DocUnderstanding Pipeline

class DocUnderstanding:
    """
    Document understanding through vision-language models.

    Uses VLM (Vision-Language Model) for visual question answering,
    document parsing, and intelligent content extraction.
    """
    def __init__(
        self,
        doc_understanding_model_name: str = None,
        doc_understanding_model_dir: str = None,
        doc_understanding_batch_size: int = None,
        paddlex_config: str = None,
        device: str = None,
        use_hpi: bool = None,
        **kwargs
    ):
        """
        Initialize DocUnderstanding pipeline.

        Args:
            doc_understanding_model_name (str, optional): VLM model name
                Default: 'PP-DocBee2-3B'
            doc_understanding_model_dir (str, optional): Model directory
            doc_understanding_batch_size (int, optional): Batch size for inference
            device (str, optional): Device for inference ('cpu', 'gpu')
            use_hpi (bool, optional): Use high-performance inference
            paddlex_config (str or dict, optional): Configuration file or dict
        """

def predict(
    self,
    input,
    **kwargs
) -> list:
    """
    Understand document content and answer questions.

    Args:
        input: Can be:
            - Image/PDF path
            - Numpy array or PIL Image
            - Directory path
            - List of paths/images
            - Dict with 'image' and 'prompt' keys for VQA

    Returns:
        list: Understanding results with answers and analysis
    """

def predict_iter(
    self,
    input,
    **kwargs
):
    """
    Iterate over understanding results for memory efficiency.

    Args:
        input: Same as predict()

    Yields:
        dict: Understanding result for each input
    """

def close(self) -> None:
    """Close the pipeline and free resources."""

def export_paddlex_config_to_yaml(self, yaml_path: str) -> None:
    """Export configuration to YAML."""

from paddleocr import DocUnderstanding

# Initialize pipeline
model = DocUnderstanding()

# Ask questions about an image
vqa_input = {
    'image': 'invoice.jpg',
    'prompt': 'What is the total amount on this invoice?'
}

result = model.predict(vqa_input)
print(result[0]['answer'])

model.close()

from paddleocr import DocUnderstanding

model = DocUnderstanding()

# Prepare multiple questions
questions = [
    {'image': 'contract.pdf', 'prompt': 'Who are the parties in this contract?'},
    {'image': 'contract.pdf', 'prompt': 'What is the contract date?'},
    {'image': 'contract.pdf', 'prompt': 'What is the contract value?'}
]

# Get answers
results = model.predict(questions)

for q, r in zip(questions, results):
    print(f"Q: {q['prompt']}")
    print(f"A: {r['answer']}\n")

model.close()

from paddleocr import DocUnderstanding

model = DocUnderstanding()

# General document understanding
result = model.predict('diagram.png')

# Get document description
print(result[0]['description'])

model.close()

from paddleocr import DocUnderstanding

model = DocUnderstanding(
    doc_understanding_batch_size=4
)

# Process multiple documents with questions
inputs = [
    {'image': 'receipt1.jpg', 'prompt': 'Extract the date and total'},
    {'image': 'receipt2.jpg', 'prompt': 'Extract the date and total'},
    {'image': 'receipt3.jpg', 'prompt': 'Extract the date and total'},
    {'image': 'receipt4.jpg', 'prompt': 'Extract the date and total'}
]

results = model.predict(inputs)

for i, result in enumerate(results):
    print(f"Receipt {i+1}: {result['answer']}")

model.close()

from paddleocr import DocUnderstanding

model = DocUnderstanding()

# Extract structured information
vqa_input = {
    'image': 'resume.pdf',
    'prompt': '''Extract the following information in JSON format:
    - Name
    - Email
    - Phone
    - Work Experience (company, role, duration)
    - Education
    '''
}

result = model.predict(vqa_input)
print(result[0]['answer'])

model.close()

from paddleocr import DocUnderstanding

model = DocUnderstanding()

# Ask about table contents
vqa_input = {
    'image': 'financial_table.png',
    'prompt': 'What are the top 3 performing products by revenue?'
}

result = model.predict(vqa_input)
print(result[0]['answer'])

model.close()

from paddleocr import DocUnderstanding

model = DocUnderstanding()

# Process large dataset efficiently
questions = [
    {'image': f'doc_{i}.pdf', 'prompt': 'Summarize the main points'}
    for i in range(100)
]

# Use iterator to avoid loading all results in memory
for result in model.predict_iter(questions):
    # Process one at a time
    with open('summaries.txt', 'a') as f:
        f.write(f"{result['input_path']}: {result['answer']}\n")

model.close()

# Classify document types
result = model.predict({
    'image': 'document.pdf',
    'prompt': 'What type of document is this? (invoice, contract, receipt, etc.)'
})

# Verify specific information
result = model.predict({
    'image': 'id_card.jpg',
    'prompt': 'Is this person over 18 years old?'
})

# Analyze visualizations
result = model.predict({
    'image': 'sales_chart.png',
    'prompt': 'What trend does this chart show?'
})

# Handle multiple languages (PP-DocBee2-3B supports many languages)
result = model.predict({
    'image': 'chinese_form.jpg',
    'prompt': 'What is the applicant name? (答案用中文)'
})

[
    {
        "input_path": "path/to/document.jpg",
        "prompt": "What is the total amount?",
        "answer": "$1,234.56",
        "confidence": 0.95,
        "model": "PP-DocBee2-3B"
    }
]

[
    {
        "input_path": "path/to/document.jpg",
        "description": "This image shows a sales invoice with...",
        "detected_elements": ["text", "table", "logo"],
        "confidence": 0.92,
        "model": "PP-DocBee2-3B"
    }
]