Answers
What is a context graph?
A context graph is a connected model of a company's entities, metric definitions, and supporting evidence, maintained so that people and AI resolve the same question the same way every time. It ties each customer to one verified identity across systems, each metric to a single governed definition, and each number back to the records it came from — so an answer is looked up rather than inferred.
The three things it connects
Entities. One customer usually exists as several records — a CRM account, a billing subscriber, a support organisation, a product workspace — with no key that works across all of them. The graph holds the resolved identity, so the join is a fact rather than a guess made from a name or a domain.
Definitions. Each metric has exactly one meaning attached to it: what counts as an active account, which revenue figure finance would defend, how churn is measured and over what window. Without this, the same question returns different totals depending on who asked and when.
Evidence. Every figure keeps a path back to the rows, calls, and tickets behind it. This is what makes an answer checkable, and an answer nobody can check is not usable for a decision however accurate it happens to be.
How it differs from a knowledge graph
A knowledge graph models what is true in a domain: entities and the relationships between them, usually drawn from documents and reference data. It is a representation of knowledge.
A context graph is narrower and more operational. It models the agreements needed to interpret a specific company's data — identity, definition, freshness, provenance — and it is governed, because those agreements change and someone has to own them. The test is not whether a fact is in the graph but whether two different consumers asking the same thing get the same answer from it.
The practical consequence is that a context graph has to stay current. A knowledge graph can be a useful snapshot; a context graph that is a week stale will confidently answer questions about this quarter using last quarter's state.
Why AI needs one
A language model arrives with no knowledge of your company and no way to acquire the agreements above. Given access to your systems and a question, it reconstructs them on the fly: it decides which records belong together, picks a definition, and assembles a number. The reasoning is usually fine. The inputs it invented are the problem.
This is why the failure is quiet. The model returns a specific figure with a confident explanation, and nothing in the output distinguishes a verified join from an assumed one. Ask the same question twice, worded differently, in two sessions — if the numbers move, the model was reconstructing rather than looking up.
A context graph removes the reconstruction step. The agreements already exist, so the model retrieves them instead of inventing a fresh set per query, and two assistants pointed at the same graph agree with each other.
Where it sits in your stack
Between your systems and whatever consumes them. The warehouse remains the store, the CRM remains the system of record, and the catalog remains the inventory of what exists. None of those decide that two records are the same customer or that a metric has one enforced meaning, which is the gap the graph fills.
On the consumption side it is reached like any other tool. Riley speaks MCP, so Claude, Copilot, ChatGPT, and internal agents keep the connections they already have — what changes is that the other end of the connection is a governed graph rather than a set of systems the model is left to reconcile alone.
Related questions
Is a context graph the same as a semantic layer?
A semantic layer is part of one. It supplies governed metric definitions, which is one of the three pieces, but it does not resolve entity identity across systems and does not usually carry unstructured evidence like calls and tickets. A context graph covers all three and keeps the lineage attached.
Do we need a context graph if we already have a data warehouse?
Usually yes. Centralising storage puts the data in one place without establishing what it means — the same ambiguity about identity and definitions survives the migration. The graph treats the warehouse as a source and adds the agreements on top of it.