nornicdb-vector-search

Vector & Full-Text Search (Cypher API)

NornicDB exposes Neo4j-compatible vector and full-text indexes plus search procedures. Most queries that work on Neo4j 5 vector indexes work unchanged on NornicDB; the extensions are documented inline.

Vector indexes

Create

DDL form (Neo4j-compatible):

CREATE VECTOR INDEX docEmbeddings IF NOT EXISTS
FOR (n:Document) ON (n.embedding)
OPTIONS { indexConfig: {
  `vector.dimensions`: 1024,
  `vector.similarity_function`: 'cosine'   -- 'cosine' | 'euclidean' | 'dot'
}}

Procedure form (handy from drivers that prefer CALL):

CALL db.index.vector.createNodeIndex(
  'docEmbeddings',     -- index name
  'Document',          -- node label
  'embedding',         -- property holding the vector
  1024,                -- dimensions
  'cosine'             -- similarity function
)

For relationship vectors:

CALL db.index.vector.createRelationshipIndex(
  'edgeEmbeddings', 'CO_ACCESSED', 'embedding', 1024, 'cosine'
)

Drop:

DROP INDEX docEmbeddings IF EXISTS
-- or
CALL db.index.vector.drop('docEmbeddings')

Similarity functions

Name	Range	Use when
cosine (default)	[-1, 1] mapped to [0, 1] for ranking	Most embeddings — the model trained with normalized cosine
dot	unbounded; valid only on normalized vectors	Speed-optimized cosine on guaranteed-unit-norm vectors
euclidean	distance, not similarity	Use when the model was trained with L2 distance

Dimensions

Every node indexed under a vector index must have the same dimension as the index's configured vector.dimensions. Mismatched dimensions are silently skipped on write, and produce 400 Bad Request from the search HTTP endpoint when queried.

Model	Dim
all-MiniLM-L6-v2	384
bge-m3, mxbai-embed-large, e5-large	1024
text-embedding-3-small	1536
text-embedding-3-large	3072

Query

The flagship procedure accepts three input shapes. Pick whichever fits the call site: