SpencerCPurdy's picture
Create app.py
e8538b0 verified
# Multi-Modal Document Intelligence System
# Author: Spencer Purdy
# Description: An advanced document analysis tool that combines LayoutLMv3 for document understanding
# with efficient language models to extract information, summarize, and answer questions about documents.
# Optimized for Google Colab Pro performance.
import subprocess
import sys
import os
import io
from typing import List, Dict, Tuple, Optional
import json
import re
import hashlib
import time
# Install required packages function
def install_packages():
"""Install all required packages for the document intelligence system"""
packages = [
'gradio',
'transformers',
'torch',
'torchvision',
'Pillow',
'pytesseract',
'pdf2image',
'opencv-python',
'sentencepiece',
'accelerate'
]
print("Installing required packages...")
for package in packages:
subprocess.check_call([sys.executable, '-m', 'pip', 'install', package, '-q'])
# Install system dependencies for PDF processing and OCR
print("Installing system dependencies...")
subprocess.check_call(['apt-get', 'update', '-qq'])
subprocess.check_call(['apt-get', 'install', '-y', '-qq', 'poppler-utils', 'tesseract-ocr'])
# Try importing, install if needed
try:
import gradio as gr
from transformers import (
AutoProcessor, AutoModelForTokenClassification,
AutoTokenizer, AutoModelForSeq2SeqLM,
pipeline
)
import torch
from PIL import Image
import pytesseract
from pdf2image import convert_from_path
import cv2
import numpy as np
except ImportError:
print("Installing required packages...")
install_packages()
# Re-import after installation
import gradio as gr
from transformers import (
AutoProcessor, AutoModelForTokenClassification,
AutoTokenizer, AutoModelForSeq2SeqLM,
pipeline
)
import torch
from PIL import Image
import pytesseract
from pdf2image import convert_from_path
import cv2
import numpy as np
# Configure device for optimal performance
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")
# Model initialization with optimized settings
print("Loading models...")
# Load LayoutLMv3 for document structure understanding
print("Loading LayoutLMv3...")
layoutlm_processor = AutoProcessor.from_pretrained("microsoft/layoutlmv3-base", apply_ocr=False)
layoutlm_model = AutoModelForTokenClassification.from_pretrained(
"microsoft/layoutlmv3-base",
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
).to(device)
layoutlm_model.eval() # Set to evaluation mode for faster inference
# Load efficient T5 model for text generation (much faster than Phi-2)
print("Loading T5 model for summarization and Q&A...")
t5_model_name = "google/flan-t5-base" # 250M parameters, efficient and effective
t5_tokenizer = AutoTokenizer.from_pretrained(t5_model_name)
t5_model = AutoModelForSeq2SeqLM.from_pretrained(
t5_model_name,
torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
).to(device)
t5_model.eval() # Set to evaluation mode
print("Models loaded successfully!")
class DocumentProcessor:
"""
Main document processing class that handles OCR, text extraction,
summarization, and question answering for various document types.
