RostaingOCR

Production-Grade Layout-Aware OCR for LLMs and RAG Systems

RostaingOCR is a high-performance Python library designed to extract text from PDFs, Scanned PDFs, and images while preserving complex layouts. Unlike standard OCR tools that output a "soup" of words, this library uses Deep Learning and geometric reconstruction to maintain tables, columns, and document structure.

It is specifically optimized for Retrieval-Augmented Generation (RAG) pipelines where maintaining the visual structure of data (like invoice tables) is critical for LLM comprehension.

Key Features

• mj-layout-aware: Uses geometric clustering to reconstruct tables and columns. Data stays on the correct line, visually aligned.
• 🧹 Noise Filtering: Automatically detects and removes low-confidence text such as messy handwriting, signatures, and stamps to keep the output clean.
• ⚡ Local Processing: Runs 100% locally (CPU or GPU). No external APIs, no data leaving your server.
• 📄 Universal Input: Handles PDFs (digital & scanned) and common image formats via a robust Base64 architecture.
• 🔒 Privacy Focused: Temporary files are handled securely and deleted immediately after extraction.

Installation

pip install rostaing-ocr

(Note: The first run will automatically download the necessary OCR models ~300MB)

Usage

1. Basic Usage (Default Behavior)

By default, the extractor prints the result to the console and saves it to output.txt.

from rostaing_ocr import ocr_extractor

# This immediately runs the extraction using DocTR
extractor = ocr_extractor("documents/invoice.pdf")

# The extracted text is now in 'output.txt'
print(extractor) # Prints status summary (Time taken, pages processed)

2. Custom Output File

You can specify a different filename. The file will be created or overwritten automatically.

from rostaing_ocr import ocr_extractor

extractor = ocr_extractor(
    "data/report.png",
    output_file="results/report_analysis.txt"
)

3. Silent Mode (Background Processing)

Useful for batch processing or server backends where you don't want console logs.

from rostaing_ocr import ocr_extractor

extractor = ocr_extractor(
    "financial_statement.pdf",
    print_to_console=False,
    save_file=True
)

4. Direct Integration (RAG Pipelines)

Access the text variable directly without reading the file.

from rostaing_ocr import ocr_extractor

extractor = ocr_extractor("scan.jpg", print_to_console=False)

if extractor.status == "Success":
    clean_text = extractor.extracted_text
    # Send 'clean_text' to GPT, Mistral, Gemini, Claude, Grok, Groq, Llama... or your Vector DB

How It Works (Architecture)

1. Input Normalization: Converts PDF pages or Images into High-Res Base64 streams.
2. Deep Learning Inference: DBNet for detection + CRNN for recognition.
3. Noise Filtering: Scans confidence scores. Text with low confidence (e.g., < 0.4), such as signatures or stamps, is discarded.
4. Geometric Reconstruction:
   • Flattens the document hierarchy.
   • Clusters words into visual lines based on Y-axis alignment.
   • Calculates horizontal gaps to insert dynamic spacing (tabs vs spaces) to simulate columns.
5. Output: Returns a clean, structured string that looks like the original document.

License

MIT License

Useful Links

• Author's LinkedIn
• Author's YouTube Channel
• GitHub Repository
• PyPI Project Page