kdrant-core/dev.kdrant.model/VectorData

VectorData

@Serializable(with = VectorDataSerializer::class)

sealed interface VectorData

The vector(s) attached to a point, in any shape Qdrant accepts or returns:

Dense — one anonymous dense vector: [0.1, 0.2, ...].
Sparse — a sparse vector (always used under a name): {"indices":[...],"values":[...]}.
MultiDense — a multi-vector / late-interaction (ColBERT) vector: [[...],[...]].
Named — a map of vector-name to any of the above; dense, sparse and multi-vectors may be mixed.
Inference — what the server should embed, rather than a vector: see InferenceInput.
Raw — a shape this version does not model, kept verbatim so decoding a response never fails.

Inheritors

Types

Dense

data class Dense(val values: List<Float>) : VectorData

One anonymous dense vector.

DenseArray

class DenseArray(val values: FloatArray) : VectorData

One anonymous dense vector backed by a FloatArray — a zero-boxing fast path for the hot path (values are never wrapped in Float objects, and serialization writes the array directly). Prefer this over Dense for large vectors. Equality is by content.

Inference

data class Inference(val input: InferenceInput) : VectorData

Text, an image or a custom object for the server to embed, instead of a vector the caller computed. Needs a Qdrant with an inference provider configured. See InferenceInput.

MultiDense

data class MultiDense(val vectors: List<List<Float>>) : VectorData

A multi-vector / late-interaction (ColBERT) vector: several equal-length dense vectors.

Named

data class Named(val vectors: Map<String, VectorData>) : VectorData

Named vectors; each entry may be Dense, Sparse or MultiDense.

Raw

data class Raw(val json: JsonElement) : VectorData

A vector shape this version does not model, kept verbatim so response decoding degrades gracefully.

Sparse

data class Sparse(val indices: List<Int>, val values: List<Float>) : VectorData

A sparse vector; indices must be unique and the same length as values. Used under a name.