Spaces:
Sleeping
Sleeping
| import os | |
| import io | |
| import re | |
| import logging | |
| import subprocess | |
| from datetime import datetime | |
| import urllib.parse | |
| import tempfile | |
| import json # For streaming JSON messages | |
| import time # For gevent.sleep | |
| from flask import Flask, request, render_template, Response, stream_with_context | |
| from werkzeug.utils import secure_filename | |
| # Ensure gevent is imported and monkey patched if needed for other libraries | |
| # that might not be gevent-friendly. For built-in libs and requests (with Gunicorn gevent worker), | |
| # this is often handled by Gunicorn. | |
| # from gevent import monkey | |
| # monkey.patch_all() # Apply this early if you suspect issues with other libs | |
| import requests | |
| import pdfplumber | |
| from pdf2image import convert_from_path, convert_from_bytes | |
| import pytesseract | |
| from PIL import Image | |
| from huggingface_hub import HfApi, create_repo, HfHubHTTPError | |
| # --- Flask App Initialization --- | |
| app = Flask(__name__) | |
| app.config['UPLOAD_FOLDER'] = tempfile.gettempdir() | |
| app.config['MAX_CONTENT_LENGTH'] = 50 * 1024 * 1024 # 50 MB limit for uploads, adjust as needed | |
| # --- Logging Configuration --- | |
| logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s") | |
| logger = logging.getLogger(__name__) | |
| # --- Hugging Face Configuration --- | |
| HF_TOKEN = os.getenv("HF_TOKEN") | |
| HF_DATASET_REPO_NAME = os.getenv("HF_DATASET_REPO_NAME", "pdf-images-extracted") | |
| hf_api = HfApi() | |
| # --- Helper to yield messages for streaming --- | |
| def yield_message(type, data): | |
| """Helper to format messages as JSON strings for streaming.""" | |
| # Add a newline so client can easily split messages | |
| return json.dumps({"type": type, **data}) + "\n" | |
| # --- PDF Processing Helper Functions (Adapted for Streaming) --- | |
| def check_poppler(): | |
| # (Same as before) | |
| try: | |
| result = subprocess.run(["pdftoppm", "-v"], capture_output=True, text=True, check=False) | |
| version_info_log = result.stderr.strip() if result.stderr else result.stdout.strip() | |
| if version_info_log: | |
| logger.info(f"Poppler version check: {version_info_log.splitlines()[0] if version_info_log else 'No version output'}") | |
| else: | |
| logger.info("Poppler 'pdftoppm -v' ran. Assuming Poppler is present.") | |
| return True | |
| except FileNotFoundError: | |
| logger.error("Poppler (pdftoppm command) not found. Ensure poppler-utils is installed and in PATH.") | |
| return False | |
| except Exception as e: | |
| logger.error(f"An unexpected error occurred during Poppler check: {str(e)}") | |
| return False | |
| def ensure_hf_dataset(): | |
| # (Same as before, but logs info useful for streaming if an error occurs) | |
| if not HF_TOKEN: | |
| msg = "HF_TOKEN is not set. Cannot ensure Hugging Face dataset. Image uploads will fail." | |
| logger.warning(msg) | |
| return "Error: " + msg | |
| try: | |
| repo_id_obj = create_repo(repo_id=HF_DATASET_REPO_NAME, token=HF_TOKEN, repo_type="dataset", exist_ok=True) | |
| logger.info(f"Dataset repo ensured: {repo_id_obj.repo_id}") | |
| return repo_id_obj.repo_id | |
| except HfHubHTTPError as e: | |
| if e.response.status_code == 409: | |
| logger.info(f"Dataset repo '{HF_DATASET_REPO_NAME}' already exists.") | |
| # Attempt to construct the full repo_id (namespace/repo_name) | |
| try: | |
| user_info = hf_api.whoami(token=HF_TOKEN) | |