"""
def __init__(self):
"""Initialize the document processor with empty state"""
self.extracted_text = ""
self.document_metadata = {}
self.page_contents = []
self.processing_cache = {} # Cache for processed documents
def _get_file_hash(self, file_path: str) -> str:
"""Generate a hash for the file to use as cache key"""
with open(file_path, 'rb') as f:
return hashlib.md5(f.read()).hexdigest()
def process_pdf(self, pdf_path: str, max_pages: int = 20) -> List[Image.Image]:
"""
Convert PDF pages to images for OCR processing
Args:
pdf_path: Path to the PDF file
max_pages: Maximum number of pages to process (for memory management)
Returns:
List of PIL Images representing PDF pages
"""
try:
# Convert PDF to images with resolution optimization
images = convert_from_path(
pdf_path,
dpi=150, # Balance between quality and performance
first_page=1,
last_page=min(max_pages, 100) # Limit pages for memory
)
return images
except Exception as e:
print(f"Error processing PDF: {e}")
return []
def extract_text_from_image(self, image: Image.Image) -> Dict[str, any]:
"""
Extract text and layout information from an image using OCR
Args:
image: PIL Image to process
Returns:
Dictionary containing extracted text and metadata
"""
try:
# Resize image if too large to improve performance
max_dimension = 2000
if max(image.size) > max_dimension:
ratio = max_dimension / max(image.size)
new_size = tuple(int(dim * ratio) for dim in image.size)
image = image.resize(new_size, Image.Resampling.LANCZOS)
# Convert to numpy array for OCR
image_np = np.array(image)
# Perform OCR with Tesseract
ocr_config = '--oem 3 --psm 6' # Use LSTM engine with uniform block detection
ocr_data = pytesseract.image_to_data(
image_np,
output_type=pytesseract.Output.DICT,
config=ocr_config
)
# Extract words and bounding boxes
words = []
boxes = []
confidences = []
for i in range(len(ocr_data['text'])):
if ocr_data['text'][i].strip() and ocr_data['conf'][i] > 30: # Filter by confidence
words.append(ocr_data['text'][i])
boxes.append([
ocr_data['left'][i],
ocr_data['top'][i],
ocr_data['left'][i] + ocr_data['width'][i],
ocr_data['top'][i] + ocr_data['height'][i]
])
confidences.append(ocr_data['conf'][i])
# Join words to form complete text
text = ' '.join(words)
# Process with LayoutLMv3 for structure understanding (if text found)
structured_text = text
if words and len(words) < 400: # Limit for performance
try:
# Prepare inputs for LayoutLMv3
encoding = layoutlm_processor(
image,
words[:400], # Limit words
boxes=boxes[:400],
return_tensors="pt",
truncation=True,
padding="max_length",
max_length=512
)
# Move to device and run inference
encoding = {k: v.to(device) for k, v in encoding.items()}
with torch.no_grad():
outputs = layoutlm_model(**encoding)
# Get predictions
predictions = outputs.logits.argmax(-1).squeeze().tolist()
if isinstance(predictions, int):
predictions = [predictions]
# Structure text based on layout
structured_text = self._structure_text(words[:len(predictions)], boxes[:len(predictions)])
except Exception as e:
print(f"LayoutLM processing skipped: {e}")
structured_text = self._simple_structure_text(words, boxes)
else:
structured_text = self._simple_structure_text(words, boxes)
return {
'raw_text': text,
'words': words,
'boxes': boxes,
'structured_text': structured_text,
'num_words': len(words),
'avg_confidence': sum(confidences) / len(confidences) if confidences else 0
}
except Exception as e:
print(f"Error extracting text: {e}")
return {
'raw_text': "",
'words': [],
'boxes': [],
'structured_text': "",
'num_words': 0,
'avg_confidence': 0
}
def _simple_structure_text(self, words: List[str], boxes: List[List[int]]) -> str:
"""
Simple text structuring based on spatial layout
Groups words into lines based on vertical position
"""
if not words:
return ""
# Group words by lines
lines = []
current_line = []
last_y = None
for word, box in zip(words, boxes):
y_pos = box[1] # Top position
if last_y is None or abs(y_pos - last_y) < 15: # Same line threshold
current_line.append(word)
else:
if current_line:
lines.append(' '.join(current_line))
current_line = [word]
last_y = y_pos
if current_line:
lines.append(' '.join(current_line))
return '\n'.join(lines)
def _structure_text(self, words: List[str], boxes: List[List[int]]) -> str:
"""Enhanced text structuring with better line detection"""
return self._simple_structure_text(words, boxes)
def process_document(self, file_path: str) -> str:
"""
Process any document type (PDF or image) and extract text
Args:
file_path: Path to the document file
Returns:
Status message indicating success or failure
"""
# Reset state
self.extracted_text = ""
self.page_contents = []
self.document_metadata = {
'filename': os.path.basename(file_path),
'pages': 0,
'total_words': 0
}
# Check cache
file_hash = self._get_file_hash(file_path)
if file_hash in self.processing_cache:
cached_data = self.processing_cache[file_hash]
self.extracted_text = cached_data['text']
self.page_contents = cached_data['pages']
self.document_metadata = cached_data['metadata']
return f"βœ… Loaded from cache: {self.document_metadata['filename']}\n" \
f"πŸ“„ Pages: {self.document_metadata['pages']}\n" \
f"πŸ“ Words: {self.document_metadata['total_words']}"
try:
start_time = time.time()
if file_path.lower().endswith('.pdf'):
# Process PDF document
images = self.process_pdf(file_path)
self.document_metadata['pages'] = len(images)
for i, image in enumerate(images):
print(f"Processing page {i+1}/{len(images)}...")
