The problem this solves: generic GraphRAG approaches extract whatever relationships an LLM thinks are relevant, which often misses domain-specific semantics. If you're working in healthcare, finance, or intelligence analysis, you likely already have ontologies (or can adapt standards like SOSA, FIBO, etc.) that define what matters. TrustGraph uses these to constrain extraction, so the resulting graph reflects your domain model rather than the LLM's interpretation.

How it works: The ontology defines classes and properties. During extraction, the LLM is prompted to identify instances of those classes and relationships matching those properties. The output is validated against the schema before being written to the graph store.

Built on Apache Pulsar for scalability, supports multiple graph backends (Memgraph, FalkorDB, others), and runs locally or in cloud. Apache 2.0 licensed.

Repo: https://github.com/trustgraph-ai/trustgraph

Ontology RAG docs: https://docs.trustgraph.ai/guides/ontology-rag/

Happy to answer questions about the extraction approach or architecture.