| namespace = user_info.get('name') if user_info else None | |
| if namespace: | |
| return f"{namespace}/{HF_DATASET_REPO_NAME}" | |
| except Exception as whoami_e: | |
| logger.error(f"Could not determine namespace for existing repo via whoami: {whoami_e}") | |
| return f"hf://datasets/{HF_DATASET_REPO_NAME}" # Fallback, might not be full id | |
| logger.error(f"Hugging Face dataset error (HTTP {e.response.status_code}): {str(e)}") | |
| return f"Error: Failed to access or create dataset '{HF_DATASET_REPO_NAME}': {str(e)}" | |
| except Exception as e: | |
| logger.error(f"Hugging Face dataset error: {str(e)}", exc_info=True) | |
| return f"Error: Failed to access or create dataset '{HF_DATASET_REPO_NAME}': {str(e)}" | |
| def upload_image_to_hf_stream(image_pil, filename_base, page_num_for_log=""): | |
| # (Adapted to potentially yield status during this sub-process if it were longer) | |
| # For now, it's synchronous but part of the larger stream. | |
| repo_id_or_error = ensure_hf_dataset() | |
| if isinstance(repo_id_or_error, str) and repo_id_or_error.startswith("Error"): | |
| return repo_id_or_error | |
| repo_id = repo_id_or_error | |
| temp_image_path = None | |
| try: | |
| timestamp = datetime.now().strftime("%Y%m%d_%H%M%S_%f") | |
| repo_filename = f"images/{filename_base}_{page_num_for_log}_{timestamp}.png" | |
| with tempfile.NamedTemporaryFile(delete=False, suffix=".png", dir=app.config['UPLOAD_FOLDER']) as tmp_file: | |
| temp_image_path = tmp_file.name | |
| image_pil.save(temp_image_path, format="PNG") | |
| logger.info(f"Attempting to upload {temp_image_path} to {repo_id}/{repo_filename}") | |
| file_url = hf_api.upload_file( | |
| path_or_fileobj=temp_image_path, path_in_repo=repo_filename, | |
| repo_id=repo_id, repo_type="dataset", token=HF_TOKEN | |
| ) | |
| logger.info(f"Successfully uploaded image: {file_url}") | |
| return file_url | |
| except Exception as e: | |
| logger.error(f"Image upload error for {filename_base}{page_num_for_log}: {str(e)}", exc_info=True) | |
| return f"Error uploading image {filename_base}{page_num_for_log}: {str(e)}" | |
| finally: | |
| if temp_image_path and os.path.exists(temp_image_path): | |
| try: os.remove(temp_image_path) | |
| except OSError as ose: logger.error(f"Error removing temp image file {temp_image_path}: {ose}") | |
| def format_page_text_to_markdown_chunk(page_text_content): | |
| """Formats a single page's text content into a markdown chunk. | |
| More complex formatting logic can be applied here page by page. | |
| """ | |
| chunk_md = "" | |
| # Normalize newlines: multiple consecutive newlines become a single blank line (two \n chars) | |
| page_text_content = re.sub(r'\n\s*\n+', '\n\n', page_text_content.strip()) | |
| lines = page_text_content.split('\n') | |
| is_in_list = False | |
| for line_text in lines: | |
| line_stripped = line_text.strip() | |
| if not line_stripped: | |
| chunk_md += "\n" | |
| is_in_list = False | |
| continue | |
| list_match = re.match(r'^\s*(?:(?:\d+\.)|[*+-])\s+(.*)', line_stripped) | |
| is_heading_candidate = line_stripped.isupper() and 5 < len(line_stripped) < 100 | |
| if is_heading_candidate and not list_match: | |
| chunk_md += f"## {line_stripped}\n\n" | |
| is_in_list = False | |
| elif list_match: | |
| list_item_text = list_match.group(1) | |
| chunk_md += f"- {list_item_text}\n" | |
| is_in_list = True | |
| else: | |
| if is_in_list: chunk_md += "\n" | |
| chunk_md += f"{line_text}\n\n" | |
| is_in_list = False | |
