oomol-lab/pdf-craft

pdf-craft is a Python library for converting PDF files into Markdown or EPUB format. It is built specifically for scanned books, where pages are images rather than searchable text. The library uses a document recognition model called DeepSeek OCR to read text from those scanned pages, handling complex content such as tables, mathematical formulas, and footnotes. The conversion runs entirely on your own machine without sending files to any outside server. To use it, you need a compatible NVIDIA graphics card with CUDA configured, plus Poppler, a tool for parsing PDF files. The library itself is installed via pip, the standard Python package installer. During conversion, pdf-craft analyzes the document's layout: it pulls out the main body text while discarding headers, footers, and other repeated page elements. For EPUB output specifically, it builds a table of contents automatically. Footnotes, embedded images, and other assets attached to footnotes are carried through to the final file intact. Starting with version 1.0.0, the library dropped the large language model it previously used for text correction. The older approach made network calls and introduced delays or occasional failures. The current version relies entirely on the local OCR model, so the process runs faster and works without an internet connection. Users who depended on the LLM correction step can still use the older v0.2.8 release. An online demo lets anyone try the conversion workflow in a browser without installing anything locally. The Python API exposes functions for Markdown and EPUB conversion, each accepting optional parameters for DPI settings, image size limits, language, table rendering format, and formula display style.