Table Extraction OCR.

Which OCR model actually preserves table structure? We tested Claude, GPT-5.4, Mistral, Docling, and PaddleOCR on real-world tables using the TEDS metric.

import anthropic
import base64
client = anthropic.Anthropic()

with open("table.png", "rb") as f:
    image_data = base64.b64encode(f.read()).decode()

response = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=4096,
    messages=[{
        "role": "user",
        "content": [
            {
                "type": "image",
                "source": {
                    "type": "base64",
                    "media_type": "image/png",
                    "data": image_data
                }
            },
            {
                "type": "text",
                "text": """Extract the table from this image.
Return as markdown table with exact cell values.
Preserve merged cells using colspan notation."""
            }
        ]
    }]
)
print(response.content[0].text)

from paddleocr import PaddleOCR
from paddleocr.ppstructure import PPStructure

# Initialize table recognition
table_engine = PPStructure(table=True, ocr=True)

# Process image
result = table_engine("table.png")

# Extract table structure
for item in result:
    if item['type'] == 'table':
        html_table = item['res']['html']
        print(html_table)

        # Convert to markdown if needed
        # from markdownify import markdownify
        # print(markdownify(html_table))

from docling import DocumentConverter
from docling.datamodel.base_models import InputFormat

# Initialize converter
converter = DocumentConverter()

# Process document
result = converter.convert("document.pdf")

# Extract tables from all pages
for page in result.document.pages:
    for table in page.tables:
        # Get as markdown
        print(table.export_to_markdown())

        # Or as pandas DataFrame
        # df = table.export_to_dataframe()
        # print(df)

from mineru import MinerU

# Initialize with table extraction focus
miner = MinerU(
    enable_table=True,
    table_format="markdown"  # or "html", "latex"
)

# Extract from PDF
result = miner.extract("research_paper.pdf")

# Process tables
for page in result:
    for block in page.blocks:
        if block.category == 'table':
            print(block.to_markdown())
            # LaTeX output for scientific docs
            # print(block.to_latex())

Bottom line: Open source models (PaddleOCR-VL, MinerU) offer 50-60× cost savings over API models with comparable or better accuracy. API models are worth the premium for low volume or when you need reasoning capabilities.