Surya — Document OCR for 90+ Languages

What it is

Surya is an open-source document OCR toolkit that performs text recognition in over 90 languages. It goes beyond basic character recognition by providing layout analysis, table detection, reading order extraction, and LaTeX OCR. The project has accumulated over 19,000 GitHub stars and benchmarks favorably against cloud OCR services.

Surya is built for developers and researchers who need to extract structured text from scanned documents, PDFs, or images. It runs locally, which matters for teams handling sensitive documents that cannot be sent to third-party cloud APIs.

How it saves time or tokens

Without Surya, extracting text from complex documents typically requires a cloud OCR service (Google Cloud Vision, AWS Textract) or manual preprocessing pipelines. Surya consolidates OCR, layout analysis, and table detection into a single local toolkit.

For AI pipelines, Surya's structured output means cleaner text going into LLMs. When you feed an LLM poorly extracted text with broken tables and scrambled reading order, you waste tokens on confused context. Surya's layout-aware extraction preserves document structure, reducing downstream token waste.

How to use

1. Install Surya:

pip install surya-ocr

1. Run OCR on a document image:

surya_ocr image.png

1. For specific tasks, use the dedicated commands:

# Detect text lines
surya_detect image.png

# Analyze layout (tables, headers, images)
surya_layout image.png

# Table recognition
surya_table image.png

Example

Using Surya in a Python script to extract text and export results:

from surya.ocr import run_ocr
from surya.model.detection.model import load_det_model
from surya.model.recognition.model import load_rec_model
from PIL import Image

image = Image.open('invoice.png')
det_model = load_det_model()
rec_model = load_rec_model()

results = run_ocr(
    [image],
    det_model=det_model,
    rec_model=rec_model,
    langs=['en']
)

for line in results[0].text_lines:
    print(line.text)

This extracts all text lines from an invoice image with English language detection.

Related on TokRepo

• Document processing tools -- more tools for document parsing and extraction
• RAG tools -- combine OCR output with retrieval-augmented generation

Common pitfalls

• Surya's models require GPU for reasonable speed on large documents. CPU inference works but is significantly slower for batch processing.
• Language detection is not automatic. You need to specify the expected languages via the langs parameter for best accuracy.
• Table detection and OCR are separate steps. Running surya_table identifies table regions, but you still need surya_ocr to extract the cell text.