"Input file not found: {original_input_path}")
            input_path = original_input_path
        elif hasattr(file_input, 'name'): # Input is a Gradio File object
            original_input_path = file_input.name
            logger.info(f"Processing Gradio file input: {original_input_path}")
            input_path = original_input_path # Gradio usually provides a temp path
        else:
            logger.error(f"Unsupported input type for transcription: {type(file_input)}")
            raise TypeError("Invalid input type for transcription. Expected file path or Gradio File object.")

        file_extension = os.path.splitext(input_path)[1].lower()

        # Check if it's a video file that needs conversion
        if file_extension in ['.mp4', '.avi', '.mov', '.mkv', '.webm']:
            logger.info(f"Detected video file ({file_extension}), converting to audio...")
            converted_audio_path = convert_video_to_audio(input_path)
            temp_files_to_clean.append(converted_audio_path)
            audio_file_to_process = converted_audio_path
        elif file_extension in ['.mp3', '.wav', '.ogg', '.flac', '.m4a']:
             logger.info(f"Detected audio file ({file_extension}).")
             audio_file_to_process = input_path
        else:
            logger.error(f"Unsupported file extension for transcription: {file_extension}")
            raise ValueError(f"Unsupported file type: {file_extension}")

        # Preprocess the audio (optional, could be skipped if causing issues)
        try:
            preprocessed_audio_path = preprocess_audio(audio_file_to_process)
            # If preprocessing creates a new file different from the input, add it to cleanup
            if preprocessed_audio_path != audio_file_to_process:
                 temp_files_to_clean.append(preprocessed_audio_path)
            audio_file_to_transcribe = preprocessed_audio_path
        except Exception as preprocess_err:
            logger.warning(f"Audio preprocessing failed: {preprocess_err}. Using original/converted audio.")
            audio_file_to_transcribe = audio_file_to_process # Fallback

        logger.info(f"Transcribing audio file: {audio_file_to_transcribe}")
        if not os.path.exists(audio_file_to_transcribe):
            raise FileNotFoundError(f"Audio file to transcribe not found: {audio_file_to_transcribe}")

        # Perform transcription
        with torch.inference_mode(): # Ensure inference mode for efficiency
            # Use fp16 if available on CUDA
            use_fp16 = torch.cuda.is_available()
            result = model_manager.whisper_model.transcribe(
                audio_file_to_transcribe,
                fp16=use_fp16
            )
            if not result:
                raise RuntimeError("Transcription failed to produce results")

        transcription = result.get("text", "Error: Transcription result empty")
        # Limit transcription length shown in logs
        log_transcription = (transcription[:100] + '...') if len(transcription) > 100 else transcription
        logger.info(f"Transcription completed: {log_transcription}")

        return transcription

    except FileNotFoundError as e:
        logger.error(f"File not found error during transcription: {e}")
        return f"Error: Input file not found ({e})"
    except ValueError as e:
         logger.error(f"Value error during transcription: {e}")
         return f"Error: Unsupported file type ({e})"
    except TypeError as e:
         logger.error(f"Type error during transcription setup: {e}")
         return f"Error: Invalid input provided ({e})"
    except RuntimeError as e:
         logger.error(f"Runtime error during transcription: {e}")
         logger.error(traceback.format_exc())
         return f"Error during processing: {e}"
    except Exception as e:
        logger.error(f"Unexpected error during transcription: {str(e)}")
        logger.error(traceback.format_exc())
        return f"Error processing the file: An unexpected error occurred."

    finally:
        # Clean up all temporary files created during the process
        for temp_file in temp_files_to_clean:
            try:
                if os.path.exists(temp_file):
                    os.remove(temp_file)
                    logger.info(f"Cleaned up temporary file: {temp_file}")
            except Exception as e:
                logger.warning(f"Could not remove temporary file {temp_file}: {str(e)}")
        # Optionally reset models if idle (might be too aggressive here)
        # model_manager.reset_models()

