Struct DenseVectorConfig 

pub struct DenseVectorConfig {
    pub dim: usize,
    pub index_type: VectorIndexType,
    pub quantization: DenseVectorQuantization,
    pub num_clusters: Option<usize>,
    pub ivf_routing: IvfRoutingMode,
    pub nprobe: usize,
    pub unit_norm: bool,
    pub soar: Option<SoarConfig>,
}

Configuration for dense vector fields using exact Flat accumulation or the single production IVF-PQ ANN format.

Indexes operate in two states:

• Flat (accumulating): Brute-force search over raw vectors before build_vector_index is called.
• Built (ANN): Fast approximate nearest neighbor search using trained structures. Centroids and codebooks are trained from index-wide data and shared by every segment; segment payloads contain only assignments and PQ codes.

Fields

dim: usize

Dimensionality of vectors

index_type: VectorIndexType

Target vector index algorithm (Flat or IVF-PQ). When in accumulating state, search uses brute-force regardless of this setting.

quantization: DenseVectorQuantization

Storage quantization for vector elements (f32, f16, uint8)

num_clusters: Option<usize>

Number of IVF leaf clusters. If omitted, a billion-scale cost model and the available training sample determine the value. If None, automatically determined based on dataset size.

ivf_routing: IvfRoutingMode

Coarse-codebook routing strategy. This setting is metric agnostic and is applied to every IVF-backed dense index.

nprobe: usize

Number of leaf clusters to probe during search (default: 64)

unit_norm: bool

Whether stored vectors are pre-normalized to unit L2 norm. When true, scoring skips per-vector norm computation (cosine = dot / ||q||), reducing compute by ~40%. Common for embedding models (e.g. OpenAI, Cohere). Default: true (most embedding models produce L2-normalized vectors).

soar: Option<SoarConfig>

SOAR spilled cluster assignments for IVF-PQ. Assigns vectors to a secondary cluster with an orthogonality-amplified residual, improving recall at the same nprobe for ~1.2-2x assignment storage. Default: None (disabled). Ignored while a field remains flat.

Implementations

impl DenseVectorConfig

pub fn new(dim: usize) -> Self

pub fn with_ivf_pq( dim: usize, num_clusters: Option<usize>, nprobe: usize, ) -> Self

Create IVF-PQ configuration.

pub fn flat(dim: usize) -> Self

Create Flat (brute-force) configuration - no ANN index

pub fn with_quantization(self, quantization: DenseVectorQuantization) -> Self

Set storage quantization

pub fn with_unit_norm(self) -> Self

Mark vectors as pre-normalized to unit L2 norm

pub fn with_num_clusters(self, num_clusters: usize) -> Self

Set number of IVF clusters

pub fn with_ivf_routing(self, routing: IvfRoutingMode) -> Self

Set flat, two-level, or HNSW IVF centroid routing explicitly.

pub fn with_soar(self, soar: SoarConfig) -> Self

Enable SOAR spilled secondary cluster assignments (IVF-based indexes only)

pub fn uses_ivf(&self) -> bool

Check if this config uses IVF

pub fn is_flat(&self) -> bool

Check if this config is flat (brute-force)

pub fn optimal_num_clusters(&self, num_vectors: usize) -> usize

Calculate optimal number of clusters for given vector count

Trait Implementations

impl Clone for DenseVectorConfig

fn clone(&self) -> DenseVectorConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for DenseVectorConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for DenseVectorConfig

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for DenseVectorConfig

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.