Introduction

The DSH (DeepSeek Harness) plugin ecosystem emphasizes “everything is a plugin.” When developing or working with DeepSeek models, developers often struggle with consolidating information: existing knowledge-base plugins are mostly static repositories that require manually dragging documents in, and conversational-memory plugins can record context, but the recorded data becomes stale over time. dsh-intelhub solves how to automatically turn dynamically incoming information (such as social media, web pages, and local files) into a searchable knowledge base, while also supporting feedback into workflows such as Obsidian.

Plugin Overview

dsh-intelhub is the first intel hub plugin natively built on zvec. It is maintained by IKEASven69, and its core value lies in turning “information you encounter” into “knowledge you can query.” The plugin does not rely on cloud APIs, runs entirely locally, and indexes and filters collected data through semantic and keyword hybrid retrieval.

Core Features

The plugin provides a complete pipeline for intelligence collection and retrieval:

  • Hybrid retrieval engine: Uses the RRF (Reciprocal Rank Fusion) algorithm, with a vector retrieval weight of 0.75 and a jieba tokenization keyword retrieval weight of 0.25. It supports semantic + keyword hybrid retrieval with source citations.
  • All-scenario data sources: Supports collection scripts, local folders, web URLs, and casually pasted text.
  • Local execution: No daemon processes, no API keys, and documents never leave the local machine. On first installation, it automatically downloads a ~30MB local embedding model (e5-small) to ~/.dsh/dsh-intelhub/.
  • Metadata filtering: Supports scalar filtering by fields such as author, stage, tags, and like count.
  • Workflow integration: Supports exporting to Obsidian and scheduled task orchestration.

Installation and Configuration

Install it using the DSH command-line tool:

dsh plugin add dsh-intelhub

On first use, the plugin automatically downloads the local model files. After installation, the plugin provides an “Intel Hub” settings entry in the settings panel, including drag-and-drop import, resident directory management, scheduled task settings, and metadata filtering options.

Common Commands

The plugin provides a set of core commands for controlling the lifecycle of the intel hub:

  • kb_watch: Register a resident directory. It works with collection scripts and automatically indexes file changes incrementally every 3 hours.
  • kb_search: Run hybrid retrieval. Input natural language, and it returns matching results with information such as author, like count, and file chunk ID.
  • kb_note: Capture a judgment. After writing text in a session, the content is indexed and can be recalled directly next time.
  • kb_export: Export with feedback. Writes search results to plain Markdown files on disk, which Obsidian can index directly without plugins.
  • kb_schedule: Scheduled scans. Supports setting automatic scans at 09:00 daily or at other intervals.

Notes

While using the plugin, keep the following limitations in mind:

  • Version compatibility: An index directory schema upgrade triggers a one-time rebuild, and the old directory is deprecated.
  • Concurrent writes: Simultaneous writes from two machines to the same index directory are not supported.
  • Model limitations: For extremely difficult cases (such as adjacent operations concepts like “timeout” vs. “rollback”), discrimination can be low under the current local model. You can configure an upgraded embedding layer.
  • Installation issues: For pnpm users, if the ERR_PNPM_IGNORED_BUILDS error appears during installation, add ignoredBuiltDependencies: [@zvec/zvec, onnxruntime-node, protobufjs, sharp] to pnpm-workspace.yaml in the profile.

Summary

dsh-intelhub is suitable for developers who need to build a dynamic knowledge base locally, prioritize privacy protection, and rely on the DeepSeek ecosystem. Through zero daemon processes and a local model, it creates a complete workflow from information collection to retrieval and then to Obsidian export. You can find more details and the source code in the GitHub repository.