from transformers import MllamaForConditionalGeneration, AutoProcessor, TextIteratorStreamer from PIL import Image import requests import torch from threading import Thread import gradio as gr from gradio import FileData import time import spaces import fitz # PyMuPDF import io import numpy as np import logging # Set up logging logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__name__) # Load model and processor ckpt = "Daemontatox/DocumentCogito" model = MllamaForConditionalGeneration.from_pretrained(ckpt, torch_dtype=torch.bfloat16).to("cuda") processor = AutoProcessor.from_pretrained(ckpt) class DocumentState: def __init__(self): self.current_doc_images = [] self.current_doc_text = "" self.doc_type = None def clear(self): self.current_doc_images = [] self.current_doc_text = "" self.doc_type = None doc_state = DocumentState() def process_pdf_file(file_path): """ Convert PDF to images and extract text using PyMuPDF with improved error handling and image quality settings. """ try: doc = fitz.open(file_path) images = [] text = "" for page_num in range(doc.page_count): try: page = doc[page_num] # Extract text with better formatting page_text = page.get_text("text") if page_text.strip(): # Only add non-empty pages text += f"Page {page_num + 1}:\n{page_text}\n\n" # Improved image extraction with error handling try: # Use higher DPI for better quality zoom = 2 # Increase zoom factor for better resolution mat = fitz.Matrix(zoom, zoom) pix = page.get_pixmap(matrix=mat, alpha=False) # Convert to PIL Image with proper color handling img_data = pix.tobytes("png") img = Image.open(io.BytesIO(img_data)) # Ensure RGB mode and reasonable size img = img.convert("RGB") # Resize if image is too large (keeping aspect ratio) max_size = 1600 if max(img.size) > max_size: ratio = max_size / max(img.size) new_size = tuple(int(dim * ratio) for dim in img.size) img = img.resize(new_size, Image.Resampling.LANCZOS) images.append(img) except Exception as e: logger.error(f"Error processing page {page_num} image: {str(e)}") continue except Exception as e: logger.error(f"Error processing page {page_num}: {str(e)}") continue doc.close() if not images: raise ValueError("No valid images could be extracted from the PDF") return images, text except Exception as e: logger.error(f"Error processing PDF file: {str(e)}") raise def process_file(file): """Process either PDF or image file with improved error handling.""" try: doc_state.clear() if isinstance(file, dict): file_path = file["path"] else: file_path = file if file_path.lower().endswith('pdf'): doc_state.doc_type = 'pdf' try: doc_state.current_doc_images, doc_state.current_doc_text = process_pdf_file(file_path) return f"PDF processed successfully. Total pages: {len(doc_state.current_doc_images)}. You can now ask questions about the content." except Exception as e: return f"Error processing PDF: {str(e)}. Please try a different PDF file or check if the file is corrupted." else: doc_state.doc_type = 'image' try: img = Image.open(file_path).convert("RGB") # Resize if necessary max_size = 1600 if max(img.size) > max_size: ratio = max_size / max(img.size) new_size = tuple(int(dim * ratio) for dim in img.size) img = img.resize(new_size, Image.Resampling.LANCZOS) doc_state.current_doc_images = [img] return "Image loaded successfully. You can now ask questions about the content." except Exception as e: return f"Error processing image: {str(e)}. Please try a different image file." except Exception as e: logger.error(f"Error in process_file: {str(e)}") return "An error occurred while processing the file. Please try again." @spaces.GPU() def bot_streaming(message, history, max_new_tokens=8192): try: txt = message["text"] messages = [] # Process new file if provided if message.get("files") and len(message["files"]) > 0: result = process_file(message["files"][0]) if "Error" in result: yield result return # Process history with better error handling for i, msg in enumerate(history): try: if isinstance(msg[0], dict): user_content = [{"type": "text", "text": msg[0]["text"]}] if "files" in msg[0] and len(msg[0]["files"]) > 0: user_content.append({"type": "image"}) messages.append({"role": "user", "content": user_content}) messages.append({"role": "assistant", "content": [{"type": "text", "text": msg[1]}]}) elif isinstance(msg[0], str): messages.append({"role": "user", "content": [{"type": "text", "text": msg[0]}]}) messages.append({"role": "assistant", "content": [{"type": "text", "text": msg[1]}]}) except Exception as e: logger.error(f"Error processing history message {i}: {str(e)}") continue # Include document context if doc_state.current_doc_images: context = f"\nDocument context:\n{doc_state.current_doc_text}" if doc_state.current_doc_text else "" current_msg = f"{txt}{context}" messages.append({"role": "user", "content": [{"type": "text", "text": current_msg}, {"type": "image"}]}) else: messages.append({"role": "user", "content": [{"type": "text", "text": txt}]}) # Process inputs texts = processor.apply_chat_template(messages, add_generation_prompt=True) try: if doc_state.current_doc_images: inputs = processor( text=texts, images=doc_state.current_doc_images[0:1], return_tensors="pt" ).to("cuda") else: inputs = processor(text=texts, return_tensors="pt").to("cuda") streamer = TextIteratorStreamer(processor, skip_special_tokens=True, skip_prompt=True) generation_kwargs = dict(inputs, streamer=streamer, max_new_tokens=max_new_tokens) thread = Thread(target=model.generate, kwargs=generation_kwargs) thread.start() buffer = "" for new_text in streamer: buffer += new_text time.sleep(0.01) yield buffer except Exception as e: logger.error(f"Error in model processing: {str(e)}") yield "An error occurred while processing your request. Please try again." except Exception as e: logger.error(f"Error in bot_streaming: {str(e)}") yield "An error occurred. Please try again." def clear_context(): """Clear the current document context.""" doc_state.clear() return "Document context cleared. You can upload a new document." # Create the Gradio interface with gr.Blocks() as demo: gr.Markdown("# Document Analyzer with Chat Support") gr.Markdown("Upload a PDF or image and chat about its contents. For PDFs, all pages will be processed for visual analysis.") chatbot = gr.ChatInterface( fn=bot_streaming, title="Document Chat", examples=[ [{"text": "Which era does this piece belong to? Give details about the era.", "files":["./examples/rococo.jpg"]}, 200], [{"text": "Where do the droughts happen according to this diagram?", "files":["./examples/weather_events.png"]}, 250], [{"text": "What happens when you take out white cat from this chain?", "files":["./examples/ai2d_test.jpg"]}, 250], [{"text": "How long does it take from invoice date to due date? Be short and concise.", "files":["./examples/invoice.png"]}, 250], [{"text": "Where to find this monument? Can you give me other recommendations around the area?", "files":["./examples/wat_arun.jpg"]}, 250], ], textbox=gr.MultimodalTextbox(), additional_inputs=[ gr.Slider( minimum=10, maximum=2048, value=8192, step=10, label="Maximum number of new tokens to generate", ) ], cache_examples=False, stop_btn="Stop Generation", fill_height=True, multimodal=True ) clear_btn = gr.Button("Clear Document Context") clear_btn.click(fn=clear_context) chatbot.textbox.file_types = ["image", "pdf", "text"] # Launch the interface demo.launch(debug=True)