knowledge-graph

knowledge-graph

Query and traverse an Obsidian vault as a knowledge graph. Semantic search, path finding, community detection, and graph analysis — all local, no cloud APIs.

What it does

Parses an Obsidian vault into an untyped graph (files = nodes, wiki links = edges), indexes it into SQLite with vector embeddings and full-text search, and exposes 10 operations via CLI and MCP server.

Search: Semantic search via local embeddings, full-text keyword search via FTS5.

Traverse: Find paths between nodes, shared connections, N-hop neighborhoods, local subgraphs.

Analyze: Community detection (Louvain), bridge nodes (betweenness centrality), central nodes (PageRank).

Install

git clone https://github.com/obra/knowledge-graph.git
cd knowledge-graph
npm install

Set your vault path:

export KG_VAULT_PATH=/path/to/your/obsidian/vault

Optionally set the data directory (defaults to ~/.local/share/knowledge-graph):

export KG_DATA_DIR=/path/to/data

CLI usage

Index your vault (first run downloads a 22MB embedding model):

npx tsx src/cli/index.ts index

Then query:

# Look up a node (brief mode — metadata + connections)
npx tsx src/cli/index.ts node "Alice Smith"

# Full content + edge context
npx tsx src/cli/index.ts node "Alice Smith" --full

# Semantic search
npx tsx src/cli/index.ts search "distributed systems framework"

# Full-text keyword search
npx tsx src/cli/index.ts search "distributed systems" --fulltext

# Find paths between two nodes
npx tsx src/cli/index.ts paths "Alice Smith" "Widget Theory"

# Shared connections
npx tsx src/cli/index.ts common "Alice Smith" "Bob Jones"

# Local neighborhood
npx tsx src/cli/index.ts neighbors "Alice Smith" --depth 2

# Subgraph extraction
npx tsx src/cli/index.ts subgraph "Widget Theory" --depth 1

# Community detection
npx tsx src/cli/index.ts communities

# Bridge nodes (connectors between clusters)
npx tsx src/cli/index.ts bridges --limit 10

# Central nodes (PageRank)
npx tsx src/cli/index.ts central --limit 10

All commands return JSON. Names are fuzzy-matched (title, aliases, substring). You can also pass full node IDs (file paths).

Claude Code plugin

This repo is also a Claude Code plugin. Add it to your plugins and the MCP server starts automatically, exposing all 10 operations as tools (kg_node, kg_search, kg_paths, etc.).

The prove-claim skill teaches the agent a structured workflow for investigating claims: decompose into entities, find them via search, trace connections via path traversal, read the evidence, assess and report with citations.

No LLM inside the tool — the agent does the reasoning, the tool provides the data infrastructure.

How it works

• Parser: Walks the vault for .md files, extracts YAML frontmatter (via gray-matter), wiki links, inline #tags, and enclosing paragraphs as edge context. Handles malformed YAML gracefully.
• Store: SQLite with sqlite-vec for vector search and FTS5 for full-text search. Single file database.
• Embedder: Xenova/all-MiniLM-L6-v2 via @huggingface/transformers. Runs locally, 22MB quantized model, 384-dimensional embeddings computed from title + tags + first paragraph.
• Graph: graphology for in-memory graph algorithms — Louvain community detection, betweenness centrality, PageRank (with degree centrality fallback for disconnected graphs), BFS traversal, all-simple-paths via DFS.
• Indexing: Incremental by default — tracks file mtimes, only reprocesses changed files. Community detection re-runs on the full graph. Use --force for a full rebuild.

Tech stack

Role	Package
Graph algorithms	graphology
Persistence	better-sqlite3
Vector search	sqlite-vec
Full-text search	SQLite FTS5
Embeddings	@huggingface/transformers
MCP server	@modelcontextprotocol/sdk
Tests	vitest (76 tests)

Known quirks

• sqlite-vec requires BigInt rowids when used with better-sqlite3.
• sqlite-vec KNN queries use WHERE embedding MATCH ? AND k = ?, not LIMIT.
• PageRank may fail to converge on large graphs with many disconnected components. Falls back to degree centrality automatically.
• Wiki link resolution uses Obsidian's "shortest unique path" algorithm. Ambiguous links (same filename in multiple directories) resolve to the first match with a warning.

License

MIT