@eqman00003/knowlp-rag
KnowLP-RAG：面向 Markdown 笔记的双知识图谱检索——DeepSeek Harness (dsh) 套件：MCP 服务器 + 原生 Cordis 插件
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:wly8691-jpg/knowlp-rag#e9a60b7fdd7c14ee3c7b924a4b5b529d068c5906Knowlp Rag 以 @eqman00003/knowlp-rag 发布，当前版本为 3.0.11。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。
Agent-first knowledge retrieval — turn your Markdown notes into a self-maintaining knowledge graph that is "use it or lose it". Agents (DSH / Claude Code) install, build the graph, and self-check it; only the vault path must be provided by the human. Retrieval returns reading paths: which notes to read, in what order, and which are similar substitutes.
All three steps are agent-runnable; only KNOWLP_VAULT (your notes directory) must be provided by the human.
# 1. Install (official npm registry)
dsh plugin add "@eqman00003/knowlp-rag"
# 2. Set the two required env vars (without them the dual-graph engine idles and only full-text search works)
export KNOWLP_VAULT="$HOME/Notes" # your Markdown notes directory
export KNOWLP_GRAPH_DIR="$HOME/.knowlp-dsh" # writable index directory
# 3. Restart dsh web — the first search triggers Python env bootstrap (~30s, don't interrupt)
A .mcp.json ships at the repo root — copy it next to your project (or into your host's config), then provide the two required env vars (KNOWLP_VAULT, KNOWLP_GRAPH_DIR; the file reads them via ${...}, so setting them in your shell is enough).
{
"mcpServers": {
"knowlp": {
"type": "stdio",
"command": "npx",
"args": ["--yes", "--package", "@eqman00003/knowlp-rag", "knowlp-mcp"],
"env": { "KNOWLP_VAULT": "${KNOWLP_VAULT}", "KNOWLP_GRAPH_DIR": "${KNOWLP_GRAPH_DIR}" }
}
}
}
Two optional env vars matter when your notes live outside a plain-vault setup:
KNOWLP_SKILL_INDEX — path to a skill index (enables skill_search);HERMES_HOME — only needed for the chroma skill index when your agent home is not ~/.hermes; the chroma DB resolves to $HERMES_HOME/skills/.chroma/chroma.sqlite3.The config shape is verified over stdio (initialize → tools/list → a real knowlp_search); any stdio MCP host works.
A 7-note bilingual demo vault ships with the repo. From clone to first search:
POSIX:
pip install -e .
export KNOWLP_VAULT="$PWD/examples/demo-vault"
export KNOWLP_GRAPH_DIR="$PWD/.demo-graph"
python build_graph.py && python -m vector_index --build
python knowlp_search.py "How should I understand this RAG architecture?" --hybrid
Windows PowerShell:
pip install -e .
$env:KNOWLP_VAULT = "$PWD\examples\demo-vault"
$env:KNOWLP_GRAPH_DIR = "$PWD\.demo-graph"
python build_graph.py; python -m vector_index --build
python knowlp_search.py "How should I understand this RAG architecture?" --hybrid
Five bilingual verification queries and what each demonstrates: docs/demo.md. Agent onboarding instructions: examples/agent-setup.md.
| Tool | Purpose |
|---|---|
knowlp_search | Four-engine fan-out retrieval (dual-graph P/S-Agent + vector + full-text) |
knowlp_get_note | Read note content (read-only, path-traversal safe) |
knowlp_stats | Engine/graph health self-check (first stop for troubleshooting) |
knowlp_status | Index lifecycle status (missing/fresh/stale + reason + exact fix command) |
knowlp_record_feedback | Explicit feedback (the only entry point of the weight loop) |
knowlp_record_correction | Explicit preference pairs (chosen ≻ rejected) — the input to preference learning |
skill_search | Skill index retrieval |
PixelRAG is an optional visual-retrieval engine — it embeds visual content for retrieval instead of relying on text tokens alone. It runs on a separate GPU machine on your network: the agent offloads the visual-embedding work to that box over Tailscale rather than computing it on the laptop. Retrieval falls back through three tiers:
Configure it via KNOWLP_PIXELRAG_DESKTOP / pixelrag_local. Unconfigured, it stays off — retrieval still works in n-gram / embedding mode.
Reproducing this: it is deployment-specific — you need your own GPU machine running a PixelRAG service, a network path to it (e.g. Tailscale), and its endpoint address. No bundled service ships with KnowLP.
Grep for "RAG architecture" gives you 105 files. KnowLP gives you 3 ranked hits with dependency context.
grep | Naive vector store | KnowLP | |
|---|---|---|---|
| Result ranking | ❌ | ✅ | ✅ |
| Dependency chain (P-Agent) | ❌ | ❌ | ✅ |
| Similar substitutes (S-Agent) | ❌ | ❌ | ✅ |
| Works without GPU | ✅ | ❌ | ✅ (n-gram mode) |
| Improves with use (feedback) | ❌ | ❌ | ✅ (weight loop) |
| Paragraph-level matching | ❌ | ❌ | ✅ |
| Decay & forgetting (use it or lose it) | ❌ | ❌ | ✅ (three half-life tiers) |
The difference: vector search finds documents that "contain keywords"; KnowLP finds documents you should read given your query, with reading paths. Edge weights between notes evolve with usage — consumed edges strengthen, unused edges decay by half-life (ephemeral 1 day / default 30 days / declarative never).
Works with Chinese note vaults out of the box (Chinese time-anchor queries and Chinese full-text search are supported).
$ knowlp_search "RAG architecture"
1. [HIT] RAG Architecture.md (score 0.77)
2. [LINK] Vector Database Selection.md (score 0.61) ← prerequisite chain
3. [LINK] Retrieval Eval Pitfalls.md (score 0.42)
4. [LINK] _Index-Reading Order (depth 1) ← tells you where to start reading
5. [LINK] _Index-Related Concepts.md (depth 2)
git clone https://github.com/wly8691-jpg/knowlp-rag.git
cd knowlp-rag
pip install -e . # provides knowlp-mcp / knowlp-build / knowlp-search
# configure vault in config.yaml → build graph → search
python build_graph.py
python knowlp_search.py "RAG architecture"
pip install -e ".[mcp,modeling]" pytest pypdf
python -m pytest -q
Nothing in the suite needs a GPU, a real Office install, or a private vault, so
CI runs all of it (.github/workflows/ci.yml, on every push and PR).
| Layer | What it covers | In CI |
|---|---|---|
tests/test_packaging_closure.py | The two manifests against the real import graph: every top-level module is covered, and everything the MCP entry can reach is in npm files. A deliberately empty files entry turns it red and names the importers. | yes |
the rest of tests/ | Retrieval and routing, dual graph, modality pools, the preference loop, evidence/result contracts, and a real MCP stdio integration test. | yes |
| the skips | Four need symlink privileges (developer mode on Windows). The fifth — the real-vault parity case — only runs when KNOWLP_VAULT points at a real vault. | — |
Against a real vault locally: 303 passed, 4 skipped. On CI, where the vault is
deliberately unreachable: 302 passed, 5 skipped. The one test difference is the
real-vault parity case, and it says so in the skip reason.
test_packaging_closure.py exists because 3.0.10 shipped without it: the npm
tarball was missing 16 top-level modules, six of them in the MCP entry's import
closure, and knowlp_search could not run at all from an npm install.