dsh-knowledge-base
适用于 DeepSeek Harness (DSH) 的通用知识库插件：导入文档（md/txt/json/yml/docx/pdf）、按文件夹整理，并使用 FTS5 进行搜索。通用知识库插件：文件导入、目录化管理、全文检索、知识库管理界面。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add dsh-knowledge-base@0.1.5Knowledge Base 以 dsh-knowledge-base 发布，当前版本为 0.1.5。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
A general-purpose knowledge base plugin for DeepSeek Harness (DSH)
Import documents · organize in folders · full-text search · manage in the Web UI
dsh-knowledge-base turns documents into a searchable knowledge base for DeepSeek Harness agents. Drop in a PDF, Word, or Markdown file — it is parsed, chunked by section, and instantly searchable by your agent with FTS5 (BM25) ranked retrieval. A built-in Web UI lets you manage the knowledge base like a file manager: create/rename/delete folders, move files, and browse content.
md / txt / json / yml / docx / pdf, no size limit; PDF is parsed in-process with pdfjs-dist (cross-platform, zero system dependencies)kb_query / kb_import / kb_list / kb_update / kb_delete, usable directly by the model in conversationdsh plugin --profile web add dsh-knowledge-base
Dependencies: DeepSeek Harness (
dsh) provides the@deepseek-ai/*runtime; Node ≥ 22.5 (built-innode:sqlite).
Via the Web UI (recommended)
dsh web (or dsh --profile web)Via agent tools (headless / any profile)
import: Use kb_import to import /path/to/manual.pdf, category "Documents", tags ["manual"]
search: Use kb_query to search for "transformer"
list: Use kb_list / kb_list category=Documents
update: Use kb_update to change id=3's category to "Documents"
delete: Use kb_delete to delete id=3
# In the profile's cordis.patch.yml or when installing the bundle
# No categories are preset (fits any domain); entries without a category go to "Uncategorized".
# You can also create folders directly in the UI (persisted, no config edit needed).
- id: knowledge-base
name: 'dsh-knowledge-base'
config:
categories: # configure as needed, e.g.:
- Documents
- Manuals
| Method | Path | Purpose |
|---|---|---|
| POST | /api/kb/import | Upload a file (base64 JSON) → parse, chunk, store |
| GET | /api/kb/list | List entries and categories |
| GET | /api/kb/search?q= | Full-text search (FTS5 + BM25) |
| POST | /api/kb/update | Update an entry's category/tags |
| POST | /api/kb/rename-category | Rename a category |
| POST | /api/kb/create-category | Create a folder (persisted) |
| POST | /api/kb/delete-category | Delete an empty folder |
| POST | /api/kb/move-file | Move a file (change category) |
| POST | /api/kb/rename-file | Rename a file |
| POST | /api/kb/delete-file | Delete an entire file |
| POST | /api/kb/delete-entry | Delete a single entry |
dsh-knowledge-base (one npm package, three plugin rows)
├── dsh-knowledge-base host tools: kb_query / kb_import / kb_list / kb_update / kb_delete
├── dsh-knowledge-base/web Web endpoints: /api/kb/* (web composition only)
└── (client half) "Knowledge Base" conversation view tab + directory browser UI
Data storage (default):
$DSH_HOME/knowledge-base/kb.sqlite # entries + FTS5 index + meta (dynamic categories)
$DSH_HOME/knowledge-base/inbox/ # upload temp dir (cleaned up after import)
Tables: kb(id, category, name, summary, payload, tags, source, updated_at) + kb_fts (FTS5 external-content table) + meta (dynamic categories).
npm run build # tsc type-check + tsdown bundle (host half + client bundle)
# Local verification (use a workspace-local test home, never touch ~/.dsh)
DSH_HOME=$PWD/.dsh-home DSH_TELEMETRY_DISABLED=1 \
dsh --profile headless --patch dev-headless.cordis.yml \
"Use kb_import to import /tmp/test.md, then kb_query to search for 'keyword'"
The standalone repo ships
dev.cordis.yml/dev-headless.cordis.ymlfor local verification. During development the@deepseek-ai/*dependencies are symlinked from an official checkout viascripts/link-official-deps.mjs— see AGENTS.md → "Dependencies".
ctx.llm structured calls on import)PRs welcome! Please read AGENTS.md (agent development guide) first. Before submitting:
npm run build passes.dsh-home/, .test-workspace/, etc. — see .gitignore)MIT © dsh-knowledge-base contributors