quivrhq/megaparse

MegaParse is a Python library that converts documents into text in a format suited for use with AI language models. The focus is on preserving as much information as possible during conversion, so that when the resulting text is fed to an AI, nothing important has been dropped or garbled. It handles PDF files, Word documents (Docx), and PowerPoint presentations. Beyond basic text extraction, it also captures tables, tables of contents, headers, footers, and images embedded in those files. The library is installable via pip and requires Python 3.11 or newer. There are two main ways to use it. The standard parser works with a library called LangChain and connects to an OpenAI or Anthropic API key. A second mode called MegaParse Vision uses multimodal AI models (such as GPT-4o or Claude 3.5 and later) to visually interpret pages rather than parsing the document structure directly. The vision approach scored higher on the project's own benchmark, achieving a similarity ratio of 0.87 compared to 0.77 for the next-best alternative tested. It can also run as a local API server. Running one make command at the project root starts a server, and the endpoints are documented at localhost:8000/docs. This lets other tools send documents to MegaParse over HTTP instead of importing it as a Python library directly. The project is open source. The README lists a few features still in progress, including modular post-processing and structured output support.