Website | Docs | YouTube | Launch TrustGraph | Discord | Blog

The open source ontology runtime for AI agents

Every platform now sells you an ontology. TrustGraph is the one you own.

For enterprises deploying AI agents where decisions carry real risk, TrustGraph is the open-source ontology runtime that turns your documents and data into ontology-grounded knowledge, and governs what every agent and user can see from it.

Unlike other platform ontologies, TrustGraph is built on W3C standards (RDF, OWL, SHACL, PROV-O) for total ownership and portability of your knowledge and business logic. Your ontology, your knowledge, and your audit trail stay yours, portable across any cloud, model or deployment.

            Your ontology                       Your documents & data
                 │                                       │
                 └──────────────────┬────────────────────┘
                                    ▼
                         ┌──────────────────────┐
              1. EXTRACT │     OntologyRAG      │  Triples match classes and properties,
                         │                      │  checked against constraints
                         └──────────┬───────────┘
                                    ▼
                         ┌──────────────────────┐
              2. GROUND  │    Context Graph     │  RDF knowledge in named graphs,
                         │                      │  with PROV-O lineage to every source
                         └──────────┬───────────┘
                                    ▼
                         ┌──────────────────────┐
              3. GOVERN  │  SHACL-AF Policies   │  Every query evaluated per node:
                         │                      │  Allowed · Filtered · Violation
                         └──────────┬───────────┘
                                    ▼
                     GraphRAG · SPARQL · Agents · MCP
           + a queryable trace of how each answer was produced

Additional Technical Diagrams

• Knowledge is stored as RDF, including language tags, XSD datatypes and RDF-star quoted triples.
• Load knowledge as Turtle or TriG (named graphs preserved); export the graph as Turtle.
• Package knowledge into Context Cores that carry triples, embeddings, provenance and source documents between deployments.
• Move whole workspaces between deployments with portable .tgx bundles (tg-export-workspace / tg-import-workspace).

• OntologyRAG uses your ontology to guide extraction. A similarity-based selector picks the relevant part of the ontology for each chunk, with a bypass mode for small ontologies.
• Extracted triples are validated against ontology domain and range constraints, including the subclass hierarchy. Triples that break the rules are rejected.
• rdf:type attributes are stored with graph embeddings, so you can filter vector searches by entity type.
• Schema-free GraphRAG and hybrid (BM25 + vector) DocumentRAG are also available, both with cross-encoder reranking.

• Write access policies as SHACL-AF shapes, stored in the urn:graph:policy named graph.
• Every graph query is evaluated node by node with a tri-state result: Allowed, Filtered or Violation.
• Policies apply across the REST API, SPARQL and GraphRAG, using a per-request user context.
• Each evaluation publishes a structured policy.evaluation audit event.
• With no policies loaded, the system runs in open mode.

• Supply grounding seeds or graph seeds to skip LLM concept extraction.
• Restrict each hop by type, relationship or named graph, and filter edges by language tag.

• Extraction emits PROV-O lineage from document to page, chunk and edge (urn:graph:source).
• GraphRAG, DocumentRAG and agent queries record full reasoning traces (urn:graph:retrieval).
• Inspect them with tg-show-explain-trace and tg-show-extraction-provenance, or query them with SPARQL.

• Decision Scoring (Laya, ONNX Runtime, ~35 ms per call): pick an option, score on an ordered scale, or get a yes/no confidence.
• Cloze Classification: auto-train a DistilBERT classifier from labeled examples and serve it with confidence scores.

• Agent orchestrator with ReAct, Plan-then-Execute and Supervisor patterns.
• MCP server with tools for graph queries, SPARQL, GraphQL over structured data and embeddings, with bearer-token authentication.
• SPARQL 1.1 query service with 30+ built-in functions and streaming results.
• REST, WebSocket, Python SDK, TypeScript libraries and a full tg-* CLI.

• Have Questions? Join our Discord
• Found a Bug? Open an issue
• Need Help? Check the documentation
• Ready to Contribute? See the contributing guide