dsh-knowledge: A Knowledge Base & RAG Plugin for DeepSeek Harness

soren-abt/dsh-knowledge

Data & KnowledgeVerified
Listed on 2026-08-20
Page last updated 2026-08-20

Knowledge base & RAG plugin for DeepSeek Harness: chunking, local embeddings, hybrid search, management panel.

Bring structured knowledge management and hybrid retrieval to DeepSeek Harness, letting DSH plugin users build local knowledge bases with ease.

install
dsh plugin --profile web add github:soren-abt/dsh-knowledge
Category
Data & Knowledge
Platform
DSH-Plugin
Author
soren-abt
Distribution
Plugin

dsh-knowledge Key Features

Knowledge base & doc managementSmart chunking & embeddingHybrid search & rerankModel tools & panel

dsh-knowledge Repository

soren-abt
Publisher
soren-abt
Repository
soren-abt/dsh-knowledge
7
Stars
3
Forks
0
Watchers
0
Open issues
LanguageTypeScript
LicenseGNU Affero General Public License v3.0
Website
Repo last push2026-08-20 13:37:08
Repo created2026-08-16 00:47:22
Topics
deepseek-harnessdshdsh-plugindsh-pluginsknowledge-based-systemsrag
Default branchmain

dsh-knowledge Repo Summary

dsh-knowledge is a knowledge base & RAG plugin (DSH plugin) for DeepSeek Harness (DSH), maintained by Soren-ABT under the MIT license, last updated in 2026-08. It addresses the lack of structured knowledge management and retrieval in DSH by providing knowledge base grouping, document management, intelligent chunking, local embeddings, hybrid search, reranking, and a Cherry Studio-style browser management panel. Core capabilities include per-base configuration of embedding and rerank models, heading-aware chunking, BM25+vector hybrid retrieval with RRF fusion, MMR deduplication, retrieval history replay, and 12 model tools such as knowledge_search and knowledge_read_document.

What core features does the dsh-knowledge DSH plugin support?

  • Knowledge bases & documents: Create/delete/rename knowledge bases and documents; group management (sidebar collapsible navigation, move to group); modal-based document addition (text/file/webpage/directory tabs); multi-file drag-and-drop upload (up to 20); directory recursive import (txt/md/csv/html/json/pdf/docx etc.); duplicate handling (keep all/replace) and content hash deduplication.
  • Per-base independent configuration: Each knowledge base can specify its own embedding provider/model (including local models), rerank model, chunk size, and topK; unset fields inherit global config; one-click reindex after config changes.
  • Vectorization & retrieval: Pluggable embedding providers (any OpenAI-compatible /embeddings endpoint, Ollama, or in-process local model via transformers.js); hybrid retrieval (BM25 + vector + Reciprocal Rank Fusion); rerank models (Jina/SiliconFlow/Cohere v2 style APIs); MMR deduplication; retrieval modes (auto/hybrid/vector/lexical) and similarity threshold; automatic fallback to keyword search when not configured.
  • Intelligent chunking: Heading-aware chunking that preserves Markdown heading paths and injects document title + heading path into embeddings and retrieval, improving recall.
  • Index management: Rebuild index, batch embedding, statistics (documents/chunks/characters/tokens).
  • Model tools: 12 tools, including knowledge_search, knowledge_list_bases, knowledge_create_base, knowledge_delete_base, knowledge_add_document, knowledge_list_documents, knowledge_delete_document, knowledge_import_url, knowledge_stats, knowledge_get_document, knowledge_read_document, knowledge_reindex_base.
  • Management panel: Sidebar bottom entry, workspace full-page overlay, Cherry Studio-style layout with search, group navigation, knowledge base cards, stats chips, tabular document list, chunk/original preview, reindex, retrieval test, global and per-base settings modals.
  • Local model management: Settings → 'Local Models' page with download/retry/delete buttons and real-time progress bar.
  • Persistence: Business state persisted via DSH storageDomain seam; chunk data stored in a separate SQLite file; lexical search via FTS5 trigram index; vector search scans stored vectors at query time; no full in-memory load at startup.

How to install dsh-knowledge?

Install from npm (recommended, no build required), or from a tarball/local source. Note: for pnpm 10+, you must allow build scripts for onnxruntime-node, sharp, and protobufjs in the profile's pnpm-workspace.yaml.

bash
dsh plugin --profile web add github:soren-abt/dsh-knowledge

After installation, restart the web service for host-side changes, and refresh the page to load the client panel.

Which DeepSeek Harness versions and platforms are compatible with dsh-knowledge?

DSH plugins are installed at the profile level, so the installation command is identical whether DSH is installed via npm or from a source clone. Compatibility details are in the README, but it is confirmed to work with DSH's official storageDomain seam and web profile.

What is the architecture of dsh-knowledge?

A bundle contains three plugin lines: knowledge (host core engine), tool-knowledge (host model tools), and ui-knowledge (client panel). The data model declares a knowledge domain via storageDomain, with bases, documents, chunks tables and a global slot.

Who is the author and what is the license of dsh-knowledge?

Maintained by Soren-ABT, licensed under MIT, last updated in 2026-08.

View more DSH plugins at dsh-plugin.org

DSH-Plugin FAQ