Neo4j is what you reach for when the relationships are the data. If your domain looks like a whiteboard of circles and arrows, forcing it into tables is painful. Neo4j models that picture as a property graph and queries it with Cypher. This guide covers what Neo4j is used for, ACID behaviour, Neo4j vs Amazon Neptune vs TigerGraph (and peers), GraphRAG with vector indexes, local and production install paths, ports and neo4j.conf, and copy-paste Cypher and Python patterns. For broader context on data infrastructure choices, see the Data Infrastructure for AI Systems pillar. What is Neo4j used for in production graph workloads Neo4j is for connected data where you need to ask connected questions, repeatedly, under production constraints. That is the direct answer to what is Neo4j used for in most teams. Property graph data model with nodes, relationships, and properties Neo4j uses the property graph model: nodes represent entities, relationships connect nodes, and both can have properties. Labels and relationship types give structure without locking you into a brittle schema. You can start with a thin model, ship value, and evolve the graph as new questions appear. Cypher graph query language for pattern matching without join soup Cypher is declarative and built around pattern matching. You describe subgraph shapes and let the planner execute them. If SQL is about sets, Cypher is about subgraphs. That matters for multi-hop traversal, path queries, recommendations, provenance, and “who touched what via which system” questions. Is Neo4j ACID compliant and why you should care Is Neo4j ACID compliant? Yes. Creating or updating relationships touches coherent structure; the database keeps that consistent under failure and concurrency. Design graph apps around strong transactional guarantees unless you are forced otherwise. That makes debugging and reasoning about behaviour much easier than assuming vague eventual consistency. Neo4j vs Amazon Neptune vs TigerGraph: a senior engineer comparison A “Neo4j vs X” question is usually “Which ecosystem will we live in for years?” Short, opinionated view—about engineering time, not benchmark slides. Product Core model and query style Where it wins Where it bites Neo4j Property graph and Cypher Strong ergonomics for connected data, mature tooling, graph plus vector retrieval Graph modelling is a skill you must invest in Amazon Neptune Managed graph on AWS (Gremlin, openCypher, SPARQL for RDF) AWS-centric contracts and operations Query language mix can feel platform-driven TigerGraph GSQL and OpenCypher-related patterns Analytics-style workloads and compiled query approaches Different mental model; not drop-in Cypher everywhere JanusGraph Distributed graph with external storage backends Open source with pluggable backends You operate the backend stack ArangoDB Multi-model (documents, KV, graph) One database for mixed shapes Graph depth varies versus graph-first engines Memgraph Property graph, Cypher compatible Streaming and fresh-data workflows Engine behaviour differs; compatibility is not identity What to decide before you pick a graph database Pick query language and operations model first. If your team wants Cypher and a graph-first workflow, Neo4j is a strong default. If you already have Gremlin expertise, Neptune or JanusGraph can fit. If you want one multi-model store, ArangoDB can reduce moving parts. Be honest about operations. “We will run a distributed storage backend” is easy to say until you are paged about compactions or JVM pressure at 03:00. Neo4j for RAG and GraphRAG: vector search plus graph context Many RAG stacks start as vector search plus prompt. That works until you need provenance, entity resolution, multi-hop context, or disambiguation—then you risk rebuilding a knowledge graph in application code. How does GraphRAG improve retrieval augmented generation? It uses the graph to pull structured context—entities, relationships, neighbourhoods—that similarity alone often misses, which helps grounding and trustworthiness. Neo4j vector index for embedding similarity search Can Neo4j do vector search for RAG? Yes. Neo4j supports vector indexes for similarity over embeddings (commonly HNSW-style approximate nearest neighbour search). Vectors find “things that look similar”. They do not by themselves encode “how they relate” in your domain. Neo4j lets you combine similarity with traversals. Using the SEARCH subclause for vector-constrained pattern matching Neo4j’s SEARCH subclause lets you constrain a Cypher MATCH pattern using approximate nearest neighbour hits from a vector index. That is the ergonomic bridge for hybrid retrieval. Practical pattern: vector retrieval for candidates, then graph expansion for context, filters, and explanation. GraphRAG in Python with neo4j-graphrag Neo4j’s neo4j-graphrag package