Why it exists
Vector search finds similar chunks. A knowledge graph answers how things connect: who imports whom, which package is a hub, and the path between two symbols — with EXTRACTED vs INFERRED provenance on every edge.
Open it
- Toolbar ◈ button
- Files panel → Open knowledge graph
- Command palette →
/nj-open-knowledge-graph
What you get
- Colored communities (package/directory clusters)
- God nodes ranked by degree
- Search, path mode (pick two nodes), explain inspector
- Double-click to open the file in the editor
- Ask agents about a selected node (composer prefill)
Agent grounding
Questions like “how does X relate to Y” or “path between A and B” route through the knowledge graph retriever and inject neighborhood/path context alongside @codebase search.
Local-first
Structure is built on-device from symbols and imports (Go, TypeScript/JavaScript, Python, Rust). Graphs live under ~/.neural-junkie/code-graph/. No third-party Graphify dependency.