What Is a Context Graph?

A context graph is a data model that connects entities and relationships with temporal data, provenance, and decision history. A knowledge graph focuses on the Who, What, Where, while a context graph extends that with When, Why, and How.

Traditional RAG to raw data is like a blind person trying to navigate a room by touch. You can guess where the door generally is, but you'll waste plenty of time and effort to actually find it. A context graph is like an annotated map of the room. Not only do you know where the door is, but you know the best way to get there without bumping into any furniture.

A context graph in its simplest form is a collection of nodes and edges. The nodes are important nouns, like people, teams, customers, projects, documents, systems and decisions. The edges show how these items relate. For example, an edge shows the owner of a project or the customer of a support ticket.

Nodes and edges can have properties, like source, validity, and authority.

A knowledge graph and a context graph use the same data model. They have different purposes.

• A knowledge graph records general facts. For example, a product is in a category. People usually curate it, and its like a static snapshot of the world.
• A context graph records the current state of a company. It shows who does each task, what the team decided and what changed this week. It gets data from live systems, and it changes all the time.

A knowledge graph tells you what is true. A context graph tells you what is true here, now, for this person, and how we know. AI that does work in a business needs the second type.

A context graph usually gets its data from the systems where people do their work. These systems include email, chat, documents, tickets, CRM records, code and meeting notes. There are usually four steps to build a context graph:

• Connect the graph to the source systems. Keep the permissions of each source.
• Extract the entities and relationships from records and from text.
• Resolve duplicates. For example, “Acme” in the CRM and “Acme Corp” in a contract become one customer.
• Maintain the graph when the sources change. Record when the graph learned each fact, and from which source.

When an AI agent starts a task, it asks the graph for the related context. The graph gives connected facts with sources, not a large set of loosely related text. Agents can traverse the nodes and edges to leverage multihop reasoning and traceable outputs. This method is the base of context engineering.

Today, models can reason well. Inside a company, their main problem is missing information. Companies now move from chat assistants to agents that do actions. This makes the problem more serious. An assistant with bad context gives a weak answer. An agent with bad context makes a wrong decision.

Document search helps, but documents seldom tell who made a decision or if it is still current. A context graph gives this structure. For this reason, it is becoming the base of the enterprise context layer.

Do you want to see a context graph of your company?