@lru_cache(maxsize=16)
def read_document(document_path):
    """Read the content of a document (PDF, DOCX, XLSX, CSV)."""
    try:
        logger.info(f"Reading document: {document_path}")
        if not os.path.exists(document_path):
            raise FileNotFoundError(f"Document not found: {document_path}")

        file_extension = os.path.splitext(document_path)[1].lower()

        if file_extension == ".pdf":
            doc = fitz.open(document_path)
            text = "\n".join([page.get_text() for page in doc])
            doc.close()
            return text
        elif file_extension == ".docx":
            doc = docx.Document(document_path)
            return "\n".join([paragraph.text for paragraph in doc.paragraphs])
        elif file_extension in (".xlsx", ".xls"):
            # Read all sheets and combine
            xls = pd.ExcelFile(document_path)
            text = ""
            for sheet_name in xls.sheet_names:
                df = pd.read_excel(xls, sheet_name=sheet_name)
                text += f"--- Sheet: {sheet_name} ---\n{df.to_string()}\n\n"
            return text.strip()
        elif file_extension == ".csv":
            # Try detecting separator
            try:
                df = pd.read_csv(document_path)
            except pd.errors.ParserError:
                 logger.warning(f"Could not parse CSV {document_path} with default comma separator, trying semicolon.")
                 df = pd.read_csv(document_path, sep=';')
            return df.to_string()
        else:
            logger.warning(f"Unsupported document type: {file_extension}")
            return "Unsupported file type. Please upload a PDF, DOCX, XLSX or CSV document."
    except FileNotFoundError as e:
        logger.error(f"Error reading document: {e}")
        return f"Error: Document file not found at {document_path}"
    except Exception as e:
        logger.error(f"Error reading document {document_path}: {str(e)}")
        logger.error(traceback.format_exc())
        return f"Error reading document: {str(e)}"

@lru_cache(maxsize=16)
def read_url(url):
    """Read the main textual content of a URL."""
    if not url or not url.strip().startswith('http'):
        logger.info(f"Invalid or empty URL provided: '{url}'")
        return "" # Return empty for invalid or empty URLs

    try:
        logger.info(f"Reading URL: {url}")
        headers = {
            'User-Agent': 'Mozilla/5.0 (Windows NT 10.0; Win64; x64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/91.0.4472.124 Safari/537.36'
        }
        # Increased timeout
        response = requests.get(url, headers=headers, timeout=20, allow_redirects=True)
        response.raise_for_status() # Raise HTTPError for bad responses (4xx or 5xx)

        # Check content type - proceed only if likely HTML/text
        content_type = response.headers.get('content-type', '').lower()
        if not ('html' in content_type or 'text' in content_type):
             logger.warning(f"URL {url} has non-text content type: {content_type}. Skipping.")
             return f"Error: URL content type ({content_type}) is not text/html."

        soup = BeautifulSoup(response.content, 'html.parser')

        # Remove non-content elements like scripts, styles, nav, footers etc.
        for element in soup(["script", "style", "meta", "noscript", "iframe", "header", "footer", "nav", "aside", "form", "button"]):
            element.extract()

        # Attempt to find main content area (common tags/attributes)
        main_content = (
            soup.find("main") or
            soup.find("article") or
            soup.find("div", class_=["content", "main", "post-content", "entry-content", "article-body"]) or
            soup.find("div", id=["content", "main", "article"])
        )

        if main_content:
            text = main_content.get_text(separator='\n', strip=True)
        else:
            # Fallback to body if no specific main content found
            body = soup.find("body")
            if body:
                text = body.get_text(separator='\n', strip=True)
            else: # Very basic fallback
                 text = soup.get_text(separator='\n', strip=True)

        # Clean up whitespace: replace multiple newlines/spaces with single ones
        text = '\n'.join([line.strip() for line in text.split('\n') if line.strip()])
        text = ' '.join(text.split()) # Consolidate spaces within lines

        if not text:
            logger.warning(f"Could not extract meaningful text from URL: {url}")
            return "Error: Could not extract text content from URL."

