What is TrustGraph?

The problem

AI agents are transforming how organisations work. But giving an AI agent access to your knowledge — and getting accurate, trustworthy answers back — is harder than it looks.

The standard approach is Retrieval Augmented Generation (RAG): chunk your documents into pieces, embed them as vectors, and retrieve the most similar chunks when a question is asked. It works for simple cases. But it breaks down fast:

Too much text causes hallucination. When you dump large chunks of text into an LLM context window, the model struggles to find the relevant signal. It starts making things up — confidently, convincingly, and wrong.

No control over what the LLM sees. Existing RAG frameworks retrieve text by similarity, but you have no say in what actually gets presented to the model. You get whatever the vector search returns — relevant or not. There’s no way to be precise about the knowledge an LLM reasons over.

Information management has no precision. Vector search gives you chunks of text, not structured knowledge. You can’t query for specific facts, relationships, or entities. You can’t ask “who reports to whom” or “which service depends on which” — you can only ask for text that looks similar to your question.

Consider the classic “Who’s on First?” comedy routine: a baseball team has players named Who, What, and I Don’t Know. Ask a vector search “Who is playing on first base?” and it breaks completely — the embedding space maps “Who” to a generic identity question, not the name of a player. Semantic similarity operates on statistical probability. It cannot distinguish between the linguistic use of a word as a pronoun and its use as a proper noun within a specific context.

No explainability. When the AI gives you an answer, you can’t trace why. What sources were used? What reasoning led to the response? In a demo this doesn’t matter. In production — especially in regulated, legal, or compliance-sensitive environments — it’s a dealbreaker.

The solution

TrustGraph is an open-source context orchestration layer that takes a fundamentally different approach to giving AI agents knowledge.

Instead of treating your documents as bags of text chunks, TrustGraph extracts structured knowledge — entities, relationships, and facts — and stores them in a context hypergraph alongside vector embeddings. A hypergraph goes beyond simple binary relationships (A → B) by connecting multiple entities into complex, real-world events — linking a document to its author, approving manager, compliance policy, and time/location metadata as a single conceptual unit.

When an agent asks a question, TrustGraph retrieves precise, structured context rather than raw text. Returning to “Who’s on First?” — a hypergraph knows that :Who is a :Player whose :playsPosition is :FirstBase. There is no ambiguity, no hallucination, because context is structured, not inferred via probability.

This means:

  • Reduced hallucination — the LLM receives focused, relevant knowledge instead of large blocks of loosely related text
  • Precise information control — you control exactly what knowledge is extracted and how it’s structured, using ontologies (BYOO — Bring Your Own Ontology) to define what matters in your domain
  • Relationship-aware retrieval — answers can follow connections across your knowledge: who owns what, what depends on what, how things relate
  • Full explainability — every answer can be traced back to the specific hypergraph nodes, source documents, and reasoning steps that produced it — not just traceable, but cryptographically verifiable

TrustGraph isn’t just a better RAG pipeline. It’s the context orchestration layer for building AI systems that reason accurately over your organisation’s knowledge — and can prove how they got there.

Next

How does TrustGraph work? — the approach and architecture.