Introduction

DeepSeek Harness (DSH) adopts a plugin-based architecture, allowing capabilities to be extended through extensions. For developers who need to access local documents in conversations, DSH-RAG provides a solution to turn local folders into a retrievable knowledge base.

What It Is

DSH-RAG is a local knowledge base plugin maintained by imkelt. It turns explicitly permitted local folders into named knowledge bases and provides answers grounded in sources when the user explicitly invokes retrieval.

Core Features

DSH-RAG provides the following capabilities:
* Named knowledge bases: Build reusable knowledge bases from local folders.
* Explicit /rag retrieval: Retrieval is completed before the model answers, rather than relying on the model to decide whether to call a tool.
* Session-scoped search: By default, searches the knowledge base selected for the current session, and supports one-turn override with @knowledge-base-name.
* Verifiable evidence: Displays source cards, including file name, structural context, excerpt, and precise page number/line number.
* Hybrid retrieval and offline fallback: Local multilingual BM25 works without an API Key or network access; optional SiliconFlow can be used for vectorized retrieval and reranking.
* Local document support: Supports PDF, DOCX, PPTX, Markdown, and UTF-8 text, with incremental indexing, scan cancellation, and resumable scanning.
* Bilingual and theme-aware UI: Follows DSH language settings and adapts to light/dark/system appearance.

Installation and Enablement

Prerequisites:
* Node.js ^22.19.0 or >=24.0.0
* pnpm available and in PATH
* DeepSeek Harness (tested version 0.1.1-rc.2)

The installation command is as follows:

npx @deepseek-ai/dsh plugin --profile web add @kai232/dsh-rag

After installation, start the Web profile.

Typical Usage

The basic usage is /rag <question>.

/rag What does the project note say about the launch window?

If you need to temporarily switch the knowledge base scope, you can use the @ selector:

/rag @"Project notes" Which Python version does the virtual environment use?

The command completes retrieval first and then passes the evidence to the model. Source cards are displayed in the UI.

Use Cases and Notes

  • Permissions and security: The plugin runs with the permissions of the current DSH process. Ensure read/write permissions for the target folder.
  • Compatibility: Currently verified on DeepSeek Harness 0.1.1-rc.2; other versions may be compatible but are untested.
  • License: Based on the MIT License.

Short Conclusion

DSH-RAG fits the Web profile plugin ecosystem and lowers the barrier to integrating local knowledge bases through explicit commands and source traceability. Its offline BM25 retrieval capability makes it still available in network-less environments.

More information and source code:
* GitHub repository
* Plugin directory