        # Limit content size to avoid overwhelming the LLM
        max_chars = 15000
        if len(text) > max_chars:
            logger.info(f"URL content truncated to {max_chars} characters.")
            text = text[:max_chars] + "... [content truncated]"

        return text
    except requests.exceptions.RequestException as e:
        logger.error(f"Error fetching URL {url}: {str(e)}")
        return f"Error reading URL: Could not fetch content ({e})"
    except Exception as e:
        logger.error(f"Error parsing URL {url}: {str(e)}")
        logger.error(traceback.format_exc())
        return f"Error reading URL: Could not parse content ({e})"

def process_social_media_url(url):
    """Process a social media URL, attempting to get text and transcribe video/audio."""
    if not url or not url.strip().startswith('http'):
        logger.info(f"Invalid or empty social media URL: '{url}'")
        return None

    logger.info(f"Processing social media URL: {url}")
    text_content = None
    video_transcription = None
    error_occurred = False

    # 1. Try extracting text content using read_url (might work for some platforms/posts)
    try:
        text_content = read_url(url)
        if text_content and text_content.startswith("Error:"):
             logger.warning(f"Failed to read text content from social URL {url}: {text_content}")
             text_content = None # Reset if it was an error message
    except Exception as e:
        logger.error(f"Error reading text content from social URL {url}: {e}")
        error_occurred = True

    # 2. Try downloading and transcribing potential video/audio content
    downloaded_audio_path = None
    try:
        downloaded_audio_path = download_social_media_video(url)
        if downloaded_audio_path:
            logger.info(f"Audio downloaded from {url}, proceeding to transcription.")
            video_transcription = transcribe_audio_or_video(downloaded_audio_path)
            if video_transcription and video_transcription.startswith("Error"):
                logger.warning(f"Transcription failed for audio from {url}: {video_transcription}")
                video_transcription = None # Reset if it was an error
        else:
             logger.info(f"No downloadable audio/video found or download failed for URL: {url}")
    except Exception as e:
        logger.error(f"Error processing video content from social URL {url}: {e}")
        logger.error(traceback.format_exc())
        error_occurred = True
    finally:
         # Clean up downloaded file if it exists
        if downloaded_audio_path and os.path.exists(downloaded_audio_path):
            try:
                os.remove(downloaded_audio_path)
                logger.info(f"Cleaned up downloaded audio: {downloaded_audio_path}")
            except Exception as e:
                logger.warning(f"Failed to cleanup downloaded audio {downloaded_audio_path}: {e}")

    # Return results only if some content was found or no critical error occurred
    if text_content or video_transcription or not error_occurred:
        return {
            "text": text_content or "", # Ensure string type
            "video": video_transcription or "" # Ensure string type
        }
    else:
        logger.error(f"Failed to process social media URL {url} completely.")
        return None # Indicate failure


@spaces.GPU(duration=300) # Allow more time for generation
def generate_news(instructions, facts, size, tone, *args):
    """Generate a news article based on provided data using an LLM."""
    request_start_time = time.time()
    logger.info("Received request to generate news.")
    try:
        # Ensure size is integer
        try:
            size = int(size) if size else 250 # Default size if None/empty
        except ValueError:
            logger.warning(f"Invalid size value '{size}', defaulting to 250.")
            size = 250

        # Check if models are initialized, load if necessary
        model_manager.check_llm_initialized() # LLM is essential
        # Whisper might be needed later, check/load if audio sources exist

        # --- Argument Parsing ---
        # The order *must* match the order components are added to inputs_list in create_demo
        # Fixed inputs: instructions, facts, size, tone (already passed directly)
        # Dynamic inputs from *args:
        # Expected order in *args based on create_demo:
        # 5 Documents, 15 Audio-related, 5 URLs, 9 Social-related
        num_docs = 5
        num_audio_sources = 5
        num_audio_inputs_per_source = 3
        num_urls = 5
        num_social_sources = 3
        num_social_inputs_per_source = 3

        total_expected_args = num_docs + (num_audio_sources * num_audio_inputs_per_source) + num_urls + (num_social_sources * num_social_inputs_per_source)

        args_list = list(args)
        # Pad args_list with None if fewer arguments were received than expected
        args_list.extend([None] * (total_expected_args - len(args_list)))

        # Slice arguments based on the expected order
        doc_files = args_list[0:num_docs]
        audio_inputs_flat = args_list[num_docs : num_docs + (num_audio_sources * num_audio_inputs_per_source)]
        url_inputs = args_list[num_docs + (num_audio_sources * num_audio_inputs_per_source) : num_docs + (num_audio_sources * num_audio_inputs_per_source) + num_urls]
        social_inputs_flat = args_list[num_docs + (num_audio_sources * num_audio_inputs_per_source) + num_urls : total_expected_args]

        knowledge_base = {
            "instructions": instructions or "No specific instructions provided.",
            "facts": facts or "No specific facts provided.",
            "document_content": [],
            "audio_data": [],
            "url_content": [],
            "social_content": []
        }
        raw_transcriptions = "" # Initialize transcription log

