Knowledge graphs for LLM applications

What is a knowledge graph in an LLM application?

Nodes are entities (people, accounts, products); edges are relationships (owns, reports to, renewed on). An LLM can extract triples from text into the graph, and query the graph to ground answers.

What a knowledge graph does well

• Multi-hop questions: “Who manages the account that raised this ticket?”
• Consistency: one node per entity, not many near-duplicate passages.
• Explainability: the path through the graph is the reasoning.

What it costs to build

• Designing an ontology (entity and edge types) and evolving it.
• Entity resolution: deciding two mentions are the same thing.
• Extraction quality monitoring as models change.
• Operating a graph database.

Adding time to a knowledge graph

Static graphs store what is true now. For agents, edges need validity intervals so that a changed relationship closes the old edge rather than deleting it. This is the idea behind temporal graph approaches such as Graphiti, and behind Auld’s validity intervals on memories.

Do you need to design an ontology?

Not always. Start with a small default set of entity types and let extraction propose relationships; customize only where your domain needs precise types (for example, legal parties or financial instruments). Auld offers entity type and edge customization on Enterprise.

Frequently asked questions