The case for agentic OCR's iterative approach to parsing documents

LlamaIndex argues that traditional one-pass OCR can flatten tables, lose charts and scramble multi-column layouts. It describes agentic OCR as a loop that follows document layout, routes difficult elements to suitable models, and checks and corrects output over multiple passes. The company says its head of open source, Logan Markewich, wrote a breakdown that also covers where the approach still struggles.