| return re.sub(r'\n\s*\n+', '\n\n', chunk_md.strip()) + "\n\n" | |
| # --- Main PDF Processing Logic (Generator Function for Streaming) --- | |
| def generate_pdf_conversion_stream(pdf_input_source_path_or_url): | |
| """ | |
| Processes the PDF incrementally and yields status messages and markdown content. | |
| `pdf_input_source_path_or_url` is a local file path or a URL string. | |
| """ | |
| try: | |
| # Initial Markdown Title | |
| yield yield_message("markdown_replace", {"content": "# Extracted PDF Content\n\n"}) | |
| time.sleep(0.01) # Give gevent a chance to yield | |
| # 1. Text and Table Extraction (Page by Page) | |
| yield yield_message("status", {"message": "Opening PDF for text extraction..."}) | |
| time.sleep(0.01) | |
| source_is_url = isinstance(pdf_input_source_path_or_url, str) and \ | |
| pdf_input_source_path_or_url.startswith(('http://', 'https://')) | |
| pdf_handle_for_text = None | |
| pdf_bytes_for_images = None # Store bytes if downloaded from URL for image extraction | |
| if source_is_url: | |
| try: | |
| response = requests.get(pdf_input_source_path_or_url, stream=True, timeout=60) # Increased timeout | |
| response.raise_for_status() | |
| pdf_bytes_for_images = response.content # Read all content for pdf2image | |
| pdf_handle_for_text = io.BytesIO(pdf_bytes_for_images) # Use BytesIO for pdfplumber | |
| yield yield_message("status", {"message": f"PDF downloaded from URL ({len(pdf_bytes_for_images)/1024:.2f} KB)."}) | |
| time.sleep(0.01) | |
| except requests.RequestException as e: | |
| logger.error(f"URL fetch error for PDF processing: {str(e)}", exc_info=True) | |
| yield yield_message("error", {"message": f"Error fetching PDF from URL: {str(e)}"}) | |
| return # Stop generation | |
| else: # Local file path | |
| pdf_handle_for_text = pdf_input_source_path_or_url # pdfplumber takes path | |
| total_text_pages = 0 | |
| try: | |
| with pdfplumber.open(pdf_handle_for_text) as pdf: | |
| total_text_pages = len(pdf.pages) | |
| yield yield_message("status", {"message": f"Found {total_text_pages} page(s) for text extraction."}) | |
| time.sleep(0.01) | |
| for i, page in enumerate(pdf.pages): | |
| yield yield_message("status", {"message": f"Extracting text from page {i+1}/{total_text_pages}..."}) | |
| time.sleep(0.01) # gevent yield | |
| page_text = page.extract_text(layout=True, x_density=1, y_density=1) or "" | |
| page_tables_md = "" | |
| tables = page.extract_tables() | |
| if tables: | |
| for table_idx, table_data in enumerate(tables): | |
| if table_data: | |
| yield yield_message("status", {"message": f" Processing table {table_idx+1} on page {i+1}..."}) | |
| header = [" | ".join(str(cell) if cell is not None else "" for cell in table_data[0])] | |
| separator = [" | ".join(["---"] * len(table_data[0]))] | |
| body = [" | ".join(str(cell) if cell is not None else "" for cell in row) for row in table_data[1:]] | |
| table_md_lines = header + separator + body | |
| page_tables_md += f"**Table (Page {i+1}):**\n" + "\n".join(table_md_lines) + "\n\n" | |
| formatted_page_text_md = format_page_text_to_markdown_chunk(page_text) | |
| yield yield_message("markdown_chunk", {"content": formatted_page_text_md}) | |
| if page_tables_md: | |
| yield yield_message("markdown_chunk", {"content": page_tables_md}) | |
| time.sleep(0.01) # gevent yield | |
| except Exception as e: | |
| logger.error(f"Error during PDF text/table extraction: {str(e)}", exc_info=True) | |
