Xiaomi-OCR-0

Use this Skill when a task needs OCR, document parsing, KIE, or OCR-VQA. Inference runs on the user's machine with the Hugging Face checkpoint SeerRay-Lab/Xiaomi-OCR-0 and SGLang, vLLM, or the experimental Ollama backend. The MCP service owns prompts, image/PDF preparation, optional layout segmentation, crop processing, concurrency, and output formatting. Call the task tool; do not recreate these steps with separate tools or direct model requests.

This repository's MCP sends inference requests to a loopback HTTP endpoint on the MCP server host; it does not call Mify. Other installed skills or services may share the xiaomi-ocr name. Identify the selected MCP from its tool descriptions and configuration before describing where files are processed. With a remote MCP, image_path and pdf_path refer to that server's filesystem, not the agent's machine.

Set up local inference

Choose the inference backend

Respect an explicit backend/model preference and the user's accuracy requirements. Otherwise: