LLM-Aided OCR is a Python tool that enhances Tesseract OCR output using large language models for error correction, smart text chunking, and markdown formatting of scanned PDFs.
The tool addresses the problem that raw OCR output from Tesseract contains errors and lacks proper formatting. It works by converting PDFs to images, running Tesseract OCR, then splitting the text into chunks and passing each through an LLM for error correction and optional markdown formatting. The system can remove duplicate content and suppress headers and page numbers. It supports both local LLMs via llama_cpp and cloud-based providers including OpenAI and Anthropic, with asynchronous processing and GPU acceleration available for local inference.
This tool suits developers working with scanned documents who need high-quality text extraction and are willing to invest in LLM processing to achieve it. It works best for documents where OCR errors significantly impact readability, such as historical letters or technical papers. The project includes example outputs showing before-and-after transformations. Anyone adopting it should be prepared to configure LLM access, either by running a local model or providing API credentials for a cloud provider.
Development activity shows consistent engagement with the codebase through regular commits and active issue management. The project maintains detailed documentation of its technical implementation, including specific function names and processing pipeline details. Error handling and retry logic are implemented for API-based LLM calls, indicating attention to reliability in production scenarios. The inclusion of quality assessment features and comprehensive logging suggests the maintainers prioritize observability and debugging capabilities.
