imaused/dsh-kb-lite

A host-only knowledge base plugin with five tools - ingest, search, ask, stats and clear - over a local index with BM25, optional dense retrieval and reciprocal-rank fusion, needing no database, daemon or port.

A host-only plugin with five tools and no browser UI: kb_ingest indexes files or directories (.md, .markdown, .txt, .pdf, recursive by default, incremental by sha256, max_files default 50), kb_search returns scored snippets with top_k default 8, kb_ask returns verbatim passages within a max_chars budget (default 6000) for the model to cite, kb_stats reports document and chunk counts plus up to 50 indexed paths, and kb_clear deletes the index after confirm: true. Retrieval modes are auto, bm25, dense and hybrid: BM25 is a pure-Python implementation with Latin tokenisation, CJK bigrams, k1 = 1.2 and b = 0.75; dense retrieval uses sentence-transformers with BAAI/bge-small-zh-v1.5; hybrid fuses both with reciprocal rank fusion at k = 60. Citations carry path, title, heading, start_line/end_line and, for PDFs, page. Markdown is chunked at ATX headings, 1200 characters with 120 overlap. One directory per knowledge base, no database, daemon or port.

Memory ★ 1
View on GitHub ↗

Install

dsh plugin --profile web add github:imaused/dsh-kb-lite

Copied verbatim from the README's "Or straight from GitHub" bash block, which follows the dsh-kb-lite npm form. The the npm registry check (2026-10-17) for dsh-kb-lite returned HTTP 404 Not Found, so no npm package is published and the GitHub form is shipped; the README's GitHub block also tells you to replace imaused with the account that publishes the repository.

Compatibility

Requires Node.js 18 or newer for the host side and Python 3.9 or newer for the engine; pymupdf is optional for PDF ingest and numpy plus sentence-transformers optional for dense retrieval, both installed into the interpreter the plugin launches. After install you must restart DSH and open a new session, because tools are bound when a session is created. Peer ranges are written with an explicit prerelease branch so 0.2.0-rc.x matches; the README states Linux and macOS are untested. The engine opens no sockets and makes no network requests, and all state lives under kb_root (default ~/.kb-lite).

Details

Alternatives

ICCuse/dsh-knowledge · ICCuse/dsh-file-memory · madage/dsh-self-improved

More plugins in Memory

Browse more in Memory

Guides for Memory plugins