Preface

In developing DeepSeek Harness, enabling agents to query local documents is a common requirement. Existing solutions often rely on cloud vector databases or require complex local environment setups. The dsh-move-rag plugin provides a lightweight solution: it does not require downloading model weights or relying on cloud services; instead, it runs the knowledge base on the Windows desktop and uses an always-on-top window for quick ingestion and retrieval.

What It Is

This is a DeepSeek Harness plugin maintained by xingmen-1, providing local knowledge-base memory. Through a persistent WinForms icon on the desktop, it supports drag-and-drop ingestion, local RAG retrieval, and exposes the retrieval results as a tool to the Agent.

Core Features

  • Desktop temporary panel: A WinForms window that supports always-on-top and dragging. When dragged to the edge of the screen, the window automatically collapses into an arrow.
  • Drag-and-drop ingestion: Supports PDF, Word, txt, md, csv, code files, and images. Dragging a file onto the desktop icon or into the panel completes ingestion.
  • Local RAG: Uses fixed-length chunking + hash embeddings (512-dimensional) + cosine similarity recall for retrieval. You can directly test retrieval quality in the panel.
  • Tool interface: Provides the search tool knowledge_search and the development tool kb_dev to the DSH Agent.
  • Web synchronization: The DSH web sidebar provides a “Knowledge Base” entry, allowing you to control the visibility of the desktop icon.

Installation and Enablement

Use the official installation command to add the plugin. After installation, restart dsh web to load the configuration.

dsh plugin --profile web add github:xingmen-1/dsh-move-rag

Typical Usage

  1. Open always-on-top: In the web sidebar, click “Knowledge Base” and turn on the “Always on Top” switch. A floating icon will appear in the top-right corner of the desktop.
  2. Ingest: Drag a file onto the desktop icon or into the opened panel.
  3. Retrieve: Enter a question in the panel and press Enter. Results are sorted by similarity.
  4. Let the Agent use it: Ask directly in the conversation; the Agent will automatically invoke the knowledge_search tool.

Applicable Scenarios and Notes

  • System limitations: The plugin currently only supports Windows. Linux/macOS requires replacing the desktop panel and file-writing implementation.
  • Embedding principle: Hash embeddings provide keyword-level semantic approximation, not neural embeddings. They are not sensitive to synonymous paraphrases, but they are offline, zero-dependency, and fast.
  • File indexing: Images are indexed only by filename; there is no OCR capability.
  • File size: The maximum size for a single uploaded file is 16 MB (base64 via JSON); the limit for ingesting local paths is 256 MB.

Conclusion

This plugin is suitable for developers who need to quickly build a local knowledge base in a Windows environment and grant agents retrieval capabilities. The plugin directory and GitHub link are as follows: