MegaParse

Quivr's team built MegaParse as the ingestion layer for LLM applications that cannot afford to lose information in parsing. It converts PDF, DOCX, PPTX, XLSX, CSV, text, and image files into clean LLM-ready output while keeping tables, headers, footers, and tables of contents intact. Two modes cover most needs: the standard parser runs locally on top of system tools, while MegaParseVision routes page images through multimodal models including GPT-4o, GPT-4, and Claude 3.5 or Claude 4 for the hardest layouts, which requires OpenAI or Anthropic API keys. Installation is pip install megaparse on Python 3.11 or newer, with Poppler, Tesseract OCR, and libmagic as system dependencies, and the repository ships a FastAPI server plus Docker files for running it as a service. The project is Apache-2.0 licensed with about 7,400 stars, though commits stopped in February 2025, so treat it as stable rather than evolving. RAG builders use it as a single entry point for heterogeneous document collections instead of stitching together format-specific extractors, and the vision mode stays optional for cost control.