        # --- Process Inputs ---
        logger.info("Processing document inputs...")
        for i, doc_file in enumerate(doc_files):
            if doc_file and hasattr(doc_file, 'name'):
                try:
                    content = read_document(doc_file.name) # doc_file.name is the temp path
                    if content and not content.startswith("Error"):
                        # Truncate long documents for the knowledge base summary
                        doc_excerpt = (content[:1000] + "... [document truncated]") if len(content) > 1000 else content
                        knowledge_base["document_content"].append(f"[Document {i+1} Source: {os.path.basename(doc_file.name)}]\n{doc_excerpt}")
                    else:
                        logger.warning(f"Skipping document {i+1} due to read error or empty content: {content}")
                except Exception as e:
                    logger.error(f"Failed to process document {i+1} ({doc_file.name}): {e}")
            # No cleanup needed here, Gradio handles temp file uploads

        logger.info("Processing URL inputs...")
        for i, url in enumerate(url_inputs):
             if url and isinstance(url, str) and url.strip().startswith('http'):
                try:
                    content = read_url(url)
                    if content and not content.startswith("Error"):
                         # Content is already truncated in read_url if needed
                        knowledge_base["url_content"].append(f"[URL {i+1} Source: {url}]\n{content}")
                    else:
                         logger.warning(f"Skipping URL {i+1} ({url}) due to read error or empty content: {content}")
                except Exception as e:
                     logger.error(f"Failed to process URL {i+1} ({url}): {e}")

        logger.info("Processing audio/video inputs...")
        has_audio_source = False
        for i in range(num_audio_sources):
            start_idx = i * num_audio_inputs_per_source
            audio_file = audio_inputs_flat[start_idx]
            name = audio_inputs_flat[start_idx + 1] or f"Source {i+1}"
            position = audio_inputs_flat[start_idx + 2] or "N/A"

            if audio_file and hasattr(audio_file, 'name'):
                 # Store info for transcription later
                 knowledge_base["audio_data"].append({
                    "file_path": audio_file.name, # Use the temp path
                    "name": name,
                    "position": position,
                    "original_filename": os.path.basename(audio_file.name) # Keep original for logs
                })
                 has_audio_source = True
                 logger.info(f"Added audio source {i+1}: {name} ({position}) - File: {knowledge_base['audio_data'][-1]['original_filename']}")

        logger.info("Processing social media inputs...")
        has_social_source = False
        for i in range(num_social_sources):
             start_idx = i * num_social_inputs_per_source
             social_url = social_inputs_flat[start_idx]
             social_name = social_inputs_flat[start_idx + 1] or f"Social Source {i+1}"
             social_context = social_inputs_flat[start_idx + 2] or "N/A"

             if social_url and isinstance(social_url, str) and social_url.strip().startswith('http'):
                 try:
                     logger.info(f"Processing social media URL {i+1}: {social_url}")
                     social_data = process_social_media_url(social_url)
                     if social_data:
                         knowledge_base["social_content"].append({
                            "url": social_url,
                            "name": social_name,
                            "context": social_context,
                            "text": social_data.get("text", ""),
                            "video_transcription": social_data.get("video", "") # Store potential transcription
                        })
                         has_social_source = True
                         logger.info(f"Added social source {i+1}: {social_name} ({social_context}) from {social_url}")
                     else:
                         logger.warning(f"Could not retrieve any content for social URL {i+1}: {social_url}")
                 except Exception as e:
                     logger.error(f"Failed to process social URL {i+1} ({social_url}): {e}")


