Skip to main content

Vector Operations

Vector Type Mapping​

The JDBC driver converts vector parameters using the collection schema. FLOAT_VECTOR(n) accepts numeric lists or one-dimensional primitive numeric arrays and returns List<Float>. Values must have n dimensions. For binary and half-precision encodings, see Vector Types.

Vector fields cannot be NULL in Milvus 2.6.2.

KNN Ordering​

Create an index and LOAD TABLE before searching. The query metric must match the index metric_type; an L2 index cannot serve a COSINE request.

LambdaTemplate lambda = new LambdaTemplate(dataSource);
List<Float> target = List.of(0.1f, 0.2f);
List<BookVector> rows = lambda.query(BookVector.class)
        .gt(BookVector::getWordCount, 500)
        .orderByL2(BookVector::getBookIntro, target)
        .initPage(5, 0)
        .queryForList();

This sends a Search request with a scalar filter and takes five nearest candidates. Each ORDER BY accepts one query vector, not a list of multiple query vectors.

For BM25, configure a BM25 function and index, then pass search text with orderByBM25("sparse", "hybrid search"). Results are ordered from highest score to lowest.

When writing a parameterized BM25 command for JdbcTemplate, escape the question mark inside <?>:

List<Map<String, Object>> rows = jdbc.queryForList(
        "SELECT id FROM docs ORDER BY sparse <\\?> ? LIMIT 10",
        new Object[] { "hybrid search" });

The driver still receives <?>; only the following ? binds the text. The builder API handles this escaping automatically. Direct JDBC calls do not need the backslash either.

Distance Range Filtering​

Milvus COSINE and IP return similarity scores: larger is better. vectorByCosine and vectorByIP take a lower score bound. For example, to require similarity above 0.8:

List<BookVector> rows = lambda.query(BookVector.class)
        .vectorByCosine(BookVector::getBookIntro, target, 0.8)
        .initPage(10, 0)
        .queryForList();

This requires a COSINE index and generates book_intro <=> ? > ?, with 0.8 passed directly as the threshold. vectorByL2 takes an upper distance bound and generates a less-than comparison.

vectorByBM25 also takes a lower score bound. Hamming and Jaccard use upper distance bounds for binary vectors.

Combined Queries​

Only one vector range may be combined with scalar filters through AND. Do not combine it with OR, NOT, another vector range, or vector ORDER BY.

Multiple Search Paths​

A scalar filter plus one vector search is not hybrid search. To merge multiple candidate lists, use HYBRID and choose the reranker in the command:

Common Builder API usage is described in Vector Query.