Context Graph vs GraphRAG
Better Retrieval Is Not Governed Decisions
GraphRAG fixed a real weakness of vector-only retrieval. By extracting entities and relations into a graph and summarizing communities, it lets models answer questions that require connecting facts across documents, not just finding similar passages.
Because both approaches put “graph” and “context” in the same sentence, they get conflated. They should not be.
GraphRAG improves what a model knows: it is a retrieval architecture, and its output is better grounding for an answer. A context graph governs what an agent may do: it is a decision architecture, and its output is a judgment on a proposed action, with the trace to prove it. Knowing is not deciding.
The Core Distinction
The graph structure is shared. The semantics are not. GraphRAG edges encode relatedness: this entity co-occurs with that one, this passage supports that claim, this community of nodes shares a theme. Those semantics rank retrieval. They cannot compute validity.
Context graph edges encode governance: this policy version applies to this entity class over this time window, this exception supersedes that default for this customer, this fact came from the system with authority over it. Those semantics answer a different question: not “what is known and related” but “what is valid, applicable, and authorized right now, for this action.”
Side-by-Side Comparison
| Dimension | GraphRAG | Context Graph |
|---|---|---|
| Core question | Which entities, relations, and passages best ground this answer? | Is this proposed action valid now, in this scope, under these rules? |
| Graph semantics | Relatedness: entities, co-occurrence, communities, summaries | Governance: applicability, temporal validity, exceptions, scope, provenance |
| Control point | Retrieval step before generation | Per-action decision boundary before execution |
| Primary artifact | Ranked, structured context for the model's answer | Applicability result, allow, escalate, or block decision, causal decision trace |
| Failure caught | Ungrounded answers, missed connections across documents | Well-grounded actions that are still invalid: stale entitlement, superseded policy, out-of-scope write |
What GraphRAG Does Well
| GraphRAG capability | Good at | Does not prove |
|---|---|---|
| Entity extraction | Building an entity graph from unstructured corpora | Which of those entities' facts are currently valid and governing |
| Community summaries | Answering global questions across a large corpus | Whether an answer justifies an action for this specific case |
| Graph traversal retrieval | Following relations that vector similarity misses | Applicability of a rule to an entity at decision time |
| Hybrid ranking | Combining structural and semantic relevance signals | Exception hierarchies: which rule supersedes which, for whom |
| Provenance of passages | Citing which documents grounded the answer | Provenance of authority: which source governs the decision |
Why Relevance Is Not Applicability
Retrieval, however structured, optimizes relevance: the best available grounding for a response. Applicability is a stricter property: the rule that governs is the one that applies to this entity, in this scope, at this moment, after exceptions are resolved.
GraphRAG can retrieve three policy documents that all mention refunds. Applicability logic determines which one is in force for this customer today, and whether an exception supersedes it. Retrieval hands the model candidates. Governance requires a verdict.
The same gap appears in provenance. GraphRAG cites which passages grounded an answer. A causal decision trace records which authoritative facts, policy versions, and exceptions produced a decision, in a form an auditor can replay. Citing sources is not the same as proving authority.
Production Scenarios
Refund decision
GraphRAG: GraphRAG retrieves the customer's history, the product's return policy text, and related support tickets, giving the model rich grounding to draft a resolution.
Context graph: The decision context graph evaluates the proposed refund itself: the return window state, the policy version in force, the entitlement's currency, active exceptions, and scope, and returns allow, escalate, or block with a trace.
Contract question vs contract action
GraphRAG: Asked what discount terms a customer has, GraphRAG traverses from the account entity to its master service agreement and answers with the negotiated schedule.
Context graph: When the agent moves to apply a discount, the context graph determines which schedule version governs today, whether a legal exception is active, and whether this deal is inside the agent's decision scope, before the CRM write happens.
Incident response
GraphRAG: GraphRAG connects an alert to related services, past incidents, and runbooks, assembling the best available picture for the on-call agent.
Context graph: Before the agent executes a rollback, the context graph checks the maintenance window, the change freeze, the blast-radius scope, and the approval state of the runbook version, and records why the action was allowed.
Where This Fits in the Agent Stack
GraphRAG belongs in the knowledge path: it raises the quality of what an agent believes while reasoning, and for corpus-scale question answering it is a real advance over similarity search alone.
A decision context graph belongs in the action path: at the decision boundary, where each proposed write is evaluated against governing state through pre-execution enforcement before it reaches a system of record.
Agents that act need both paths, and they need them kept distinct. Better retrieval reduces wrong beliefs. Only a decision boundary prevents wrong actions.