Connected ideas.
Practical intelligence.

A knowledge graph connects entities through meaningful relationships. Think of suppliers, products, contracts, and the dependencies between them. It gives AI a structured model of a business domain.

A context graph emphasises the circumstances surrounding a task or decision. It may connect goals, events, permissions, evidence, and earlier decisions. The term is evolving; this is a useful distinction, not a rigid boundary.

The two approaches complement each other. A knowledge graph can also contain history and context. A context graph brings those connections into focus for a particular situation.

In Neo4j, both can be represented using nodes, relationships, and properties. GraphRAG can combine semantic retrieval with graph traversal to assemble relevant evidence. An agent can use that evidence to choose a better next step.

Consider a delayed order: dependencies reveal affected customers, while context identifies applicable commitments. Together, these connections support more relevant answers and explainable actions. Their value still depends on current data, provenance, and enforced access controls.

A fixed RAG pipeline follows a predefined retrieval path. Agentic GraphRAG lets an agent select tools and adapt its investigation. Neo4j provides the connected data those tools query.

The agent first identifies what the question requires. A similarity search can locate relevant text, while Cypher can retrieve precise relationships. Combining retrieval methods helps when a question needs both narrative evidence and structured facts.

Suppose we ask which customers face disruption from a supplier delay. The agent can find the event, follow product and order dependencies, and retrieve delivery commitments. It can then check whether the collected evidence covers the question.

Neo4j’s GraphRAG Python package provides vector, hybrid, and Cypher-based retrievers. Its ToolsRetriever supports LLM-guided tool selection. A surrounding orchestration loop can add follow-up retrieval, evaluation, and stopping conditions.

Use bounded, authorised queries and retain the sources behind each claim. If evidence is missing or contradictory, retrieve again within a budget or state the uncertainty. The goal is a grounded answer, not an endlessly busy agent.