MinerU

MinerU is an open-source project developed by opendatalab, designed to convert complex documents (such as PDFs and Office documents) into LLM-ready Markdown/JSON formats to support agentic workflows. Its version 4.0 integrates document parsing, a local document library, and service tools into a single workflow for document conversion, application integration, and agent reading.

Purpose MinerU is primarily used as a command-line document reader for agents. It parses local documents into readable content, allows agents to continue reading by page or block, and returns stable locators for subsequent reads and citations. It is not a RAG framework, vector database, or chat document application.

Main Capabilities

• Four parsing levels: Flash for fast preview and indexing; Basic for OCR and model-based parsing; Standard / Advanced for more demanding layout and quality requirements.
• Multiple input formats: PDF, images, DOC/DOCX, PPT/PPTX, XLS/XLSX, RTF, ODT/ODS/ODP, EPUB, OFD, HTML/MHTML, and CSV/TSV. DocVortex provides native document parsing.
• Document library and agent reading: discover files, cache results, search content, continue by page or block, and retain stable citation locators.
• Independent model configuration: small models support ONNX or Torch; VLM supports llama.cpp, vLLM, or LMDeploy.
• Unified tools: Python SDK, V1 API, stateless batch conversion, multi-service Router, and a Gradio-based WebUI.
• Structured results and rendering: one document model supports nine rendering targets: Markdown, HTML, LaTeX, DOCX, EPUB, PDF, Structured Content, and Content List V1/V2. Each CLI/API exposes its own export subset.

Usage It can be installed through agent skills: send an installation message to the agent and have the agent install MinerU for you. Manual installation requires Python >=3.10,<3.15, installing the MinerU 4.0 stable version in a virtual environment. The default installation works out of the box: small models run ONNX CPU inference, and VLM runs llaVulkan mode in ma.cpp has good compatibility on the vast majority of devices. If you have an NVIDIA GPU, installing mineru[full]>=4.0 provides the best throughput. On Windows, the GPU version of torch must be installed separately; on macOS, the default installation is already the best-throughput package, and [full] is not needed. Other non-NVIDIA devices require manually installing an accelerated version of torch plus vllm/lmdeploy to obtain the best inference speed.

Common commands include: mineru parse <file> to parse a document; mineru read <locator> to read a locator; mineru find to find by filename; mineru search to search indexed documents; mineru show or mineru list to view status; mineru watch or mineru scan to manage watched folders; mineru forget to forget a file or folder; mineru parse --force or mineru invalidate to force re-parsing; mineru usage --json to view remote API usage.

Applicable scenarios Suitable for reading, checking, summarizing, citing, quoting, or answering questions about local documents; converting document content to Markdown for analysis; OCR of scanned PDFs or images; extracting content from supported document formats; continuing to process long documents using pages/blocks rather than loading the entire file into context; searching indexed documents; retrieving page or block images for visual inspection; preserving stable references using locators.

Privacy and telemetry MinerU is privacy-first and parses locally by default. Documents are uploaded only when a command uses the --remote parameter. Local failures do not silently fall back to remote; remote failures may fall back to local when the request can still be satisfied locally. For sensitive, confidential, legal, medical, financial, personal, or proprietary requests, remain local unless the user explicitly permits remote. MinerU may collect anonymous, locally aggregated usage and diagnostic telemetry, but does not collect document content, extracted text or images, filenames, paths, search queries, prompts, snippets, API keys, usernames, hostnames, raw stack traces, or exact hardware identifiers. Users can inspect and explicitly enable or disable telemetry.

Limitations

• Documents are not automatically uploaded to the official service; remote parsing requires explicit configuration.
• Plain text files (.txt, .md, .markdown, .rst, .tex) are not parsed and are read directly.
• PDFs and images support all four tiers; Office, OpenDocument, EPUB, OFD, HTML/MHTML, and CSV/TSV use local Flash native parsing.
• Docker deployment for non-NVIDIA devices is pending an update.
• Remote parsing requires explicit user consent, and when quality, privacy, download, or configuration choices are involved, a recovery path must be presented and the user must be waited on to choose; there must be no silent fallback to non-MinerU parsers.

The above information is all based on the official text provided; content not provided has not been mentioned for now.