        # --- Transcribe Audio/Video ---
        # Only initialize Whisper if needed
        transcriptions_for_prompt = ""
        if has_audio_source or any(sc.get("video_transcription") == "[NEEDS_TRANSCRIPTION]" for sc in knowledge_base["social_content"]): # Check if transcription actually needed
            logger.info("Audio sources detected, ensuring Whisper model is ready...")
            try:
                model_manager.check_whisper_initialized()
            except Exception as whisper_init_err:
                 logger.error(f"FATAL: Whisper model initialization failed: {whisper_init_err}. Cannot transcribe.")
                 # Add error message to raw transcriptions and continue without transcriptions
                 raw_transcriptions += f"[ERROR] Whisper model failed to load. Audio sources could not be transcribed: {whisper_init_err}\n\n"
                 # Optionally return an error message immediately?
                 # return f"Error: Could not initialize transcription model. {whisper_init_err}", raw_transcriptions

        if model_manager.whisper_model: # Proceed only if whisper loaded successfully
             logger.info("Transcribing collected audio sources...")
             for idx, data in enumerate(knowledge_base["audio_data"]):
                 try:
                     logger.info(f"Transcribing audio source {idx+1}: {data['original_filename']} ({data['name']}, {data['position']})")
                     transcription = transcribe_audio_or_video(data["file_path"])
                     if transcription and not transcription.startswith("Error"):
                         quote = f'"{transcription}" - {data["name"]}, {data["position"]}'
                         transcriptions_for_prompt += f"{quote}\n\n"
                         raw_transcriptions += f'[Audio/Video {idx + 1}: {data["original_filename"]} ({data["name"]}, {data["position"]})]\n"{transcription}"\n\n'
                     else:
                         logger.warning(f"Transcription failed or returned error for audio source {idx+1}: {transcription}")
                         raw_transcriptions += f'[Audio/Video {idx + 1}: {data["original_filename"]} ({data["name"]}, {data["position"]})]\n[Error during transcription: {transcription}]\n\n'
                 except Exception as e:
                     logger.error(f"Error during transcription for audio source {idx+1} ({data['original_filename']}): {e}")
                     logger.error(traceback.format_exc())
                     raw_transcriptions += f'[Audio/Video {idx + 1}: {data["original_filename"]} ({data["name"]}, {data["position"]})]\n[Error during transcription: {e}]\n\n'
                 # Gradio handles cleanup of the uploaded temp file audio_file.name

        logger.info("Adding social media content to prompt data...")
        for idx, data in enumerate(knowledge_base["social_content"]):
            source_id = f'[Social Media {idx+1}: {data["url"]} ({data["name"]}, {data["context"]})]'
            has_content = False
            if data["text"] and not data["text"].startswith("Error"):
                # Truncate long text for the prompt, but keep full in knowledge base maybe?
                text_excerpt = (data["text"][:500] + "...[text truncated]") if len(data["text"]) > 500 else data["text"]
                social_text_prompt = f'{source_id} - Text Content:\n"{text_excerpt}"\n\n'
                transcriptions_for_prompt += social_text_prompt # Add text content as if it were a quote/source
                raw_transcriptions += f"{source_id}\nText Content:\n{data['text']}\n\n" # Log full text
                has_content = True
            if data["video_transcription"] and not data["video_transcription"].startswith("Error"):
                social_video_prompt = f'{source_id} - Video Transcription:\n"{data["video_transcription"]}"\n\n'
                transcriptions_for_prompt += social_video_prompt
                raw_transcriptions += f"{source_id}\nVideo Transcription:\n{data['video_transcription']}\n\n"
                has_content = True
            if not has_content:
                 raw_transcriptions += f"{source_id}\n[No usable text or video transcription found]\n\n"


        # --- Prepare Final Prompt ---
        # Combine document and URL summaries
        document_summary = "\n\n".join(knowledge_base["document_content"]) if knowledge_base["document_content"] else "No document content provided."
        url_summary = "\n\n".join(knowledge_base["url_content"]) if knowledge_base["url_content"] else "No URL content provided."
        transcription_summary = transcriptions_for_prompt if transcriptions_for_prompt else "No usable transcriptions available."

        # Construct the prompt for the LLM
        prompt = f"""<s>[INST] You are a professional news writer. Your task is to synthesize information from various sources into a coherent news article.

