On this page

Query language

An openCypher subset that reads and writes — and a precise account of what is not in it.

The query language lives in its own module, which depends on the storage engine’s public API and nothing inside it. That is deliberate: the engine’s specification says a query language is not a goal of the storage layer, and the module boundary is what keeps that literally true.

What is supported

MATCH (p:Person)-[:KNOWS*1..2]->(friend)
WHERE p.city = $city AND friend.age IS NOT NULL
RETURN friend.name AS name, count(*) AS routes
ORDER BY routes DESC LIMIT 10

  • MATCH with variable-length paths, shortestPath and allShortestPaths

  • WHERE, RETURN, WITH, UNWIND, CASE, UNION [ALL]

  • CREATE, MERGE with ON CREATE / ON MATCH, SET, REMOVE, DETACH DELETE

  • Aggregation: count (including count(*)), sum, avg, min, max, collect, each with DISTINCT

  • Parameters as $name, comments as // and /* */, backticks around a name with spaces

  • One procedure: CALL vector.query(label, key, vector, k) YIELD node, score

What is not, and why

Each of these is refused at parse time, by name, with a caret — so "this build cannot do that" and "you typed it wrong" are different messages.

Construct Reason

OPTIONAL MATCH

Needs null-padding semantics through the rest of the pipeline

FOREACH

UNWIND covers the common case

List comprehensions, map projections, pattern comprehensions

Not built

ALL / ANY / NONE / SINGLE, EXISTS subqueries

Not built

User-defined procedures, LOAD CSV

There is no procedure registry; CALL itself exists with exactly one procedure

Temporal and spatial types

The store has Instant; Cypher’s temporal algebra is not implemented

 
MATCH (n) FOREACH (x IN [1] | SET n.x = x)
          ^
FOREACH is not supported: use UNWIND, which this build does implement (1:11)

CREATE additionally refuses three patterns that MATCH accepts, each with its own message, because none of them describes a graph that could be built: a variable-length range does not say how many relationships to make, an undirected arrow does not say which way one points, and a missing or alternated type does not say what to create.

Writes are off by default

A write query is refused with 403 on the HTTP port and READ_ONLY on the binary port unless the server was started with --allow-write. No port has authentication, and the difference between handing out a copy of the data and handing out the ability to delete it is worth an explicit act.

A write query also does not stream: its rows are produced, its transaction commits, and only then does the first byte go out. A client that had already received rows when the commit failed would be holding a result from a transaction that no longer exists.

Null semantics

Comparison with null yields null, not false, and WHERE keeps a row only when the predicate is true. Use IS NULL and IS NOT NULL rather than = null.

Access paths

A WHERE on an indexed (label, key) pair is a lookup; on an unindexed one it is a full scan. The browser and the MCP server both expose the declared indexes so you can tell which you are about to get.

Looking for something else?

Graph overview All documentation