thepipe

Why we included this project

Most document parsers stumble on messy inputs, but thepipe tries a different approach: it lets a vision-language model actually read the page rather than relying on brittle layout heuristics. That means PDFs, web pages, Word files, PowerPoints, Jupyter notebooks, and even video and audio get converted into clean markdown, tables, and structured data you can feed straight into an LLM, embedding model, or vector store. Teams building retrieval pipelines or internal knowledge tools will save real engineering time, because file-type detection, layout analysis, and multimodal extraction all happen in one package, and it also includes chunking helpers plus conversions for LlamaIndex and OpenAI messages. The default install stays light for CPU-only machines and CI, with GPU dependencies as opt-in extras. If your work involves turning awkward documents into usable model inputs, this is worth trying.