page_data = self.extract_text_from_image(image)
self.page_contents.append(page_data)
self.extracted_text += f"\n\n--- Page {i+1} ---\n\n"
self.extracted_text += page_data['structured_text']
self.document_metadata['total_words'] += page_data['num_words']
else:
# Process single image
image = Image.open(file_path).convert('RGB')
page_data = self.extract_text_from_image(image)
self.page_contents.append(page_data)
self.extracted_text = page_data['structured_text']
self.document_metadata['pages'] = 1
self.document_metadata['total_words'] = page_data['num_words']
# Cache the results
self.processing_cache[file_hash] = {
'text': self.extracted_text,
'pages': self.page_contents,
'metadata': self.document_metadata
}
processing_time = time.time() - start_time
if self.document_metadata['total_words'] == 0:
return f"⚠️ No text found in {self.document_metadata['filename']}. Please ensure the document contains readable text."
return f"βœ… Successfully processed {self.document_metadata['filename']}\n" \
f"πŸ“„ Pages: {self.document_metadata['pages']}\n" \
f"πŸ“ Words extracted: {self.document_metadata['total_words']}\n" \
f"⏱️ Processing time: {processing_time:.1f}s"
except Exception as e:
return f"❌ Error processing document: {str(e)}"
def summarize_document(self) -> str:
"""
Generate a concise summary of the document using T5 model
Returns:
Document summary or error message
"""
if not self.extracted_text:
return "No document has been processed yet. Please upload and process a document first."
try:
start_time = time.time()
# Prepare text for summarization (limit to manage tokens)
text_to_summarize = self.extracted_text[:2048]
# Create prompt for T5
prompt = f"Summarize the following document:\n\n{text_to_summarize}"
# Tokenize input
inputs = t5_tokenizer(
prompt,
return_tensors="pt",
max_length=1024,
truncation=True
).to(device)
# Generate summary
with torch.no_grad():
summary_ids = t5_model.generate(
inputs.input_ids,
max_length=150,
min_length=30,
num_beams=4,
length_penalty=2.0,
early_stopping=True
)
# Decode summary
summary = t5_tokenizer.decode(summary_ids[0], skip_special_tokens=True)
generation_time = time.time() - start_time
return f"{summary}\n\n⏱️ Generated in {generation_time:.1f}s"
except Exception as e:
return f"Error generating summary: {str(e)}"
def answer_question(self, question: str) -> str:
"""
Answer questions about the document using T5 model
Args:
question: User's question about the document
Returns:
Answer to the question
"""
if not self.extracted_text:
return "Please upload and process a document first."
if not question.strip():
return "Please enter a question."