| yield yield_message("error", {"message": f"Error during text extraction: {str(e)}"}) | |
| # Continue to image extraction if possible, or return based on severity | |
| # 2. Image Extraction and OCR | |
| if not check_poppler(): | |
| yield yield_message("error", {"message": "Poppler (for image extraction) not found or not working."}) | |
| else: | |
| yield yield_message("status", {"message": "Starting image extraction..."}) | |
| yield yield_message("markdown_chunk", {"content": "## Extracted Images\n\n"}) | |
| if not HF_TOKEN: | |
| yield yield_message("markdown_chunk", {"content": "**Note:** `HF_TOKEN` not set. Images will be described but not uploaded.\n\n"}) | |
| time.sleep(0.01) | |
| extracted_pil_images = [] | |
| try: | |
| if source_is_url and pdf_bytes_for_images: | |
| # Use the already downloaded bytes | |
| extracted_pil_images = convert_from_bytes(pdf_bytes_for_images, dpi=150) # Lower DPI for speed/memory | |
| elif not source_is_url: # local file path | |
| extracted_pil_images = convert_from_path(pdf_input_source_path_or_url, dpi=150) | |
| yield yield_message("status", {"message": f"Found {len(extracted_pil_images)} image(s) in PDF (these are rasterized pages for now)."}) | |
| time.sleep(0.01) | |
| # TODO: Implement more granular image extraction if pdf2image supports it, | |
| # or if you integrate a library that can extract embedded images directly. | |
| # For now, convert_from_path/bytes often gives full pages as images. | |
| for i, img_pil in enumerate(extracted_pil_images): | |
| page_num_for_log = f"page_{i+1}" # Assuming one image per page from convert_from_path | |
| yield yield_message("status", {"message": f"Processing image {i+1}/{len(extracted_pil_images)} (OCR & Upload)..."}) | |
| time.sleep(0.01) | |
| ocr_text = "" | |
| try: | |
| ocr_text = pytesseract.image_to_string(img_pil).strip() | |
| if ocr_text: | |
| yield yield_message("status", {"message": f" OCR successful for image {i+1}."}) | |
| except Exception as ocr_e: | |
| logger.error(f"OCR error for image {i+1}: {str(ocr_e)}") | |
| ocr_text = f"OCR failed: {str(ocr_e)}" | |
| image_md_chunk = "" | |
| if HF_TOKEN: | |
| image_url_or_error = upload_image_to_hf_stream(img_pil, "pdf_image", page_num_for_log) | |
| if isinstance(image_url_or_error, str) and not image_url_or_error.startswith("Error"): | |
| image_md_chunk += f"\n" | |
| yield yield_message("status", {"message": f" Image {i+1} uploaded."}) | |
| else: | |
| image_md_chunk += f"**Image {i+1} (Upload Error):** {str(image_url_or_error)}\n\n" | |
| yield yield_message("error", {"message": f"Failed to upload image {i+1}: {str(image_url_or_error)}"}) | |
| else: | |
| image_md_chunk += f"**Image {i+1} (not uploaded due to missing HF_TOKEN)**\n" | |
| if ocr_text: | |
| image_md_chunk += f"**Image {i+1} OCR Text:**\n```\n{ocr_text}\n```\n\n" | |
| yield yield_message("image_md", {"content": image_md_chunk}) | |
| time.sleep(0.01) # gevent yield | |
| except Exception as e: | |
| logger.error(f"Error during image extraction/processing: {str(e)}", exc_info=True) | |
| yield yield_message("error", {"message": f"Error during image extraction: {str(e)}"}) | |
| yield yield_message("final_status", {"message": "All processing stages complete."}) | |
| except Exception as e: | |
| logger.error(f"Unhandled error in PDF conversion stream: {str(e)}", exc_info=True) | |
| yield yield_message("error", {"message": f"Critical processing error: {str(e)}"}) | |
| # --- Flask Routes --- | |
| def index(): | |
| return render_template('index.html') | |
