Pulpie
A Python library that extracts the main content from raw HTML using small encoder models, stripping navigation, ads and boilerplate for feeding cleaned pages to language models or search pipelines.
Pulpie extracts the main content from raw HTML pages, removing navigation, advertisements, sidebars and footers. It uses small encoder models that label every block of a page in a single forward pass, rather than the slower autoregressive extractors it is compared against, and returns the result as clean Markdown or HTML along with counts of kept and dropped blocks.
Three model sizes share a tokenizer and architecture, with the larger model acting as a teacher for the smaller, distilled ones. A Pipeline class overlaps CPU preprocessing with GPU inference and balances work across one or more GPUs for bulk extraction. Models download from Hugging Face on first use, and the library installs from PyPI with an optional extra for Markdown output.