local-knowledge-base
A local-first RAG knowledge base for DeepSeek Harness (DSH). Drag documents into a folder, choose per-file vector retrieval, preview the original PDF/Markdown, and let the agent answer with citations.
本地优先的 DSH 知识库插件:把文件拖进目录,即可在对话中检索、阅读并引用;支持 PDF 解析、文件级 RAG 开关,以及一个专用的「知识库问答」智能体模式。
A community plugin. Not affiliated with DeepSeek. MIT licensed.
Install
The plugin is published on npm as local-knowledge-base.
Profile CLI (no prompts, nothing to confirm):
dsh plugin --profile web add local-knowledge-base
Replace web with your profile name. Restart dsh once — a newly added bundle is not applied to a running process — then open 本地知识库 / Knowledge base in the left sidebar.
From the DSH Web UI instead:
- Open Plugins → Add plugin.
- Enter
local-knowledge-base. - Install and enable the bundle.
- Refresh the page if the sidebar icon does not appear.
Or ask the agent to install it with plugin_manager:
install_bundle target: local-knowledge-base
Pin a version
dsh plugin --profile web add local-knowledge-base@1.0.4
The explicit pin matters: pnpm holds back very recent releases, so a bare add can resolve to an older version.
Add to Discover
The package ships a format-1 market index. Add one of these URLs under Plugins → Discover → Manage index sources:
https://unpkg.com/local-knowledge-base@1.0.4/market-index.json
https://cdn.jsdelivr.net/npm/local-knowledge-base@1.0.4/market-index.json
The bundle installs two rows:
knowledge-base— the Host/Client plugin and theknowledge_basetool.preset-knowledge-qa— the 知识库问答 Agent Preset.
Features
- Drag-and-drop import — drop one or many files anywhere on the knowledge-base page. Large PDFs stream to the Host instead of being read into browser memory.
- PDF text extraction — uses
pdfjs-dist; the extracted text is cached under%DSH_HOME%\knowledge-base-cache, so later searches are fast. - Per-file RAG — every searchable file has a
RAGtoggle. RAG files use local TF-IDF vector cosine retrieval; non-RAG files use BM25 keyword retrieval. Results are merged and tagged[RAG]or[keyword]. - Agent tool — registers
knowledge_basewithlist,search,read,add,import, anddeleteactions. Search results includepath, line range, snippet, and retrieval mode. - Dedicated agent mode — the bundle also installs the 知识库问答 / Knowledge Q&A Agent Preset. In that mode the agent must search the library first, cite
path:startLine-endLine, and must not invent an answer when nothing matches. - Persistent folder — documents stay in the configured folder between sessions; there is no need to re-import them each time.
- Live updates — the directory is rescanned on every call; files added, edited, or deleted are reflected on the next search.
- Preview keeps format — PDFs are served as the original file through a range-capable
/rawroute and rendered by the browser's PDF viewer; Markdown is rendered with headings, lists, code blocks, and links, with a source toggle. - Custom directory — change the folder from the management page; the choice is saved under
%DSH_HOME%\knowledge-base-settings.json.DSH_KB_DIRor the plugin row'sconfig.directorytakes precedence. - Local only — no cloud embedding service, no file upload to a third party. The Host reads files from disk.
Use
Manage files
Open 本地知识库 / Knowledge base in the sidebar:
- Drop files onto the page.
- Toggle 新文件启用 RAG / RAG for new files before dropping.
- Use the row-level
RAGbutton to change one file later. - Add, preview, filter, and delete files from the same page.
- Change the directory with 更改目录 / Change directory.
Default directory:
%DSH_HOME%\knowledge-base
On Windows this is usually C:\Users\<name>\.dsh\knowledge-base.
Ask questions
Start a new session and select the 知识库问答 preset, then ask normally:
根据知识库解释反向传播和梯度下降的区别。
知识库里有哪些关于动态规划的内容?
The agent calls knowledge_base automatically. No per-session import is needed.
Supported files
- Text:
.md,.markdown,.mdx,.txt,.rst,.json,.jsonl,.csv,.tsv,.yaml,.yml,.toml,.ini,.log,.html,.xml, and common source-code extensions. - PDF:
.pdf(text layer required; scanned image-only PDFs need OCR). - Limits: 32 MB per text file, 256 MB per PDF.
- DOCX, images, and other binary formats are not indexed yet.
RAG details
RAG means retrieval-augmented generation. This plugin owns the retrieval layer; the selected DSH model performs the generation.
- RAG files: chunks are tokenized into terms (latin words + CJK bigrams), weighted by TF-IDF, and ranked by cosine similarity.
- Non-RAG files: BM25 keyword ranking.
- Merged results: each result carries
retrieval: 'rag' | 'keyword'. - No external embeddings: the current vectorizer is local and lexical, not a neural embedding model. It is lightweight and private, but it does not provide deep semantic matching. A future version can add an OpenAI-compatible or Ollama embedding backend.
- File-level setting:
%DSH_HOME%\knowledge-base-rag.json, grouped by knowledge-base directory.
Privacy and security
- Files never leave the machine unless you explicitly send their content to a model in a conversation.
- The management API is same-origin and refuses
Origin/Sec-Fetch-Sitecross-site requests. pathvalues are resolved inside the configured knowledge-base directory; traversal is rejected.- PDFs are served inline with
X-Content-Type-Options: nosniff; the raw route supports byte ranges for the browser PDF viewer. - The plugin does not execute uploaded files.
Requirements
- DSH
0.2.x(the Web UI profile). - Node.js 20 or newer on the DSH host, for the bundled
pdfjs-dist@4. - Desktop or
dsh web; the browser half mounts into@deepseek-ai/dsh-client-ui-layout.
Development
index.js Host plugin: scanning, chunking, PDF extraction, retrieval, tools, HTTP API
client.js Web UI: sidebar panel, drag/drop, preview, RAG toggles
cordis.patch.yml Bundle rows: the plugin and the Knowledge Q&A preset
locale/ Plugin Manager title/description
icon.svg Bundle icon
The Host requires pdfjs-dist@4.10.38 or a compatible 4.x release. It is declared as a dependency and installed with the package.
If you develop against a local checkout, link it into a profile instead of installing from npm:
dsh plugin --profile web add link:/absolute/path/to/local-knowledge-base
License
MIT — see LICENSE. pdfjs-dist is distributed under the Apache-2.0 license; see its package for details.
No comments yet. Be the first to write one.