Primary Instructions: {knowledge_base["instructions"]}
Key Facts to Include: {knowledge_base["facts"]}

Supporting Information:

Document Content Summary:
{document_summary}

Web Content Summary (from URLs):
{url_summary}

Transcribed Quotes/Content (Use these directly or indirectly):
{transcription_summary}

Article Requirements:
- Title: Create a concise and informative title for the article.
- Hook: Write a compelling 15-word (approx.) hook sentence that complements the title.
- Body: Write the main news article body, aiming for approximately {size} words.
- Tone: Adopt a {tone} tone throughout the article.
- 5 Ws: Ensure the first paragraph addresses the core questions (Who, What, When, Where, Why).
- Quotes: Incorporate relevant information from the 'Transcribed Quotes/Content' section. Aim to use quotes where appropriate, but synthesize information rather than just listing quotes. Use quotation marks (" ") for direct quotes attributed correctly (e.g., based on name/position provided).
- Style: Adhere to a professional journalistic style. Be objective and factual.
- Accuracy: Do NOT invent information. Stick strictly to the provided facts, instructions, and source materials. If information is contradictory or missing, state that or omit the detail.
- Structure: Organize the article logically with clear paragraphs.

Begin the article now. [/INST]
Article Draft:
"""

        # Log the prompt length (useful for debugging context limits)
        logger.info(f"Generated prompt length: {len(prompt.split())} words / {len(prompt)} characters.")
        # Avoid logging the full prompt if it's too long or contains sensitive info
        # logger.debug(f"Generated Prompt:\n{prompt}")

        # --- Generate News Article ---
        logger.info("Generating news article with LLM...")
        generation_start_time = time.time()

        # Estimate max_new_tokens based on requested size + buffer
        # Add buffer for title, hook, and potential verbosity
        estimated_tokens_per_word = 1.5
        max_new_tokens = int(size * estimated_tokens_per_word + 150) # size words + buffer
        # Ensure max_new_tokens doesn't exceed model limits (adjust based on model's max context)
        model_max_length = 2048 # Typical for TinyLlama, but check specific model card
        # Calculate available space for generation
        # Note: This token count is approximate. Precise tokenization is needed for accuracy.
        # prompt_tokens = len(model_manager.tokenizer.encode(prompt)) # More accurate but slower
        prompt_tokens_estimate = len(prompt) // 3 # Rough estimate
        max_new_tokens = min(max_new_tokens, model_max_length - prompt_tokens_estimate - 50) # Leave buffer
        max_new_tokens = max(max_new_tokens, 100) # Ensure at least a minimum generation length

        logger.info(f"Requesting max_new_tokens: {max_new_tokens}")

        try:
            # Generate using the pipeline
            outputs = model_manager.text_pipeline(
                prompt,
                max_new_tokens=max_new_tokens, # Use max_new_tokens instead of max_length
                do_sample=True,
                temperature=0.7, # Standard temperature for creative but factual
                top_p=0.95,
                top_k=50,         # Consider adding top_k
                repetition_penalty=1.15, # Adjusted penalty
                pad_token_id=model_manager.tokenizer.eos_token_id,
                num_return_sequences=1
            )

            # Extract generated text
            generated_text = outputs[0]['generated_text']

            # Clean up the result by removing the prompt
            # Find the end of the prompt marker [/INST] and take text after it
            inst_marker = "[/INST]"
            marker_pos = generated_text.find(inst_marker)
            if marker_pos != -1:
                news_article = generated_text[marker_pos + len(inst_marker):].strip()
                # Further clean potentially leading "Article Draft:" if model included it
                if news_article.startswith("Article Draft:"):
                    news_article = news_article[len("Article Draft:"):].strip()
            else:
                 # Fallback: Try removing the input prompt string itself (less reliable)
                 if prompt in generated_text:
                     news_article = generated_text.replace(prompt, "", 1).strip()
                 else:
                      # If prompt not found exactly, assume the output is only the generation
                      # This might happen if the pipeline handles prompt removal internally sometimes
                      news_article = generated_text
                      logger.warning("Prompt marker '[/INST]' not found in LLM output. Returning full output.")


            generation_time = time.time() - generation_start_time
            logger.info(f"News generation completed in {generation_time:.2f} seconds. Output length: {len(news_article)} characters.")

        except torch.cuda.OutOfMemoryError as oom_error:
             logger.error(f"CUDA Out of Memory error during LLM generation: {oom_error}")
             logger.error(traceback.format_exc())
             model_manager.reset_models(force=True) # Attempt to recover
             raise RuntimeError("Generation failed due to insufficient GPU memory. Please try reducing article size or complexity.") from oom_error
        except Exception as gen_error:
            logger.error(f"Error during text generation pipeline: {str(gen_error)}")
            logger.error(traceback.format_exc())
            raise RuntimeError(f"LLM generation failed: {gen_error}") from gen_error

        total_time = time.time() - request_start_time
        logger.info(f"Total request processing time: {total_time:.2f} seconds.")

