The one-time setup
- Open the causal graph.
- Settings → Connected sources → Add — pick the documents, folders, or memory spaces this graph is derived from.
- Save.
Where the graph is used
Two consumers:- Verifier — always on. The graph’s approved (LIVE) edges are folded into grounding checks for every answer.
- Retrieval — governed by the graph’s dedup policy (see below). Not automatic like verification.
Dedup policy — how the graph shows up in retrieval
Because the graph is derived from the connected documents, retrieving both the raw chunks and the graph’s facts would duplicate information. The graph’s dedup policy decides how they mix.
Configure per graph: Graph → Settings → Dedup policy.
The default
dedup_by_source matches “docs are the primary evidence, graph adds the causal logic on top” — the rule for how the graph should behave in retrieval.
What the graph’s role really is
- The verifier’s primary consumer — always folds in approved edges to check factual claims.
- Retrieval’s auxiliary logic — supplies the causal edges that documents alone don’t state, without duplicating chunks the retrieval already found.
Multiple graphs
An Agent can end up referencing multiple graphs — one per set of overlapping connected sources. Each graph’s dedup policy applies independently to its own contribution.Configuring from the CLI
nora causal for the full command surface.
Legacy explicit binding
For backward compatibility, the explicitcausal_graph data-source binding still works — an Agent that has an old-style causal-graph attachment short-circuits to its prior path. New Flows should use the connected-source model instead.
What to do next
Verify agent answers
How the verifier uses the graph — always on.
Grounding (retrieval)
How retrieval uses the graph — via dedup policy.