using-graph-databases

Use variable-length path matching like MATCH (u:User)-[:FRIEND*1..3]->(fof) with DISTINCT and LIMIT to find connections up to a bounded depth. Filter out existing friends with WHERE NOT exists((u)-[:FRIEND]-(fof)) and order by mutual friend count.

Choose Neo4j for graph-first workloads needing mature tooling and 65+ GDS algorithms like PageRank and Louvain. Choose ArangoDB when you need multi-model storage combining documents and graphs in one database with a single AQL query language.

Avoid graph databases for fixed schemas with shallow joins of two to three tables, where PostgreSQL performs better. They are also a poor fit for aggregation-heavy analytics, which suit columnar databases, or simple key-value lookups, which suit Redis or DynamoDB.

Slow traversals usually come from unbounded variable-length paths, missing indexes on filtered properties, or supernodes with thousands of relationships. Bound traversals with depth limits like *1..4, create indexes on lookup properties, and partition supernodes with intermediate time or category nodes.

Yes, hybrid search runs vector similarity in Qdrant or pgvector first, then expands results with graph relationships in Neo4j. This enriches LLM context by traversing RELATED_TO or IS_A relationships from the vector-matched concepts.

Use Cypher pattern matching to find circular money flows with MATCH path = (a:Account)-[:SENT*3..6]->(a), filtering relationships by amount thresholds. You can also detect shared devices across accounts and rapid transaction chains using timestamp duration comparisons.