[](https://hub.docker.com/u/trustgraph) [](https://pypi.org/project/trustgraph/)  
[](https://discord.gg/kT5dAsaj8v)
[**Playground**](https://docs.google.com/forms/d/e/1FAIpQLSeTnF22ZjUP20FWV--VvS5606x-5cOvnKty6AqcPdtlnPuqbQ/viewform) | [**Self-Host**](https://config-ui.demo.trustgraph.ai/) | [**Docs**](https://docs.trustgraph.ai) | [**YouTube**](https://www.youtube.com/@TrustGraphAI?sub_confirmation=1) | [**Discord**](https://discord.gg/sQMwkRz5GX) | [**Website**](https://trustgraph.ai)
### Business meaning in the graph. Business authority in the runtime.
**Open Source · Open Standards · Total Transparency**
---
[TrustGraph](https://trustgraph.ai) turns enterprise knowledge, systems, and business logic into an ontology-grounded operating graph connected to a policy runtime that governs what AI agents can do.
Agents propose actions. TrustGraph checks them against business meaning, authority, and evidence before execution. Each decision preserves the evidence and policy version behind it as queryable W3C PROV-O provenance.
### Understand. Govern. Verify.
- **Understand your business**: Build a connected operating graph of entities, relationships, policies, and evidence from documents and systems, using your own or domain-standard ontologies.
- **Govern agent actions**: Enforce explicit rules for authority, scope, risk, uncertainty, time, and value at the point of action.
- **Verify decisions**: Retain the evidence and policy versions needed to explain and reproduce each outcome.
Built on [W3C](https://www.w3.org) [RDF 1.2](https://www.w3.org/TR/rdf12-concepts/), [OWL](https://www.w3.org/OWL/), [SHACL](https://www.w3.org/TR/shacl/), and [PROV-O](https://www.w3.org/TR/prov-o/), TrustGraph keeps your business knowledge portable. Deploy with Docker or Kubernetes, in the cloud or on-prem, using hosted LLMs or open-weight models with the native TrustGraph inference stack.
## The Problem: "Common Semantic Understanding"
To understand why AI struggles in complex use cases, consider Abbott and Costello’s classic ["Who's on First?"](https://www.youtube.com/watch?v=sYOUFGfK4bU) routine.
Abbott explains the baseball lineup: `Who` is on first base, `What` is on second base, and `I Don't Know` is on third base. Costello is driven mad because he assumes Abbott is asking questions rather than stating the names of the players: `Who`, `What`, and `I Don't Know`.
Two agents cannot communicate if they do not share the same context understanding.
## Why Vector Embeddings and Keyword Search Fail Here
If you feed this scenario into a standard RAG pipeline using vector embeddings or keyword search, it breaks completely.
If a user asks: "*Who is playing on first base?*"
1. The vector database converts the query into an embedding.
2. Cosine similarity searches for vectors close to "playing," "first base," and "who."
3. Because "Who" is a common pronoun, the embedding space maps it to general inquiries about identity, not the specific name of a baseball player.
4. The LLM retrieves irrelevant documents and hallucinates, failing to understand that "Who" is an entity (a Person), not a question.
Cosine similarity operates on fuzzy, statistical probability. It cannot distinguish between the linguistic usage of a word as a pronoun and its usage as a proper noun within a specific, localized context.
## Put Ontologies in Action
An ontology defines the concepts, unique definitions, and relationships that matter in a domain. In the “Who’s on First?” example, it distinguishes a player’s name from their position, so “Who” can be the name of the player at first base without being mistaken for a question. It gives knowledge extraction and retrieval a shared understanding instead of leaving a LLM to infer entity types, definitions, and relationships from each text chunk.
TrustGraph puts that shared understanding to work. Bring an OWL ontology, build your own, or start with a domain standard. For each chunk of source text, TrustGraph retrieves the relevant part of the ontology to guide extraction, turning raw data into more consistent, ontology-enhanced knowledge. Native natural-language retrieval then makes that knowledge available to agents, with its provenance intact.
## The Operating Graph
A semantic intelligence layer built using standards like RDF and OWL, establishes explicit, unambiguous semantics. It doesn't rely on "guessing" based on word proximity; it relies on defined relationships.
Here is the "Who's on First" routine modeled in RDF with an OWL ontology. By structuring data this way, the LLM knows exactly what "Who" means in this context:
```turtle
@prefix :