Ontologies

Standard Graph RAG lets an LLM discover whatever entities and relationships it finds in your documents. This works surprisingly well for general-purpose knowledge. But when you’re working in a specific domain — legal, medical, financial, technical — you often need more control over what gets extracted.

That’s where ontologies come in. TrustGraph supports BYOO — Bring Your Own Ontology — loading your domain schema in OWL format to guide knowledge extraction.

What is an ontology?

An ontology is a structured description of the concepts and relationships that matter in a domain. Think of it as a schema for human knowledge — it defines:

  • Entity types — the kinds of things that exist (e.g. Person, Service, Role, Department)
  • Relationships — how those things connect (e.g. “manages”, “depends on”, “approves access to”)
  • Constraints — which relationships are valid between which entity types

If you’ve worked with database schemas, the concept is familiar. The difference is that ontologies describe knowledge about the world, not rows in a table.

Why use ontologies with RAG?

Without an ontology, Graph RAG extraction is opportunistic — the LLM extracts whatever it finds. This means:

  • Different documents may produce inconsistent entity types
  • Important domain-specific relationships might be missed
  • The LLM might focus on generic relationships rather than domain-relevant ones

With an ontology, extraction is guided. You tell TrustGraph what kinds of entities and relationships matter, and it extracts accordingly. The ontology-enabled hypergraph uses the provided ontology for semantic compliance on all ingested data, dramatically improving agentic accuracy and precision. The result is a more consistent, more precise, and more useful context hypergraph.

How TrustGraph uses ontologies

TrustGraph’s approach to ontology-guided extraction is pragmatic. Good ontologies can be large, and flooding an LLM context window with the full ontology would be counterproductive — exactly the kind of context overload that causes problems.

Instead, TrustGraph applies a retrieval operation on the ontology itself. For each chunk of text being processed, it selects the relevant subset of the ontology to guide extraction. This means the LLM always has the right context for the text it’s processing, without being overwhelmed.

Creating ontologies

Ontologies have a reputation for being complex and time-consuming to create. In academic and enterprise knowledge engineering, that reputation is deserved — formal ontology projects can take years.

But for practical RAG use cases, you don’t need a perfect, exhaustive ontology. You need one that’s good enough to guide extraction in your domain. Modern code assistants can generate a working OWL ontology from a description of your domain in minutes. You can refine it as you go.

TrustGraph supports OWL ontologies and includes tooling for loading, inspecting, and managing them through both the Workbench UI and CLI.

Next

How do people use TrustGraph? — real-world applications and scenarios.