| def process_pdf_stream(): | |
| pdf_file = request.files.get('pdf_file') | |
| pdf_url = request.form.get('pdf_url', '').strip() | |
| temp_pdf_path = None # To store path of uploaded file for cleanup | |
| pdf_input_source_for_generator = None | |
| def stream_processor(): | |
| nonlocal temp_pdf_path # Make it accessible in this inner function for cleanup | |
| nonlocal pdf_input_source_for_generator | |
| try: | |
| if pdf_file and pdf_file.filename: | |
| if not pdf_file.filename.lower().endswith('.pdf'): | |
| yield yield_message("error", {"message": "Uploaded file is not a PDF."}) | |
| return | |
| filename = secure_filename(pdf_file.filename) | |
| # Save to a temporary file (ensure UPLOAD_FOLDER is writable by app user) | |
| os.makedirs(app.config['UPLOAD_FOLDER'], exist_ok=True) | |
| fd, temp_pdf_path = tempfile.mkstemp(suffix=".pdf", prefix="upload_", dir=app.config['UPLOAD_FOLDER']) | |
| os.close(fd) | |
| pdf_file.save(temp_pdf_path) | |
| logger.info(f"Uploaded PDF saved to temporary path: {temp_pdf_path}") | |
| pdf_input_source_for_generator = temp_pdf_path | |
| yield yield_message("status", {"message": f"Processing uploaded PDF: {filename}"}) | |
| time.sleep(0.01) | |
| elif pdf_url: | |
| unquoted_url = urllib.parse.unquote(pdf_url) | |
| if not (unquoted_url.startswith('http://') or unquoted_url.startswith('https://')): | |
| yield yield_message("error", {"message": "Invalid URL scheme. Must be http or https."}) | |
| return | |
| # Consider a light check for .pdf extension, but content-type is more reliable | |
| pdf_input_source_for_generator = unquoted_url | |
| yield yield_message("status", {"message": f"Preparing to process PDF from URL: {unquoted_url}"}) | |
| time.sleep(0.01) | |
| else: | |
| yield yield_message("error", {"message": "No PDF file uploaded and no PDF URL provided."}) | |
| return | |
| # Yield from the main generator | |
| for message_part in generate_pdf_conversion_stream(pdf_input_source_for_generator): | |
| yield message_part | |
| # time.sleep(0.01) # Allow gevent to switch context, important for streaming | |
| except Exception as e: | |
| logger.error(f"Error setting up stream or in initial validation: {str(e)}", exc_info=True) | |
| yield yield_message("error", {"message": f"Setup error: {str(e)}"}) | |
| finally: | |
| if temp_pdf_path and os.path.exists(temp_pdf_path): | |
| try: | |
| os.remove(temp_pdf_path) | |
| logger.info(f"Cleaned up temporary PDF: {temp_pdf_path}") | |
| yield yield_message("status", {"message": f"Cleaned up temporary file."}) | |
| except OSError as ose: | |
| logger.error(f"Error removing temporary PDF {temp_pdf_path}: {ose}") | |
| yield yield_message("error", {"message": f"Could not clean temp file: {ose}"}) | |
| # Using stream_with_context for proper handling of request context within the generator | |
| return Response(stream_with_context(stream_processor()), mimetype='application/x-ndjson') | |
| # --- Main Execution --- | |
| if __name__ == '__main__': | |
| if not check_poppler(): # Check Poppler at startup for local dev | |
| logger.warning("Poppler utilities might not be installed correctly. PDF processing might fail.") | |
| os.makedirs(app.config['UPLOAD_FOLDER'], exist_ok=True) | |
| # For local dev, Flask's built-in server is fine. Gunicorn handles production. | |
| # The 'threaded=True' or using gevent server locally can also help test streaming. | |
| app.run(host='0.0.0.0', port=int(os.getenv("PORT", 7860)), debug=True, threaded=True) |