Skip to main content
Building a causal graph from scratch is a lot of typing. If your domain knowledge already lives in documents or a refinery, let Nora propose the structure and keep the review for yourself.

​
Starting points

An empty graph offers three ways to start:
  • Add by hand pure manual authoring.
  • From a folder pick a Documents folder and a chat model; Build ontology has the LLM bootstrap variables and edges from that folder’s content.
  • From a refinery pick a refinery and a model; its blocks seed the ontology.
Both LLM paths need a chat model: save a provider key under Settings → External LLMs and the default is picked up automatically.

​
Create with AI

The Create with AI button in the toolbar opens a suggest panel on any graph. Type what you want in plain language, such as “sketch a customer churn model” or “what edges are missing from this graph?”, and hit Ask. The panel returns node and edge suggestions as cards, each with the LLM’s rationale. Click Accept and they land in the graph the same way a manual add would.

​
Review and approve

LLM-built nodes and edges arrive as proposals, not live items:
  • Each proposal carries its origin and confidence in the node’s Evidence tab.
  • Approve items individually in their Review tab, or approve the staged batch with the Approve all button that appears in the toolbar when proposals are pending.
  • Only approved (live) items are used by verification and retrieval.
Two controls keep re-runs from trampling your curation:
  • Allow auto-edit (per node) lock a node and LLM re-runs won’t overwrite it.
  • Reviewed (per edge) reviewed edges are preserved through LLM auto-restructuring.

​
Evidence stays attached

When a pipeline ingests documents into the graph, the attach pass surfaces source chunks on each node’s Evidence tab. Every variable can point back to its supporting passages, and Search this subtree can build a retrieval preset from exactly that evidence. The staged/live workflow is also scriptable: see nora causal.