        # Return the generated article and the log of raw transcriptions
        return news_article, raw_transcriptions.strip()

    except Exception as e:
        total_time = time.time() - request_start_time
        logger.error(f"Error in generate_news function after {total_time:.2f} seconds: {str(e)}")
        logger.error(traceback.format_exc())
        # Attempt to reset models to recover state if possible
        try:
            model_manager.reset_models(force=True)
        except Exception as reset_error:
            logger.error(f"Failed to reset models after error: {str(reset_error)}")
        # Return error messages to the UI
        error_message = f"Error generating the news article: {str(e)}"
        transcription_log = raw_transcriptions.strip() + f"\n\n[ERROR] News generation failed: {str(e)}"
        return error_message, transcription_log

def create_demo():
    """Creates the Gradio interface"""
    with gr.Blocks(theme=gr.themes.Soft()) as demo:
        gr.Markdown("# 📰 NewsIA - AI News Generator")
        gr.Markdown("Create professional news articles from multiple information sources.")

        # Store all input components for easy access/reset
        all_inputs = []

        with gr.Row():
            with gr.Column(scale=2):
                instructions = gr.Textbox(
                    label="Instructions for the News Article",
                    placeholder="Enter specific instructions for generating your news article (e.g., focus on the economic impact)",
                    lines=2,
                    value=""
                )
                all_inputs.append(instructions)

                facts = gr.Textbox(
                    label="Main Facts",
                    placeholder="Describe the most important facts the news should include (e.g., Event name, date, location, key people involved)",
                    lines=4,
                    value=""
                )
                all_inputs.append(facts)

                with gr.Row():
                    size_slider = gr.Slider(
                        label="Approximate Length (words)",
                        minimum=100,
                        maximum=700, # Increased max size
                        value=250,
                        step=50
                    )
                    all_inputs.append(size_slider)

                    tone_dropdown = gr.Dropdown(
                        label="Tone of the News Article",
                        choices=["neutral", "serious", "formal", "urgent", "investigative", "human-interest", "lighthearted"],
                        value="neutral"
                    )
                    all_inputs.append(tone_dropdown)

            with gr.Column(scale=3):
                with gr.Tabs():
                    with gr.TabItem("📝 Documents"):
                        gr.Markdown("Upload relevant documents (PDF, DOCX, XLSX, CSV). Max 5.")
                        doc_inputs = []
                        for i in range(1, 6):
                            doc_file = gr.File(
                                label=f"Document {i}",
                                file_types=["pdf", ".docx", ".xlsx", ".csv"], # Explicit extensions for clarity
                                file_count="single" # Ensure single file per component
                            )
                            doc_inputs.append(doc_file)
                        all_inputs.extend(doc_inputs)

                    with gr.TabItem("🔊 Audio/Video"):
                         gr.Markdown("Upload audio or video files for transcription (MP3, WAV, MP4, MOV, etc.). Max 5 sources.")
                         audio_video_inputs = []
                         for i in range(1, 6):
                            with gr.Group():
                                gr.Markdown(f"**Source {i}**")
                                audio_file = gr.File(
                                    label=f"Audio/Video File {i}",
                                    file_types=["audio", "video"]
                                )
                                with gr.Row():
                                    speaker_name = gr.Textbox(
                                        label="Speaker Name",
                                        placeholder="Name of the interviewee or speaker",
                                        value=""
                                    )
                                    speaker_role = gr.Textbox(
                                        label="Role/Position",
                                        placeholder="Speaker's title or role",
                                        value=""
                                    )
                                audio_video_inputs.append(audio_file)
                                audio_video_inputs.append(speaker_name)
                                audio_video_inputs.append(speaker_role)
                         all_inputs.extend(audio_video_inputs)

                    with gr.TabItem("🌐 URLs"):
                         gr.Markdown("Add URLs to relevant web pages or articles. Max 5.")
                         url_inputs = []
                         for i in range(1, 6):
                            url_textbox = gr.Textbox(
                                label=f"URL {i}",
                                placeholder="https://example.com/article",
                                value=""
                            )
                            url_inputs.append(url_textbox)
                         all_inputs.extend(url_inputs)

