type: KnowLP document status: engine date: "2026-08-29" note: engine README (v3.0.8 repo edition, dsh-first)

KnowLP-RAG

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.

Quick start (3 steps)

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)

Use from any MCP client

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.

Try the demo vault (no private data, no embedding model)

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.

Seven tools

| 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 (optional cross-machine visual 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:

1. desktop GPU — the primary dedicated box (an RTX-class machine reachable over Tailscale)
2. local — a same-machine fallback
3. cloud API — a hosted PixelRAG endpoint

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.

Documentation

• Install & usage guide: docs/usage.md
• Troubleshooting: docs/troubleshooting.md
• dsh integration details (env vars / Cordis plugin): dsh/README.md

Why 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).

Demo

$ 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)

Local development

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"

Tests

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.

View Architecture Diagram