Auto

⌘1, and the default. Auto exists because two kinds of question need two kinds of answer: “how many concepts are there?” is a count, and “what is said about surrender?” is a search for meaning. In Auto a router reads the question and picks.

With nothing embedded there is only one way to answer, so Auto is plain Text to Cypher and no routing call is made.

The four strategies

Route	Trail line	When the router picks it
cypher	Routed to Cypher	structure or exact facts: counts, “which X connect to Y”, filters on numbers or names, sorting, everything of a type, paths. A vector search is never a substitute for a count or a filter.
semantic	Routed to vector search	ideas or content in words that may not match the graph’s wording: “what is said about…”, “anything on…”, “passages about…”. The person wants the passages. Best when no specific node is named.
semantic_then_graph	Routed to vector search, then connections	the question is semantic but also asks how the matches connect, who is behind them, what relates: “what discusses X and what supports it”. Matches are found by meaning, then expanded one hop through their relationships.
graph_then_semantic	Routed to Cypher, then vector search	the question names specific things (a node, a type, an author, a source, a number range) to search within, then asks something conceptual about them: “what does the Gita say about duty”, “among papers before 1950, which discuss control”. A Cypher query returns the nodes; vector search ranks inside them.

What the router is told

The routing prompt gives the model the shape, the list of embedded properties (“Concept.summary: the summary of each Concept”), the four strategies as above, and these rules:

Prefer cypher when the question can be answered exactly from structure.
Prefer semantic or semantic_then_graph when the person is essentially asking “find me the relevant knowledge” and the wording is free.
semantic_query is the conceptual core of the question as a short standalone search text, in the person’s language, without instructions (“the relationship between soul and body”, not “find passages about…”). Empty for cypher.
scope lists the embedded properties worth searching, by name; empty means all; narrow it only when the question clearly points at some types and not others.
For graph_then_semantic, graph_question is a standalone question whose Cypher answer returns whole nodes to search within.
explanation is one plain sentence for the person.

The router runs at low effort on the Routing model (Settings, or the mode popover’s AI models tab).

What happens next

cypher: the Text-to-Cypher flow.
semantic / semantic_then_graph: the semantic query (or your question, if the router gave none) is searched within the router’s scope, which is never wider than what you ticked under Search in. For then connections, each matched node’s relationships are gathered: Following connections → Followed connections · 14 connected nodes, and the cards show connection chips.
graph_then_semantic: the graph question is translated and run; the records it returns become the candidates; vector search runs within them (“Ran vector search within 38 results”). If the query returns no nodes, the trail says Nothing to narrow to: the query returned no nodes or relationships, so the whole graph is searched instead.

The explanation under the trail is the router’s sentence; the trail’s first line is the decision with the model and time (“Routed to vector search · Sonnet · 1.1 s”).

Routes are remembered

A route is saved with the question in history, together with a fingerprint of the context (the shape and the embedded properties). The same question in the same context routes the same way without asking: Routed to Cypher · Sonnet · reused, not asked again, and a cached Cypher query is reused too. Ask Claude again under the explanation starts fresh.

Without the AI

If Claude Code or Codex isn’t ready and something is embedded, Auto doesn’t stop: Using vector search, as Claude isn’t set up · on this Mac, with a note to set it up for Cypher and routing, and an Open Settings button.

The router sees the question and the shape, never the graph. Its decision is a small structured answer: strategy, semantic query, scope, graph question, explanation.