Vector search

On this page

Vector (ANN) index
- Declare
- Query
Removing the index

Vector (ANN) index

Store embedding vectors (list-of-number properties) and query for the k nearest neighbours. Small indexes use exact brute-force k-NN; large ones (at/above a threshold) use an approximate HNSW graph for sub-linear search, with the returned candidates re-scored exactly — returned candidates are ordered by their exact scores. The approximate candidate set may miss a true nearest neighbour; results are not guaranteed equal to a full exact search. The HNSW graph is rebuilt on reopen (only the index declaration is persisted), and is deterministic.

Declare

CREATE VECTOR INDEX embIdx FOR (n:Doc) ON (n.embedding)
OPTIONS { indexConfig: { `vector.dimensions`: 3, `vector.similarity_function`: 'cosine' } };

SHOW VECTOR INDEXES;

vector.dimensions — the expected embedding length; a wrong-length vector is silently skipped. Omit (or 0) to accept any length.
vector.similarity_function — 'cosine' (default) or 'euclidean'.
Backtick the dotted config keys (the lexer accepts backtick-quoted identifiers).

Query

Store vectors with the declared dimension before querying. An empty index returns no neighbours. For a standalone example:

CREATE (:Doc {title:'seed', embedding:[1.0,0.0,0.0]});
CREATE (:Doc {title:'near', embedding:[0.9,0.1,0.0]});

MATCH (d:Doc {title: 'seed'})
CALL db.index.vector.queryNodes('embIdx', 5, d.embedding)
YIELD node, score
RETURN node.title, score ORDER BY score DESC;

Cosine scores lie in [0, 1]; Euclidean scores lie in (0, 1] (higher = nearer). Maintained on every write and rebuilt on restore (declaration-only persistence).

Removing the index

After querying, DROP VECTOR INDEX embIdx removes the index; stored vectors remain.

EXPLAIN and PROFILE · Resident and hybrid storage · Schema and maintenance procedures