LLM-Aided OCR is an open-source pipeline that upgrades the output of the classic Tesseract OCR engine by passing it through a large language model. The LLM stage corrects recognition errors, intelligently chunks long documents, and reformats scanned PDFs into clean Markdown that is far more useful for search, archiving and downstream AI applications. The GitHub project has about 3,000 stars and is actively maintained, with updates continuing into 2026.
Raw Tesseract output is famously messy: broken words, mangled tables and lost structure make scanned documents hard to use. LLM-Aided OCR keeps Tesseract for what it does cheaply โ reading characters โ and delegates cleanup to an LLM, which is very good at reconstructing intended text from noisy input. You can run the language model locally, for example through Ollama with an open-weight model, or call a hosted API for maximum quality. The pipeline handles the awkward parts of real-world documents, splitting them into sensible chunks so long scans do not blow past context limits, and it can even generate study quizzes from processed documents. The same technology also powers the author’s hosted FixMyDocuments service for people who prefer a web interface.
The tool is free and open source. If you use a hosted LLM API for the correction stage you pay that provider’s normal usage rates; running a local model makes the pipeline entirely free.
LLM-Aided OCR is for archivists, researchers, legal and operations teams, and developers building retrieval systems who need accurate, well-formatted text out of scanned documents without buying enterprise OCR software.