Preface

DSH’s plugin ecosystem emphasizes “everything is a plugin,” and the community directory is an independent site, not belonging to the DeepSeek or Hypersphere official app stores. For developers using DSH or Claude Code, Markdown notes are often used as background material, experience records, or project context. The common problem is not completely failing to find files, but after finding many files containing keywords, one still needs manual judgment: which one to read first, which are related, and which can replace each other.

Below introduces KnowLP-RAG. It organizes Markdown notes into a retrievable knowledge structure and provides DSH / Claude Code with retrieval results containing reading paths: which notes to read, in what order, and which are similar alternatives.

What is it

KnowLP-RAG is a DeepSeek Harness (dsh) bundle containing an MCP server and a native Cordis plugin.

The verified basic information is as follows:

  • npm package name: @eqman00003/knowlp-rag
  • GitHub repository: wly8691-jpg/knowlp-rag
  • package.json version: 3.0.8
  • License: MIT
  • package.json engines: node >=18
  • README note: Python 3.11+
  • Release configuration: npm registry https://registry.npmjs.org/, public access
  • Community directory page: https://www.skillhub.cn/plugins/wly8691-jpg/knowlp-rag

Its main positioning is: dual knowledge-graph retrieval oriented towards Markdown notes, combined with decay-based forgetting, feedback weighting, vector retrieval, and full-text retrieval.

Core Capabilities

Retrieval Path

KnowLP-RAG’s retrieval results for DSH / Claude Code are not only sorted by relevance but also provide a reading path. Specifically including:

  • Which notes to read;
  • In what order to read;
  • Which notes are similar alternatives.

Its retrieval uses four-engine fan-out, including:

  • dual-graph P/S-Agent;
  • vector;
  • full-text.

Tool Interfaces

The five verified tools are as follows:

  • knowlp_search: four-engine fan-out retrieval, using dual-graph P/S-Agent, vector, and full-text;
  • knowlp_get_note: reads note content, read-only, path-traversal safe;
  • knowlp_stats: engine/graph health self-check;
  • knowlp_record_feedback: explicit feedback, the only entry point for the weight loop;
  • skill_search: skill index retrieval.

Decay & Forgetting

KnowLP-RAG supports decay-based forgetting, meaning “use strengthens, non-use decays.” The verified half-life tiers are:

  • ephemeral: 1 day;
  • default: 30 days;
  • declarative: never.

Feedback enters the weight loop via knowlp_record_feedback. After use, the system improves based on feedback rather than merely maintaining a static index.

Matching & Runtime Modes

Verified capabilities include:

  • paragraph-level matching;
  • supports Chinese note vaults out of the box, including Chinese time-anchor queries and Chinese full-text search;
  • works only in n-gram mode when there is no GPU;
  • improves through the feedback weight loop as usage increases.

Installation & Enablement

First, install the plugin.

dsh plugin add "@eqman00003/knowlp-rag"

Then set two necessary environment variables. Without these variables, the dual-graph engine will be in an idle state, leaving only full-text search available.

export KNOWLP_VAULT="$HOME/Notes"
export KNOWLP_GRAPH_DIR="$HOME/.knowlp-dsh"

Where:

  • KNOWLP_VAULT: your Markdown notes directory;
  • KNOWLP_GRAPH_DIR: writable index directory.

Finally, restart dsh web. The first search triggers a Python env bootstrap, which takes about 30 seconds; do not interrupt during this process.

Typical Usage

Search in DSH

After setting environment variables and restarting, you can first use a query to verify the retrieval results.

knowlp_search "RAG architecture"

If you need to use a tool to read a specific note, you can use:

knowlp_get_note

If you need to check the health status of the engine or graph, you can use:

knowlp_stats

If you want to record explicit feedback, use:

knowlp_record_feedback

If you want to retrieve the skill index, use:

skill_search

Local Development

Local development can start from the GitHub repository.

git clone https://github.com/wly8691-jpg/knowlp-rag.git

After entering the repository, install development dependencies:

pip install -e .

Build the graph:

python build_graph.py

Execute local search:

python knowlp_search.py "RAG architecture"

Use Cases & Notes

Suitable for the following scenarios:

  • You use DSH or Claude Code and maintain a set of Markdown notes;
  • You want retrieval results to not just hit keywords but provide a readable path;
  • You use a Chinese note vault and need Chinese time-anchor queries or Chinese full-text search;
  • You do not have a GPU but can use n-gram mode;
  • You want the retrieval weights to change with usage through explicit feedback.

Before use, you need to confirm a few things:

  • The plugin runs with the current dsh process permissions; you should check the source code and license before installing;
  • KNOWLP_VAULT and KNOWLP_GRAPH_DIR are necessary configurations for the dual-graph engine;
  • The Python env bootstrap after the first search takes about 30 seconds and should not be interrupted;
  • In scenarios without a GPU, it can only work in n-gram mode;
  • knowlp_get_note is read-only and path-traversal safe;
  • knowlp_stats can be used for engine/graph health self-check.

Links

  • Community directory page: https://www.skillhub.cn/plugins/wly8691-jpg/knowlp-rag
  • GitHub: https://github.com/wly8691-jpg/knowlp-rag