Typed edges enable AI reasoning; embeddings only find proximity — label relationships explicitly
When AI systems traverse your knowledge graph, maintain typed relationship labels with explicit predicates rather than relying on semantic similarity alone, because typed edges enable logical reasoning while embeddings only surface associative proximity.
Why This Is a Rule
Semantic similarity (embedding-based retrieval) answers "what notes are about similar topics?" Typed edges answer "what notes contradict, extend, enable, or depend on this note?" These are fundamentally different capabilities, and only typed edges enable multi-hop logical reasoning.
When an AI traverses your graph using embeddings alone, it finds notes in the same topic neighborhood — but it can't trace causal chains ("this note causes that, which enables the other"), it can't identify contradictions ("these two notes disagree about X"), and it can't follow dependency paths ("this conclusion requires that evidence, which depends on that assumption"). These logical traversals require typed edges because the traversal logic differs by edge type.
This is Treat every note link as infrastructure for AI graph traversal (treat every link as AI infrastructure) with the specific implementation guidance: the links must carry type labels, not just exist. An untyped link between notes A and B tells the AI "these are related." A typed link "A contradicts B" tells the AI something it can reason about.
When This Fires
- When building or maintaining a knowledge graph that AI systems will query
- When AI retrieval returns topically related but logically unrelated results
- During graph architecture design for RAG or GraphRAG systems
- When upgrading from embedding-only retrieval to structured graph traversal
Common Failure Mode
Relying entirely on embedding similarity: "The AI will figure out the relationships from the content." It won't — embeddings find topical proximity, not logical relationship. Two notes about "cognitive load" are nearby in embedding space regardless of whether one supports the other, contradicts it, or is completely independent.
The Protocol
For AI-traversable knowledge graphs: (1) Maintain typed edge labels on every connection (Label diagram arrows with specific verbs — 'causes' and 'blocks,' not 'affects' and 'relates to', Limit knowledge graph relationship types to 5-7 — more precision costs more than it reveals). (2) Use explicit predicates: enables, contradicts, extends, supports, depends_on, exemplifies. (3) When creating a link, don't just link — label the relationship type. (4) The typed labels serve as instructions to AI traversal algorithms: "follow 'enables' edges for causal chains," "flag 'contradicts' edges for conflict detection." Untyped links serve only topical grouping.