Introduction

DeepSeek Harness (DSH) uses a plugin-based architecture. Developers or users working with the web profile often need to handle visual tasks such as document screenshot processing and table recognition. Existing solutions often truncate long-document recognition due to token limits and lack unified export and history management tools. aura-vision is a free visual recognition plugin designed specifically for the DSH web profile. It aims to address these pain points by providing a complete workflow from recognition to export.

What Is It?

aura-vision is a free visual recognition plugin maintained by Ck-epsilon under the MIT License.

It is built specifically for the DeepSeek Harness web profile, with its core value lying in long-document recognition support and multiple export formats. By integrating free or compatible interfaces such as Zhipu GLM-4V-Flash, it makes local visual capabilities easily accessible.

Core Features

The plugin provides the following core capabilities:

  1. Multi-channel support

    • Prioritizes Zhipu GLM-4V-Flash (free tier).
    • Supports any OpenAI-compatible multimodal interface.
    • Provides an anonymous Pollinations fallback.

  2. Long-document recognition

    • Recognizes documents block by block using adaptive grid segmentation (target block size of 1100px, maximum 3×3, 8% overlap), bypassing the 1024-token output limit.
    • Applies gentle normalization to images larger than 3200px.

  3. UI and interactions

    • Uses the Aura UI design language: a highly translucent frosted-glass effect (55% base color + 28px blur) and Samsung-style small rounded corners.
    • Light and dark themes are self-consistent, with secondary information appearing subtle or barely visible.

  4. History management

    • Stores thumbnails and original images separately.
    • Supports favoriting, filtering, deleting, and clearing history.
    • Exports Markdown (original image embedded as base64, self-contained in a single file).

  5. Multi-format result export

    • Supports exporting Markdown, Word (.doc), PNG long images, and Excel (when tables are present).
    • Double-click a preview image or detail image to enlarge it to fullscreen.

Installation and Enablement

  1. Run the installation command:
    dsh plugin --profile web add aura-vision
  1. dsh must be restarted for the changes to take effect. After restarting, an “Image Recognition” button will appear to the right of the input box in any conversation, indicating successful installation.

Typical Usage

  • Basic use: Click the “Image Recognition” button, upload an image, and wait for the recognition result.
  • Long-document processing: When uploading a long image, the system automatically segments it and processes the blocks sequentially. The more blocks, the longer it takes; this is expected behavior.
  • Uninstall plugin: To remove it, run:
    dsh plugin --profile web remove aura-vision

Notes

  • Restart required: After installation or upgrade, the dsh process must be restarted to load the new features.
  • Token limit: The 1024-token output limit of glm-4v-flash is a hard limit of the free model. Although the plugin mitigates this through segmentation, more complete results require configuring a model with longer output support and setting it as current.
  • Data storage: Data is stored in the <workspaceRoot>/.aura-vision/ directory and migrates with the workspace.
  • Key management: The API key is stored in local credential storage. After switching machines, history migrates with the workspace, but the key must be entered again on the new machine.
  • Privacy and security: Images are sent only to the currently selected backend (Zhipu or your configured interface), and the local machine does not upload them to any third party.

Short Conclusion

As an open-source plugin, Aura Vision provides a practical toolchain from long-document recognition to multi-format export. For more technical details and source code, refer to its GitHub repository: https://github.com/Ck-epsilon/aura-vision