Solicitações para OCR e análise de documentos

LLMs como leitores de documentos

Instrução básica de extração de texto

import anthropic, base64

client = anthropic.Anthropic(api_key='YOUR_API_KEY')

def extract_text(image_path, extraction_prompt):
    with open(image_path, 'rb') as f:
        img_b64 = base64.standard_b64encode(f.read()).decode('utf-8')

    r = client.messages.create(
        model='claude-opus-4-5', max_tokens=1000,
        messages=[{'role': 'user', 'content': [
            {'type': 'image', 'source': {'type': 'base64', 'media_type': 'image/jpeg', 'data': img_b64}},
            {'type': 'text', 'text': extraction_prompt}
        ]}]
    )
    return r.content[0].text

# Basic extraction
basic_prompt = 'Extract all text from this document image exactly as it appears. Preserve line breaks.'

# Structure-preserving extraction
structured_prompt = 'Extract all text from this document image. Preserve: paragraph structure, line breaks, and any visible formatting. Do not add any text not present in the image.'

print('Text extraction functions defined.')

Preservação da estrutura de tabelas

table_prompt = '''
Extract all text from this document image.
If the document contains any tables, preserve the table structure using markdown table format:
| Column 1 | Column 2 | Column 3 |
|----------|----------|----------|
| Value    | Value    | Value    |

For any text outside tables, use plain text preserving paragraph structure.
Do not invent or infer any data not visible in the image.
'''

# For structured output:
table_json_prompt = '''
Extract the table from this image.
Return JSON:
{
  "headers": ["column name"],
  "rows": [["cell value", "cell value"]],
  "caption": "table caption if present or null"
}
If a cell is empty or illegible, use null.
'''

print('Table extraction prompts defined.')

Discriminação dos itens do recibo

import json

receipt_prompt = '''
Extract all information from this receipt image.
Return JSON:
{
  "merchant": {
    "name": str,
    "address": str or null,
    "phone": str or null
  },
  "transaction": {
    "date": "YYYY-MM-DD or as written",
    "time": "HH:MM or as written or null",
    "receipt_number": str or null,
    "payment_method": str or null
  },
  "items": [
    {"description": str, "quantity": number or null, "unit_price": number or null, "total": number}
  ],
  "subtotal": number or null,
  "tax": number or null,
  "tip": number or null,
  "total": number,
  "currency": "3-letter ISO code"
}
For any field not visible, use null. For numbers, use numeric type (not string).
'''

def extract_receipt(image_path):
    text = extract_text(image_path, receipt_prompt)
    return json.loads(text)

print('Receipt extraction function defined.')

Transcrição de anotações manuscritas

handwriting_prompt = '''
Transcribe the handwritten text in this image as accurately as possible.

Handling rules:
- If a word is illegible, write [ILLEGIBLE]
- If a word is partially legible, write [PARTIAL: best_guess]
- If text is crossed out, include it with strikethrough notation: ~~crossed out text~~
- Preserve line breaks as they appear
- If there are arrows, circles, or annotations, note them in brackets: [arrow pointing right]
- Do not correct spelling or grammar

After transcription, estimate overall legibility: high (>90% readable) | medium (70-90%) | low (<70%)

Format:
TRANSCRIPTION:
[transcribed text here]

LEGIBILITY: [rating]
'''

print(handwriting_prompt)

Extração de campos de formulário

form_prompt = '''
Extract all form fields and their values from this document image.

For each field:
- Field label: the printed label (e.g., "First Name:", "Date of Birth:")
- Field value: the filled-in value (handwritten or typed)
- Filled: whether the field has been filled in (true/false)

Return JSON:
{
  "form_title": str or null,
  "fields": [
    {
      "label": str,
      "value": str or null,
      "filled": true | false
    }
  ],
  "signature_present": true | false,
  "date_signed": str or null
}

If the value is illegible, use "[ILLEGIBLE]".
If the field is blank, value should be null and filled should be false.
'''

print('Form field extraction prompt defined.')
print('Handles: printed forms, questionnaires, applications.')

