Similarity & Search

Overview

Compare vectors and perform semantic search to find related content.

Similarity Functions

Compare vectors using different similarity measures:

Cosine Similarity is most common for text because it measures angle (direction), not magnitude.

Semantic search using cosine similarity:

GQL
LET query = AI.embed('graph database tutorial')
MATCH (d:Document)
RETURN d.title, AI.COSINE(d.embedding, query) AS similarity
ORDER BY similarity DESC
LIMIT 5

Find nearest products by Euclidean distance:

GQL
MATCH (p1:Product {id: 'A'}), (p2:Product)
WHERE p1 <> p2
RETURN p2.name, AI.EUCLIDEAN(p1.embedding, p2.embedding) AS distance
ORDER BY distance ASC
LIMIT 5

Find related article pairs:

GQL
MATCH (a:Article), (b:Article)
WHERE a.id < b.id  // Avoid duplicates
RETURN a.title, b.title, AI.DOT(a.embedding, b.embedding) AS similarity
ORDER BY similarity DESC
LIMIT 10

Vector Search

For large datasets, use vector indexes for efficient approximate nearest neighbor (ANN) search.

Create vector index for fast search:

GQL
CREATE VECTOR INDEX doc_search
FOR (d:Document)
ON (d.embedding)
OPTIONS {
  dimensions: 1536,
  similarity: 'cosine'
}

Efficient ANN search using index:

GQL
LET query = AI.embed('how to model relationships in graphs')
CALL db.index.vector.search('doc_search', query, 10)
YIELD node, score
RETURN node.title, score

Vector search combined with graph traversal:

GQL
LET query = AI.embed('machine learning applications')
CALL db.index.vector.search('doc_search', query, 5)
YIELD node AS doc, score
MATCH (doc)-[:AUTHORED_BY]->(author:Person)
MATCH (author)-[:WORKS_AT]->(company:Company)
RETURN doc.title, score, author.name, company.name

Hybrid Search

Combine vector similarity with traditional graph queries for best results.

Vector search with property filters:

GQL
LET query = AI.embed('cloud computing tutorial')
MATCH (d:Document)
WHERE d.category = 'Technology'
  AND d.published > date('2023-01-01')
  AND AI.COSINE(d.embedding, query) > 0.7
RETURN d.title, d.published, AI.COSINE(d.embedding, query) AS relevance
ORDER BY relevance DESC
LIMIT 10

Combine graph relationships with similarity:

GQL
MATCH (doc:Document {id: 'DOC-123'})-[:AUTHORED_BY]->(author)
MATCH (author)-[:AUTHORED_BY]-(other:Document)
WHERE other <> doc
RETURN other.title, AI.COSINE(doc.embedding, other.embedding) AS similarity
ORDER BY similarity DESC
LIMIT 5

Retrieve context for RAG applications:

GQL
LET question = 'How do I create a graph index?'
LET questionVector = AI.embed(question)
CALL db.index.vector.search('doc_search', questionVector, 3)
YIELD node, score
RETURN COLLECT_LIST(node.content) AS context