dsh-pseudo-vision

In DeepSeek Harness, text-only models often lack visual understanding capabilities, limiting their application in image processing tasks. The dsh-pseudo-vision plugin solves this problem through an innovative local tool layer, automatically decomposing images into OCR text, color statistics, pixel scans, and metadata, enabling any text-only model to 'see' images. The plugin takes over routing, generating sibling routes with image support for text-only models, and seamlessly converts images to text injection during LLM dispatch, without external vision APIs, ensuring data privacy and low latency.

Its core capability lies in a robust local processing pipeline: using tesseract.js for high-precision OCR (supporting Chinese and English recognition), combined with sharp for color analysis and pixel scanning, all executed locally. This effectively enhances the performance of text-only models in scenarios like document analysis and image description, while avoiding the costs and risks of relying on external services. The plugin works out of the box after installation, requiring no additional configuration and integrating smoothly with existing workflows.

This plugin is particularly suitable for developers, researchers, or everyday users of DeepSeek Harness who want to expand the functionality of text-only models for multimodal tasks. Whether you need image text extraction, visual analysis, or a localized solution for privacy protection, dsh-pseudo-vision provides efficient and convenient support, helping users overcome model limitations in AI applications.