Neo4j의 전문 검색과 벡터 검색

Plain property lookups in Neo4j are great for exact matches, but real applications need more. Users misspell words, search across multiple fields, and increasingly expect semantic results.

Neo4j extends its capabilities with two specialized index types:

• Full-text indexes for fuzzy, multi-field text search
• Vector indexes for similarity search over embeddings

Both are first-class features you can manage with Cypher.

A full-text index is built over one or more node labels and properties. Once created, it powers tokenized, case-insensitive search.

The example creates an index named movieSearch over the title and plot properties of Movie nodes.

You query full-text indexes with the db.index.fulltext.queryNodes procedure. It returns matching nodes plus a relevance score.

This Lucene-style syntax supports wildcards, fuzzy matching with ~, and boolean operators.

Full-text search shines with imperfect input. A few common operators:

• star~ — fuzzy match, tolerates typos
• star* — prefix wildcard
• title:matrix — restrict to one field
• matrix AND reloaded — boolean combination

These let one query handle the messy real-world queries users actually type.

A vector embedding is a list of numbers that captures the meaning of text, an image, or other data. Items with similar meaning have vectors that point in similar directions.

By storing an embedding as a property on a node, Neo4j can answer questions like find the documents most semantically similar to this one — not just keyword matches.

Vector indexes require you to declare the dimension (length of the embedding) and the similarity function (cosine or euclidean).

The db.index.vector.createNodeIndex procedure creates one over a label and property. Here we index a 1536-dimension embedding stored on Document nodes.

Embeddings are usually produced by an external model and written back to Neo4j. The db.create.setNodeVectorProperty procedure stores the float array efficiently.

In practice the array has hundreds or thousands of values; it is shortened here for readability.

To find the nearest neighbors, call db.index.vector.queryNodes with the index name, the number of results, and a query vector.

It returns nodes ordered by similarity along with a score between 0 and 1.

The real power of Neo4j is mixing search with traversal. You can find semantically similar documents, then follow relationships to enrich the results.

This query finds similar documents and returns their authors — something a pure vector database cannot do in one step.

Like any index, full-text and vector indexes can be listed and dropped. Use SHOW INDEXES to inspect them and DROP INDEX to remove one.

Always check that an index is ONLINE before relying on it in production queries.

To get the most from search indexes:

• Keep embedding dimensions consistent with your model output
• Choose cosine similarity for most text embeddings
• Re-embed and update vectors when source data changes
• Limit result counts and post-filter with Cypher for relevance

These habits keep searches fast and accurate as data grows.

You extended Neo4j with two powerful search capabilities:

• Full-text indexes — tokenized, fuzzy, multi-field keyword search via db.index.fulltext.queryNodes
• Vector indexes — semantic similarity over embeddings via db.index.vector.queryNodes

Best of all, both integrate with graph traversals, letting you blend search relevance with relationship context in a single Cypher query.