try:
start_time = time.time()
# Prepare context and question
context = self.extracted_text[:1536] # Limit context
# Format prompt for T5
prompt = f"Answer the question based on the context.\n\nContext: {context}\n\nQuestion: {question}\n\nAnswer:"
# Tokenize
inputs = t5_tokenizer(
prompt,
return_tensors="pt",
max_length=1024,
truncation=True
).to(device)
# Generate answer
with torch.no_grad():
answer_ids = t5_model.generate(
inputs.input_ids,
max_length=100,
min_length=5,
num_beams=3,
temperature=0.7,
do_sample=True,
top_p=0.9
)
# Decode answer
answer = t5_tokenizer.decode(answer_ids[0], skip_special_tokens=True)
generation_time = time.time() - start_time
return f"{answer}\n\n⏱️ Generated in {generation_time:.1f}s"
except Exception as e:
return f"Error answering question: {str(e)}"
def extract_key_information(self) -> Dict[str, List[str]]:
"""
Extract key entities from the document using regex patterns
Returns:
Dictionary of extracted entities organized by type
"""
if not self.extracted_text:
return {"message": ["No document has been processed yet."]}
try:
entities = {
'dates': [],
'emails': [],
'phone_numbers': [],
'monetary_amounts': [],
'percentages': [],
'urls': []
}
# Date extraction patterns
date_patterns = [
r'\b\d{1,2}[/-]\d{1,2}[/-]\d{2,4}\b',
r'\b\d{4}[/-]\d{1,2}[/-]\d{1,2}\b',
r'\b(?:January|February|March|April|May|June|July|August|September|October|November|December|Jan|Feb|Mar|Apr|May|Jun|Jul|Aug|Sep|Oct|Nov|Dec)\s+\d{1,2},?\s+\d{4}\b',
r'\b\d{1,2}\s+(?:January|February|March|April|May|June|July|August|September|October|November|December|Jan|Feb|Mar|Apr|May|Jun|Jul|Aug|Sep|Oct|Nov|Dec)\s+\d{4}\b'
]
for pattern in date_patterns:
matches = re.findall(pattern, self.extracted_text, re.IGNORECASE)
entities['dates'].extend(matches)
# Email extraction
email_pattern = r'\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b'
entities['emails'] = re.findall(email_pattern, self.extracted_text)
# Phone number extraction (various formats)
phone_patterns = [
r'\b\+?1?\s*\(?([0-9]{3})\)?[-.\s]?([0-9]{3})[-.\s]?([0-9]{4})\b',
r'\b\d{3}[-.\s]\d{3}[-.\s]\d{4}\b'
]
for pattern in phone_patterns:
matches = re.findall(pattern, self.extracted_text)
if isinstance(matches[0], tuple) if matches else False:
entities['phone_numbers'].extend(['-'.join(match) for match in matches])
else:
entities['phone_numbers'].extend(matches)
# Monetary amount extraction
money_patterns = [
r'\$\s*[\d,]+\.?\d*',
r'USD\s*[\d,]+\.?\d*',
r'\b\d{1,3}(?:,\d{3})*(?:\.\d{2})?\s*(?:dollars?|USD)\b'
]
for pattern in money_patterns:
matches = re.findall(pattern, self.extracted_text, re.IGNORECASE)
entities['monetary_amounts'].extend(matches)
# Percentage extraction
percent_pattern = r'\b\d+\.?\d*\s*%'
entities['percentages'] = re.findall(percent_pattern, self.extracted_text)
# URL extraction
url_pattern = r'https?://(?:www\.)?[-a-zA-Z0-9@:%._\+~#=]{1,256}\.[a-zA-Z0-9()]{1,6}\b(?:[-a-zA-Z0-9()@:%_\+.~#?&/=]*)'
entities['urls'] = re.findall(url_pattern, self.extracted_text)
# Clean up and deduplicate
for key in entities:
# Remove duplicates and limit to 10 items
unique_items = list(dict.fromkeys(entities[key])) # Preserves order
entities[key] = unique_items[:10]
# Remove empty categories
entities = {k: v for k, v in entities.items() if v}
if not entities:
entities = {"info": ["No specific entities found. The document may need better quality or contain different types of information."]}
return entities
except Exception as e:
return {"error": [f"Error extracting information: {str(e)}"]}
# Initialize global processor
processor = DocumentProcessor()
# Gradio interface handlers
def process_document_handler(file):
"""Handle document upload and processing"""
if file is None:
return "Please upload a document.", "", {}
# Process the document
status = processor.process_document(file)
# Get text preview
text_preview = processor.extracted_text[:1000] + "..." if len(processor.extracted_text) > 1000 else processor.extracted_text
# Extract key information
key_info = processor.extract_key_information()
return status, text_preview, key_info
def summarize_handler():
"""Handle document summarization request"""
return processor.summarize_document()
def qa_handler(question):
"""Handle question answering request"""
if not question:
return "Please enter a question."
