This project leverages Vision-Language Models (VLMs) to parse and linearize complex PDF documents, converting unstructured content (such as multi-column text, tables, embedded images, mixed font styles, and layouts) into continuous, structured text representations. It supports the full process of distributed multi-node parsing of millions of PDF documents, enabling the construction of high-quality datasets for Large Language Models (LLMs).