11mirror-graph-reasoning

Graph Reasoning — When to use the Knowledge Graph

When an agent needs to understand RELATIONSHIPS between entities (people, vendors, cost centers, documents), route to the graph engine. For direct content retrieval, use brain search first.

Decision Protocol

1. First: ALWAYS try fbrain search (fast, <1s, hybrid vector+keyword)

   • Works for: "find me X", "what is X", "show me the doc about X"

2. Then: If the question is about CONNECTIONS, use graph_query (5-20s, uses LLM)

   • Works for: "how is X connected to Y?", "what depends on X?", "trace the approval chain"
   • Works for: "show all entities related to cost center CC-4401"
   • Works for: "what is the provenance of this EBITDA number?"

3. Never skip brain search. Even for graph questions, brain search gives context.

Tools

Tool	When	Latency	Cost
search	Any content question	<1s	Free (no LLM)
query	Semantic + keyword hybrid	<1s	Free (no LLM)
graph_query	Entity relationships, multi-hop	5-20s	~$0.005 (LLM synthesis)
entity_search	Find specific entities in graph	3-10s	~$0.003
get_page	Read full document	<1s	Free
traverse_graph	Walk typed edges from a node	<1s	Free (no LLM)
get_backlinks	Find pages linking to a slug	<1s	Free

Rules

• NEVER call graph_query for simple "find me" questions (waste of tokens)
• ALWAYS cite the source entity/page in your answer
• If graph_query returns empty, fall back to brain search results
• If both return empty, say "I don't have information about this"
• Prefer traverse_graph over graph_query when you already know the starting entity slug
• Use get_backlinks to discover reverse relationships cheaply before escalating to graph_query

Routing Examples

Use brain search (fast path)

• "What is Project Aurora?" -> search("Project Aurora")
• "Show me the Q3 board deck" -> search("Q3 board deck")
• "What did Alice say about hiring?" -> search("Alice hiring")

Use graph query (relationship path)

• "How is Alice connected to Project Aurora?" -> search("Alice") then graph_query
• "What depends on vendor Acme Corp?" -> search("Acme Corp") then graph_query
• "Trace the approval chain for this budget" -> search("budget") then graph_query
• "Show all entities related to cost center CC-4401" -> entity_search("CC-4401") then graph_query

Use traverse_graph (known entity path)

• "Who does Alice work with?" -> traverse_graph(slug="people/alice", type="works_with")
• "What companies has Fund A invested in?" -> traverse_graph(slug="funds/fund-a", type="invested_in")

Contract

This skill guarantees:

• Brain search is attempted before any graph query
• Every answer includes source attribution (page slug or entity ID)
• Graph queries are only used when relationship/connection semantics are needed
• Latency and cost are minimized by routing to the cheapest sufficient tool