return processor.answer_question(question)
def create_interface():
"""
Create the Gradio interface for the document intelligence system
"""
with gr.Blocks(title="Multi-Modal Document Intelligence System", theme=gr.themes.Soft()) as interface:
# Header
gr.Markdown("""
# 🧠 Multi-Modal Document Intelligence System
**Upload any document (PDF or image) and unlock its insights with AI!**
This advanced system combines:
- πŸ“„ **LayoutLMv3** for understanding document structure and layout
- πŸ€– **Flan-T5** for intelligent summarization and question answering
- πŸ” **OCR Technology** for accurate text extraction from any document
### ✨ Features
- Upload PDFs or images (JPG, PNG, etc.)
- Automatic text extraction with layout understanding
- Intelligent document summarization
- Natural language Q&A about your documents
- Key information extraction (dates, emails, amounts, etc.)
""")
# Main interface layout
with gr.Row():
# Left column - Upload and processing
with gr.Column(scale=1):
file_input = gr.File(
label="πŸ“ Upload Document",
file_types=[".pdf", ".png", ".jpg", ".jpeg", ".bmp", ".tiff"],
type="filepath"
)
process_btn = gr.Button("πŸ”„ Process Document", variant="primary", size="lg")
status_output = gr.Textbox(
label="πŸ“Š Processing Status",
lines=4,
interactive=False
)
gr.Markdown("### πŸ”‘ Key Information Extracted")
key_info_output = gr.JSON(label="Extracted Entities", elem_id="key_info")
# Right column - Results and interaction
with gr.Column(scale=2):
text_preview = gr.Textbox(
label="πŸ“„ Document Text Preview",
lines=10,
max_lines=15,
interactive=False
)
with gr.Tab("πŸ“ Summary"):
summary_btn = gr.Button("Generate Summary", variant="secondary")
summary_output = gr.Textbox(
label="Document Summary",
lines=8,
interactive=False
)
with gr.Tab("❓ Q&A"):
question_input = gr.Textbox(
label="Ask a question about the document",
placeholder="e.g., What are the main points? What dates are mentioned? What is the total amount?",
lines=2
)
qa_btn = gr.Button("Get Answer", variant="secondary")
answer_output = gr.Textbox(
label="Answer",
lines=6,
interactive=False
)
# Example questions
gr.Markdown("### πŸ“š Example Questions")
gr.Examples(
examples=[
"What is the main topic of this document?",
"What dates are mentioned?",
"What is the total amount due?",
"Who are the key people mentioned?",
"What are the main findings?",
"Summarize the key points."
],
inputs=question_input
)
# Footer with instructions
gr.Markdown("""
---
### 🎯 How to Use
1. **Upload** a PDF or image document
2. **Process** the document to extract text
3. **Review** the extracted text and key information
4. **Generate** a summary or ask questions
### πŸ’‘ Tips for Best Results
- Use clear, high-quality documents
- For images, ensure good lighting and contrast
- The system works with multiple languages
- Processing time depends on document size and complexity
---
πŸ‘¨β€πŸ’» **Created by Spencer Purdy** | Computer Science @ Auburn University
[GitHub](https://github.com/spencercpurdy) | [LinkedIn](https://linkedin.com/in/spencerpurdy) | [Hugging Face](https://huggingface.co/spencercpurdy)
""")
# Connect event handlers
process_btn.click(
fn=process_document_handler,
inputs=file_input,
outputs=[status_output, text_preview, key_info_output]
)
summary_btn.click(
fn=summarize_handler,
inputs=[],
outputs=summary_output
)
qa_btn.click(
fn=qa_handler,
inputs=question_input,
outputs=answer_output
)
# Allow Enter key to submit questions
question_input.submit(
fn=qa_handler,
inputs=question_input,
outputs=answer_output
)
return interface
# Main execution
if __name__ == "__main__":
print("Starting Multi-Modal Document Intelligence System...")
# Create and launch the interface
interface = create_interface()
# Launch with public link
interface.launch(
debug=True,
share=True,
server_name="0.0.0.0",
server_port=7860
)