Enable agents to read and cite long documents locally
Agents can now find and cite specific parts of hundreds of pages of PDFs or Excel files full of tables without reading them all at once. MinerU is a tool that converts documents into machine-readable Markdown or JSON and acts as a local document library for storing and searching the converted results. Version 4.0 bundles document conversion, library management, and agent reading capabilities into a single workflow, with specific support for how agents read and continue through documents partially.
Changes in Version 4.0: Unifying Conversion and Reading
MinerU 4.0 provides a workflow that integrates a local document library and service tools, going beyond a simple document conversion tool. While previous versions focused mainly on file format conversion, this version enhances the ability to cache and search converted results, allowing agents to continue reading content at the page or block level. It also provides a stable locator for identifying document locations, reducing confusion when agents later reference or cite specific parts. These changes enable agents to access long documents step by step, targeting only the necessary parts rather than processing them all at once.
Four Quality Tiers and Supported Formats
MinerU offers four quality tiers based on document complexity and intended use.
- Selecting a Quality Tier Users can choose one of
flash,basic,standard, oradvanceddepending on the nature of the document and hardware specifications. The default isstandard, andflashmust be explicitly specified when a quick check is needed. - Model Engine Configuration Small models can be configured independently using ONNX or Torch, while VLMs can use llama.cpp, vLLM, or LMDeploy. This provides flexibility to achieve optimal inference speeds across various hardware environments.
Usage in Agent Workflows
MinerU supports the process of agents reading documents through a command-line interface (CLI) and a Python SDK. Agents parse documents using the mineru parse command and use the --json option to receive structured results for determining control flow. When processing long documents, they can continue reading content at the page or block level using the returned locators, without loading the entire file into the context. These features help agents efficiently extract the information needed to answer questions about documents or generate summaries.
Installation and Hardware Requirements
MinerU requires a Python environment of 3.10 or higher but less than 3.15, and can be installed via uv or pipx. By default, small models use ONNX CPU inference, and VLMs use llama.cpp in Vulkan mode, ensuring good compatibility on most devices. If an NVIDIA GPU is available, installing mineru[full] yields higher throughput, which requires an NVIDIA GPU from Volta onwards with at least 16GB of RAM and 8GB or more of VRAM. For Apple Silicon or low-spec GPU environments, the default package or the torch add-on package is suitable.
Privacy and Remote Parsing Policy
MinerU prioritizes privacy, parsing all documents locally by default. To parse documents remotely, the --remote option must be explicitly used, which requires user consent. It is recommended to keep sensitive or confidential documents on local parsing, and local parsing failures do not automatically switch to remote parsing. Even when remote parsing is allowed, documents are not uploaded without explicit user consent. This policy provides a safety mechanism that allows enterprises or individuals to handle sensitive information while utilizing document automation workflows.
