Live knowledge graph

While a discussion runs, FluxyChat can extract ideas, tasks, decisions, and contradictions into a queryable graph. The graph is a semantic index, not source of truth: every node links to the message that introduced it.

Enable

On the Worker:

KNOWLEDGE_GRAPH_ENABLED=true

Configure an LLM provider (same as other AI features). Tenants can opt out via hosted LLM policy.

Extract

POST /rooms/:roomId/kg/extract
Authorization: Bearer <member-jwt>
Content-Type: application/json

{}

Optional body: { "since": "2026-08-13T10:00:00.000Z" } to limit the transcript window.

Response:

{
  "nodesExtracted": 4,
  "edgesExtracted": 3,
  "nodesInserted": 2,
  "edgesInserted": 1
}

Extraction is idempotent: duplicate nodes/edges are skipped; high-confidence updates supersede older nodes with provenance preserved.

Query

GET /rooms/:roomId/kg?limit=80
Authorization: Bearer <member-jwt>

Returns nodes, edges, and stats (counts by type).

Dashboard (reagraph)

Open Rooms → Live knowledge graph. The console uses reagraph (WebGL) for force-directed layout:

• Color by node type (task, decision, concept, user, …)
• Click a node to see confidence, source message id, and properties
• Drag nodes; curved directed edges with labels

Node and edge types

Related

• Room memory (PH-132)
• Community decision simulation (PH-131)