Graph Databases

Graph query: run Cypher statements or Gremlin traversals with parameters from the input rows

Graph output: write input rows to a graph through a graph model

Graph vector search: find the nodes or relationships nearest to a vector through a vector index

Get graph schema: the labels, relationship types and properties of a database, one row each

Cypher output: write one or two nodes and a relationship per row with Cypher

Cypher builder: build Cypher statements from operations in a dialog

Split graph: split a graph value into its nodes and relationships

Generate graph CSVs: write the nodes and relationships of graphs to CSV files for bulk loading

Neo4j import: bulk load those CSV files with neo4j-admin import (Neo4j only)

Get graph logging info: the dates of earlier executions in the execution logging

Graph script: run a script of statements

Graph index and Graph constraint: manage indexes and constraints in the syntax of the database

Check graph database connections: check that databases are available

Set the variable HOP_GRAPH_LOGGING_CONNECTION to the name of a connection to log executions to it, see the logging schema. Query the history with the Graph logging perspective.

Store execution information in a graph database location, to view it in the execution information perspective.

In a graph model, give the property the type Vector. Graph output then writes the field as a vector.

In Graph query, pick Vector as the graph data type of a parameter to pass a vector to a statement, for example to a vector search. Give a return value the Hop type Vector to read a stored vector back into a Vector field.

A text field holding [0.1,0.2,0.3], as Embed text writes it with output format STRING, is converted too.

Generate graph CSVs writes a vector as a float[] column, with ; between the elements.

Neo4j: CALL db.index.vector.queryNodes('index', 5, $q) YIELD node, score, where the score is between 0 and 1: (1 + cosine similarity) / 2, or 1 / (1 + squared distance).

Memgraph: CALL vector_search.search('index', 5, $q) YIELD node, distance, where the distance is 1 - cosine similarity, or the squared distance: lower is nearer. Memgraph’s similarity column is the absolute value of 1 - distance, which ranks opposite vectors as near ones.

FalkorDB: CALL db.idx.vector.queryNodes('Label', 'property', 5, vecf32($q)) YIELD node, score, where the score is a distance, 1 - cosine similarity or the euclidean distance: lower is nearer.

Neo4j: CALL db.index.vector.queryRelationships('index', 5, $q) YIELD relationship, score

Memgraph: CALL vector_search.search_edges('index', 5, $q) YIELD edge, distance

FalkorDB: CALL db.idx.vector.queryRelationships('TYPE', 'property', 5, vecf32($q)) YIELD relationship, score, again with a distance.

Neo4j connections keep working. You can convert them to graph database connections, see Converting a Neo4j connection.

The transforms and actions have vendor neutral names, icons and the category Graph. Their plugin IDs didn’t change, so pipelines and workflows open as before:

HOP_GRAPH_LOGGING_CONNECTION replaces NEO4J_LOGGING_CONNECTION, which keeps working when the new variable isn’t set.

The Cypher builder writes parameters as $name. It wrote them as {name}, which Neo4j 4 and later reject, so statements without UNWIND work again.

Get graph logging info finds the previous executions of pipelines and workflows: it looked for execution types which the logging doesn’t write. The workflow logging now writes the status of a workflow, which these lookups need.

The graph index, graph constraint, graph script and check connections actions report their own success, also after a failure hop.

Graph output validates against the graph model on every database: the use of the model always, its indexes and unique constraints when the database can list them.