Knowledge Graph

The knowledge graph maps every connection in the wiki — explicit [[wikilinks]] and relationships inferred by the agent during ingest. It uses NetworkX for graph analysis and Louvain community detection to cluster related pages, then renders everything as a self-contained HTML file you can open in any browser.

Building the graph

Tell the agent build graph or run the script directly:

build graph                                   # via the agent
python tools/build_graph.py                   # standalone
python tools/build_graph.py --report --save   # with health report saved to wiki
python tools/build_graph.py --tag project-x   # single project subgraph

The script scans every Markdown file in wiki/, extracts [[wikilinks]] from the content, builds a directed graph, then hands it to print_graph.py which renders the interactive visualization at graph/graph.html.

CLI flags

Flag	What it does
--report	Print a full health report to the console
--save	Write the report as a wiki page (graph/report.md)
--json	Output the report as machine-readable JSON
--tag project-x	Build the graph for a single project (comma-separated for multiple)

Tag filtering

With --tag you can build a subgraph for a single project or a subset of projects. Pages are matched by their YAML frontmatter tags. Comma-separate multiple tags to include several projects in one graph:

python tools/build_graph.py --tag project-alpha
python tools/build_graph.py --tag project-alpha,project-beta --report

This is useful for comparing the density and health of different knowledge domains, or for generating project-specific reports to share with stakeholders.

Graph analysis

When you pass --report, the script runs a full structural analysis and produces a health summary with actionable findings:

Metrics

Metric	What it detects
Orphan nodes	Pages with zero connections — isolated knowledge that should be linked or removed
God nodes	Pages with degree > mean + 2×stdev — overly central hubs that may need splitting
Fragile bridges	Edges whose removal would disconnect parts of the graph — single points of failure
Phantom hubs	Nodes with high betweenness centrality but low degree — hidden structural bottlenecks
Communities	Clusters detected via Louvain algorithm — thematic groupings of related pages
Density	Ratio of actual edges to possible edges — measures overall interconnectedness

The report ends with a suggested actionssection — concrete steps the agent can take to improve graph health (e.g. “link orphan page X to entity Y”, “split god node Z”).

Health thresholds

Overall graph health is determined by the ratio of edges to nodes:

A healthy wiki should have at least 2 edges per node on average. Below 1.0, many pages are isolated or barely connected — a sign that wikilinks are being missed during ingest.

Machine-readable output

Pass --json to get the full report as structured JSON — useful for CI pipelines, dashboards, or scripted health monitoring:

python tools/build_graph.py --json

The JSON includes the same data as the text report: health level, all metrics, per-community breakdown, orphan/god/bridge/phantom lists, and suggested actions.

Interactive visualization

The graph renders as a self-contained HTML file at graph/graph.html — open it in any browser, no server needed. Built with vis.js and a JanusLM-branded interface:

graph.html UI

The interface includes a branded header with the report date, a type legend (Source, Entity, Concept), and a theme toggle. The left panel combines search with the graph insights dashboard. The entire UI is fully customizable — since the HTML is generated by a Python script, you can ask any LLM coding agent to restyle it, change the branding, add features, or adapt the layout to your needs.

UI features

Interactive elements

Feature	Description
Search	Real-time node filtering — type to highlight matching nodes and fade the rest
Node drawer	Click any node to open a detail panel with type, community, file path, content preview, and related nodes as clickable chips
Graph Insights	Collapsible stats panel showing health badge, node/edge counts, orphans, communities, god nodes, fragile bridges, density
Type legend	Color-coded circles for Source (gold), Entity (dark), Concept (light) — click to filter by type
Light/dark theme	Toggle between light and dark mode — persists across sessions
Status bar	Bottom bar showing live node/edge count

Node types

Every wiki page becomes a node, colored by its type:

Node size scales with degree (number of connections). Community membership is shown in the node drawer when you click a node.

Node detail drawer

Clicking any node opens a slide-in drawer on the right side showing:

• Title — the page name
• Type and community — color-coded badge and community ID
• File path — where the page lives on disk
• Content preview — the first portion of the Markdown content, rendered with wikilink highlighting
• Related nodes — clickable chips that navigate the graph to connected pages

Click the background to close the drawer and reset the graph view.

Layout engine

The visualization uses the Barnes-Hut physics simulation with parameters that adapt based on node count. Smaller graphs get stronger repulsion and tighter springs for a compact layout; larger graphs relax the physics to keep the visualization readable. The simulation stabilizes automatically, then disables physics to reduce CPU usage.