The Knowledge Graph

The CRG (Contextual Resource Graph) is the core IP of ConCRG. It's a structured representation of everything ConCRG knows about your application — pages, elements, workflows, roles, and the relationships between them.

Data Model

RDF Triples

All knowledge is stored as RDF triples: subject → predicate → object.

"Invoice page"   → "navigates_to"    → "/invoices/new"
"Delete button"  → "requires_role"   → "admin"
"Deal"           → "belongs_to"      → "Pipeline"
"Settlement"     → "has_workflow"    → "Bond Settlement Workflow"
"Create Invoice" → "requires"        → "Contact"

Each triple carries metadata:

source — which training source produced it (probe, code, docs, chat)
confidence — 0.0 to 1.0 score
route — which page it was extracted from
timestamp — when it was created
category — semantic category

Deduplication: Exact matches on (subject, predicate, object) are collapsed. On conflict, the triple with the highest confidence and most recent timestamp wins.

ApplicationSitemap

Triples are assembled into a structured ApplicationSitemap:

interface ApplicationSitemap {
  appId: string;
  version: number;
  generatedAt: string;
  lastUpdated: string;

  pages: Record<string, Page>;
  sections: Section[];
  roles: Role[];
  roleHierarchy: RoleInheritance[];
  workflows: Workflow[];
  triples: Triple[];
}

Each Page captures:

interface Page {
  route: string;
  title: string;
  category: string;
  elements: UIElement[];
  parent?: string;
  children: string[];
  breadcrumbs: string[];
  requiredRoles: string[];
  visibilityConditions: string[];
  discoveredVia: 'probe' | 'code' | 'docs' | 'chat';
}

Persistence

Knowledge is stored locally by the Train Service with no external dependencies required. For production deployments, knowledge optionally syncs to a Neo4j graph database, enabling persistent storage, faster queries, and change tracking over time. Sync is triggered from the training panel and can be full (rebuild) or incremental (delta only).

Graph RAG vs. Flat RAG

ConCRG uses Graph RAG rather than flat vector search. Here's why it matters:

Question	Flat RAG	Graph RAG
"What is an invoice?"	Retrieves invoice description chunk	Same
"How do I get to invoices?"	May not find navigation path	Traverses `navigates_to` edges
"What roles can delete contacts?"	Keyword match on "delete"	Traverses `requires_role` edges on Delete elements
"What's the flow from deal to invoice?"	May miss multi-hop connection	Follows relationship chain

Graph RAG excels at relational and navigational queries — exactly the kind users ask about complex enterprise software.

Viewing the Knowledge Graph

During training, the sidecar UI shows a live count of facts extracted per source — pages discovered, workflows mapped, and total triples. You can inspect, filter, and delete triples from the Knowledge tab of the training panel.

Data Model​

RDF Triples​

ApplicationSitemap​

Persistence​

Graph RAG vs. Flat RAG​

Viewing the Knowledge Graph​