Introduction

DeepSeek Harness (DSH) focuses on a plugin-based architecture. When developing agents or debugging code, quickly locating the specific logic in source code is often more efficient than simple text matching. Traditional search tools rely on keyword matching and struggle to understand code semantics. The semantic-search plugin addresses this problem by converting codebase snippets into vector indices and supporting natural-language queries.

What Is It?

This is a local semantic code search plugin for DeepSeek Harness (DSH), maintained by developer JohnXu22786. The core value of the plugin is that it enables semantic understanding and retrieval of code without requiring an internet connection, supporting snippet-level indexing and hybrid search strategies.

Core Features

The plugin provides the following main capabilities:

  • Code indexing: Supports snippet-level, symbol-aware index construction. It can recognize boundaries such as functions and classes, and supports more than 16 programming languages (including TypeScript, Python, Go, Rust, Java, C/C++, Ruby, PHP, Swift, Bash, Markdown, JSON, YAML, etc.).
  • Offline embeddings: Uses offline lexical embeddings by default, requiring no network or API key. It also supports configuring OpenAI-compatible endpoints.
  • Hybrid retrieval: Combines vector cosine similarity with the BM25 algorithm, and merges results using a reciprocal rank fusion (RRF) strategy.
  • Toolset: Provides three DSH tools (sema_search, sema_reindex, sema_stats) and a standalone sema command-line tool.
  • Incremental updates: Supports incremental refreshing and file watching, and persists the index to the <root>/.sema directory.

Installation and Enablement

Before installation, ensure the environment meets the requirements: The Node.js version must be in the range ^22.19.0 or >=24.0.0.

  1. Run the installation command:
    dsh plugin --profile demo add github:JohnXu22786/semantic-search
  1. After installation, the plugin registers the corresponding tools in the DSH toolchain.

Typical Usage

After installation, you can use the sema command to build, update, and search the index.

  • Build index: Build a full index from the workspace.
    sema index
  • Incremental update: Perform an incremental rebuild based on file changes, or use --full to perform a full rebuild.
    sema reindex [--full]
  • Semantic search: Hybrid vector and BM25 retrieval, outputting top-ranked results.
    sema search <query...>
  • Index statistics: View index health status, provider information, and data size.
    sema stats [--]

Common global options include:
* --root <dir>: Specify the workspace root directory (defaults to the current directory).
* --provider <kind>: Specify the embedding provider type (defaults to lexical).
* --dim <n>: Specify the embedding dimension.

Use Cases and Considerations

  • Offline-first: The plugin works offline by default; the built-in lexical provider does not depend on the network. If a remote provider is configured and the connection fails, it gracefully falls back to offline mode.
  • Determinism: The same source code and configuration produce deterministic indices and ranking results.
  • Permissions and security: The plugin runs with the permissions of the DSH process. It is recommended to review the source code and license (MIT) before installation.
  • Ecosystem note: This plugin is a community-maintained project, published through directories such as SkillHub, and has no official affiliation with DeepSeek or High-Flyer.

Brief Conclusion

semantic-search provides a lightweight, localized semantic code search solution for the DSH ecosystem. Through snippet-level indexing and hybrid retrieval, it can significantly improve the accuracy of code localization. For more details, refer to the plugin directory or the GitHub repository.