                    with gr.TabItem("📱 Social Media"):
                         gr.Markdown("Add URLs to social media posts (e.g., Twitter, YouTube, TikTok). Max 3.")
                         social_inputs = []
                         for i in range(1, 4):
                            with gr.Group():
                                gr.Markdown(f"**Social Media Source {i}**")
                                social_url_textbox = gr.Textbox(
                                    label=f"Post URL",
                                    placeholder="https://twitter.com/user/status/...",
                                    value=""
                                )
                                with gr.Row():
                                    social_name_textbox = gr.Textbox(
                                        label=f"Account Name/User",
                                        placeholder="Name or handle (e.g., @username)",
                                        value=""
                                    )
                                    social_context_textbox = gr.Textbox(
                                        label=f"Context",
                                        placeholder="Brief context (e.g., statement on event X)",
                                        value=""
                                    )
                                social_inputs.append(social_url_textbox)
                                social_inputs.append(social_name_textbox)
                                social_inputs.append(social_context_textbox)
                         all_inputs.extend(social_inputs)


        with gr.Row():
            generate_button = gr.Button("✨ Generate News Article", variant="primary")
            clear_button = gr.Button("🔄 Clear All Inputs")

        with gr.Tabs():
            with gr.TabItem("📄 Generated News Article"):
                news_output = gr.Textbox(
                    label="Draft News Article",
                    lines=20, # Increased lines
                    show_copy_button=True,
                    value=""
                )
            with gr.TabItem("🎙️ Source Transcriptions & Logs"):
                transcriptions_output = gr.Textbox(
                    label="Transcriptions and Processing Log",
                    lines=15, # Increased lines
                    show_copy_button=True,
                    value=""
                )

        # --- Event Handlers ---
        # Define outputs
        outputs_list = [news_output, transcriptions_output]

        # Generate button click
        generate_button.click(
            fn=generate_news,
            inputs=all_inputs, # Pass the consolidated list
            outputs=outputs_list
        )

        # Clear button click
        def clear_all_inputs_and_outputs():
            # Return a list of default values matching the number and type of inputs + outputs
            reset_values = []
            for input_comp in all_inputs:
                 # Default for Textbox, Dropdown is "", for Slider is its default, for File is None
                if isinstance(input_comp, (gr.Textbox, gr.Dropdown)):
                    reset_values.append("")
                elif isinstance(input_comp, gr.Slider):
                    # Find the original default value if needed, or just use a sensible default
                    reset_values.append(250) # Reset slider to default
                elif isinstance(input_comp, gr.File):
                    reset_values.append(None)
                else:
                    reset_values.append(None) # Default for unknown/other types

            # Add default values for the output fields
            reset_values.extend(["", ""]) # Two Textbox outputs

            # Also reset the models in the background
            model_manager.reset_models(force=True)
            logger.info("UI cleared and models reset.")

            return reset_values

        clear_button.click(
            fn=clear_all_inputs_and_outputs,
            inputs=None, # No inputs needed for the clear function itself
            outputs=all_inputs + outputs_list # The list of components to clear
        )

        # Add event handler to reset models when the Gradio app closes or reloads (if possible)
        # demo.unload(model_manager.reset_models, inputs=None, outputs=None) # Might not work reliably in Spaces

    return demo

if __name__ == "__main__":
    logger.info("Starting NewsIA application...")

    # Optional: Pre-initialize Whisper on startup if desired and resources allow
    # This can make the first transcription faster but uses GPU resources immediately.
    # Consider enabling only if transcriptions are very common.
    # try:
    #     logger.info("Attempting to pre-initialize Whisper model...")
    #     model_manager.initialize_whisper()
    # except Exception as e:
    #     logger.warning(f"Pre-initialization of Whisper model failed (will load on demand): {str(e)}")

    # Create the Gradio Demo
    news_demo = create_demo()

    # Configure the queue - remove concurrency_count and max_size
    # Use default queue settings, suitable for most Spaces environments
    news_demo.queue()

    # Launch the Gradio app
    logger.info("Launching Gradio interface...")
    news_demo.launch(
        server_name="0.0.0.0", # Necessary for Docker/Spaces
        server_port=7860,
        # share=True # Share=True is often handled by Spaces automatically, can be removed
        # debug=True # Enable for more detailed Gradio logs if needed
    )
    logger.info("NewsIA application finished.")