A semantic model
A shared definition of the entities, relationships and terms an AI system should use, kept as small as the questions require.
Services
AI context layer: knowledge graphs for LLMs and AI agents
An AI system is only as reliable as the context it retrieves. We build a semantic context layer on a knowledge graph that turns fragmented enterprise data into structured, source-traceable context for LLMs, RAG systems and AI agents. Entities, relationships and definitions are explicit, so answers can be connected, checked and explained.
Built on our knowledge graph research: ReviewGraph, presented at IEEE ICKG 2025, and our design work on a knowledge graph foundation model. See ReviewGraph research →
When an AI system needs a context layer
- An LLM retrieves passages that look similar, but the answer depends on how things are connected.
- The same term means different things in different systems.
- Answers must be traceable to a source before people will trust them.
- Known facts and model predictions get mixed together in a single answer.
What we build
A knowledge graph with provenance
Records and documents connected into a graph, with the source kept for each fact.
Retrieval interfaces
Query interfaces and APIs that return connected context, with its sources, to an LLM or agent.
Separation of evidence and prediction
Known facts and model estimates stored and labelled separately, so an answer shows which is which.
Graph-based prediction
Embeddings and link prediction on the graph, when the layer should also estimate what is missing or likely.
Use cases
Questions across several sources
Answer questions that span systems and show the source behind each part of the answer.
Grounding for AI agents
Give an agent a defined vocabulary and connected facts to plan and act on, instead of raw documents.
A shared vocabulary
One definition of key terms and entities that teams and systems can use consistently.
Explainable predictions
Show which entities and relationships a prediction relied on, as ReviewGraph does with its graph visualisation of reviews.
Research and evidence discovery
Connect publications, entities and evidence and keep known evidence apart from model predictions, as in Vink Drug.
How it works
- 01
Start from the questions
List the questions the AI system must answer and the decisions it supports.
- 02
Model and connect
Define the entities and relationships those questions need, and connect the sources with consistent identities and provenance.
- 03
Expose as context
Make the graph queryable so an LLM or agent can retrieve connected facts and their sources when it answers.
- 04
Evaluate and maintain
Test answers against cases where the correct answer is known, and keep the graph current as the data changes.
Every project follows the same four steps. Understand → Build → Validate → Deploy. About our way of working →
Where the approach comes from
Research
ReviewGraph
The paper names interpretability, visual exploration and potential integration into retrieval-augmented generation as advantages of the graph approach.
Read the ReviewGraph research →Research
Vink Drug
Biomedical knowledge graphs that keep known evidence apart from model predictions.
See Vink Drug ↗Research
Knowledge graph foundation model
Our current design for a model that infers how an unseen graph behaves from its structure. A vision, not yet a result.
Read the vision →Frequently asked questions
What is an AI context layer?
The layer between your data and an AI system. It holds the entities, relationships and definitions the AI should use, together with their sources, and serves them as context when the AI answers or acts. We build it with a knowledge graph and a semantic model.
How is this different from RAG over documents?
Standard RAG retrieves text passages that resemble the question. A context layer adds explicit entities and relationships, so the system can retrieve connected facts, combine information from several sources and show where each part came from. The two can be combined.
Are knowledge graphs useful for AI agents?
Yes. An agent needs to know what things exist, how they relate and what the terms mean. A graph provides that as structured context and, because facts keep their source, answers and actions can be checked.
Do we need an enterprise ontology first?
Not a large one. We define the smallest semantic model that answers your questions and extend it as new use cases appear.
How do you keep predictions separate from facts?
Known evidence and model estimates are stored and shown separately, so an answer states which is which. We apply this principle in the biomedical graphs of Vink Drug.