TrustGraph is an ontology-native knowledge platform that turns source documents and data into typed, traceable facts for natural-language analytics and AI agents. It targets cases where vector search over isolated text chunks leaves agents without the structure or provenance to answer consistently, such as a name that the embedding space mistakes for a common word.
Knowledge is stored as RDF 1.2 named graphs governed by OWL ontologies. You bring an existing OWL ontology, build one, or start from a domain standard, and for each chunk of source text TrustGraph retrieves the relevant part of the ontology to guide extraction. Queries run through SPARQL or GraphRAG over explicit relationships, with automated natural-language retrieval, semantic filtering, and reranking. Extraction lineage, query traces, and agent behavior are recorded as W3C PROV-O triples, and the graph exports as Turtle to any RDF-compliant system.
Agents run either on the native TrustGraph Agent Runtime, which traces each decision back to its source knowledge, or on your own framework through the TrustGraph API Gateway. Workspaces, Collections, Flows, and Knowledge Cores isolate, version, and load knowledge at runtime. The stack bundles its own storage and messaging (Cassandra, Qdrant, Garage, and Pulsar or RabbitMQ) and an inference layer that calls hosted LLM APIs or self-hosts open-weight models through vLLM, TGI, Ollama, LM Studio, or Llamafile. It deploys as Docker containers locally or on Kubernetes.
| Dimension | TrustGraph | Conventional graph databases (such as Neo4j) |
|---|---|---|
| Data model | RDF 1.2 named graphs with statement reification | Labeled property graph |
| Semantics | OWL ontology enforcement | Schema-optional, semantics in application code |
| Provenance | Built in, W3C PROV-O | Hand-modeled |
| Natural-language retrieval | Automated with filtering and reranking | Manual Cypher or add-on vector search |
| Unstructured ingest | Integrated | External ETL |
| Query language | SPARQL plus natural language | Cypher or GQL |
Features
- Ontology-guided extraction: OWL ontologies steer knowledge extraction from each chunk of source text
- RDF knowledge graph: RDF 1.2 named graphs with class hierarchies, property constraints, and logical inference
- GraphRAG and SPARQL: natural-language and SPARQL retrieval over explicit relationships
- Provenance: extraction lineage, query traces, and agent decisions stored as PROV-O triples
- Agent Runtime: a native runtime that traces agent behavior, or an API gateway for your own framework
- Knowledge Cores: reusable context bases loaded and removed at runtime, alongside Workspaces, Collections, and Flows
- Document ingest: PDF, DOCX, XLSX, PPTX, HTML, Markdown, CSV, and images
- LLM inference stack: hosted provider APIs or self-hosted open-weight models on Nvidia, AMD, or Intel hardware
- Web UI: agent console with live explainability, GraphRAG view, 3D context explorer, ontology and schema workbenches, and a prompt editor
- TypeScript libraries: client, React state, and React provider packages for building UIs
- Interoperable export: Turtle serialization portable to any RDF-compliant system
