Builder API

Data Writes

Inserts use Insert; updates use native Partial Update rather than reading and replacing whole entities. Multi-page operations do not commit or roll back as one unit.

Deleting known primary keys may count keys that do not exist. Do not use that count alone to establish existence. See Data Writes for deletion after scalar or vector selection.

For chunked deletion, repeatedly fetch the first page of remaining rows and delete those keys. Scalar ordering is unnecessary; do not increment the page number while deleting.

Write Conflicts

Strategy	Milvus behavior
Default / Into	Ordinary Insert; duplicate primary keys do not reliably cause an error
Ignore	Unsupported: there is no atomic insert-if-absent operation
Update	Partial Upsert: updates specified fields of existing entities or inserts new entities

See Data Writes. Do not use an ordinary Insert exception as a duplicate-key check.

Query

Field selection, entity and Map results, scalar results, and COUNT(*) are supported. Computed projections, DISTINCT, and general SQL aggregates are not supported.

queryForObject() returns the first matching row. Without scalar ordering, it does not mean “the row with the smallest primary key.” Use a primary-key condition to identify a specific row. See Query Operations.

Map Query Mode

Mapped Map mode uses Java property names; free Map mode uses collection field names. For a property wordCount mapped to word_count, use the respective name in each mode.

See Query Operations for entity and field mapping. Map mode does not add scalar ordering, computed columns, or grouping.

Pagination

Ordinary queries use LIMIT / OFFSET and a separate count query. The driver reads data on demand. Pagination iteration uses page queries; fetchSize controls driver fetch batches, not page numbers or total rows.

Scalar fields cannot be used for ordering. Vector queries order by distance; see Pagination.

Where Builder

Comparisons, ranges, IN, NULL, and AND/OR conditions are available. Remove nulls from collections and use isNull or isNotNull to express null predicates explicitly.

Milvus 2.6.2 does not currently support these parameterized predicates:

• LIKE patterns, including like, likeLeft, likeRight, and their negated forms.
• NOT around comparisons with integer parameters, repeated NOT, and some negated half-open ranges. Use ne to exclude one value, or combine the two outside comparisons with OR for a range.

The driver does not bypass parameter binding by concatenating strings.

An ordinary predicate example:

lambda.query(BookVector.class)
        .ge(BookVector::getWordCount, 1000)
        .isNotNull(BookVector::getTitle)
        .queryForList();

Predicate Values

Comparisons, collections, ranges, and apply pass values through SDK parameter binding. Do not concatenate or escape quotes in string values yourself. LIKE pattern parameters are currently unsupported; see Where Builder.

Do not paste SQL fragments from other databases into apply, such as 1=1 for dynamic conditions. Use the API's boolean flag for optional conditions:

lambda.query(BookVector.class)
        .eq(title != null, BookVector::getTitle, title)
        .queryForList();

Expressions passed to apply must follow Milvus expression syntax.

Group By

The current dialect does not support builder groupBy or grouped aggregation. Use queryForCount() for ordinary counts. Grouping search in vector retrieval is a separate feature, not SQL grouped aggregation.

Order By

The current dialect does not support scalar asc, desc, or orderBy, including their null-placement strategies. Use orderByL2, orderByCosine, or orderByIP for vector-distance ordering. See Vector Operations.