Classificação do documento antes da extração

import json

DOC_SCHEMAS = {
    'receipt': receipt_prompt,
    'form': form_prompt,
    'table': table_json_prompt,
    'letter': 'Extract all text preserving paragraph structure. Identify: sender, recipient, date, subject, body.',
    'label': 'Extract all text from this label. Include: product name, ingredients/contents, weight, expiry date, barcode numbers.'
}

def classify_and_extract(image_path):
    # Step 1: Classify document type
    classify_prompt = 'What type of document is this? Return JSON: {"type": "receipt|form|table|letter|label|other", "confidence": "high|medium|low"}'
    classification_text = extract_text(image_path, classify_prompt)
    doc_type = json.loads(classification_text)['type']

    # Step 2: Apply correct schema
    schema = DOC_SCHEMAS.get(doc_type, 'Extract all visible text from this document.')
    extracted = extract_text(image_path, schema)

    return {'type': doc_type, 'data': extracted}

print('Document classify-then-extract pipeline defined.')

Tratamento de imagens de baixa qualidade

low_quality_prompt = '''
Extract text from this document image. The image may be low quality, blurry, or poorly lit.

Extraction guidelines:
- Extract all text you can read with reasonable confidence
- For unclear sections, use [UNCLEAR] as a placeholder
- For completely unreadable sections, use [UNREADABLE: approximately N words]
- Do not guess or hallucinate words you cannot see clearly
- Note image quality issues at the end: "Image quality: [good/fair/poor]. Issues: [description]"

Be conservative — it is better to mark something as unclear than to guess incorrectly.
'''

print(low_quality_prompt)
print('\nConservative approach: unclear beats hallucinated.')

Resumo de documentos com várias páginas

def extract_multi_page_document(image_paths):
    # Step 1: Extract text from each page
    page_texts = []
    for i, path in enumerate(image_paths):
        page_text = extract_text(path, f'Extract all text from page {i+1} of this document. Preserve structure.')
        page_texts.append(f'=== PAGE {i+1} ===\n{page_text}')

    full_text = '\n\n'.join(page_texts)

    # Step 2: Synthesize summary and key information
    r = client.messages.create(
        model='claude-opus-4-5', max_tokens=500,
        messages=[{'role': 'user', 'content': f'''
Here is the extracted text from a {len(image_paths)}-page document:\n\n{full_text}\n\n
Provide:
1. Document type and title
2. 3-sentence summary
3. Key data points extracted
Return JSON: {{"type": str, "title": str, "summary": str, "key_data": [str]}}
'''}]
    )
    return json.loads(r.content[0].text)

print('Multi-page document pipeline defined.')

Validação após a extração por OCR

import re
from datetime import datetime

def validate_receipt_extraction(data):
    errors = []

    # Validate total is present and numeric
    if data.get('total') is None:
        errors.append('total is missing')
    elif not isinstance(data['total'], (int, float)):
        errors.append(f'total is not numeric: {data["total"]}')

    # Validate date format
    if data.get('transaction', {}).get('date'):
        date_str = data['transaction']['date']
        try:
            datetime.strptime(date_str, '%Y-%m-%d')
        except ValueError:
            errors.append(f'date format invalid: {date_str}')

    # Validate line items total approximately equals subtotal
    if data.get('items') and data.get('subtotal'):
        items_total = sum(item.get('total', 0) for item in data['items'] if item.get('total'))
        if abs(items_total - data['subtotal']) > 0.05:
            errors.append(f'Items total {items_total} does not match subtotal {data["subtotal"]}')

    return errors

print('Receipt validation function defined.')

Extração de dados estruturados de gráficos

chart_prompt = '''
Extract the data from this chart or graph image.

Identify:
1. Chart type (bar, line, pie, scatter, table)
2. Title and axis labels
3. All data series names
4. All data points with their labels/values
5. Any notable trend or pattern

Return JSON:
{
  "chart_type": str,
  "title": str or null,
  "x_axis_label": str or null,
  "y_axis_label": str or null,
  "data_series": [
    {"name": str, "values": [{"label": str, "value": number}]}
  ],
  "key_insight": str
}

If exact values are not readable, provide best estimates with a note.
'''

print(chart_prompt)

Verificação rápida

OCR e análise de documentos — Principais conclusões