Struct VectorStoragePolicy 

pub struct VectorStoragePolicy {
    pub column_name: String,
    pub dim: u32,
    pub metric: VectorMetric,
    pub precision: VectorPrecision,
    pub pq: Option<PQConfig>,
    pub keep_raw_for_reranking: bool,
    pub pre_normalize: bool,
    pub hnsw_m: Option<u32>,
    pub hnsw_ef_construction: Option<u32>,
    pub ivf_residual: bool,
    pub embedding_model: Option<EmbeddingModelInfo>,
    pub modality: Option<VectorModality>,
}

Vector storage configuration applied at table creation time. Stored in Iceberg metadata.json properties.

Fields

column_name: String

dim: u32

metric: VectorMetric

precision: VectorPrecision

pq: Option<PQConfig>

keep_raw_for_reranking: bool

pre_normalize: bool

Normalize each input vector to unit L2 length before indexing. Enables the NormalizedCosine fast path in HNSW: distance = 1 - dot(a, b), no sqrt, ~2× faster distance computation. Semantics unchanged — same top-k results as Cosine. Most embedding models (OpenAI, Cohere, etc.) produce nearly-unit vectors; enabling this adds negligible write overhead.

hnsw_m: Option<u32>

HNSW M parameter — connections per node. None = default (16). Higher M → better recall, more memory, slower build. Recommended values: 8 (low-memory), 16 (default), 32 (high-recall), 64 (max).

hnsw_ef_construction: Option<u32>

HNSW ef_construction — candidate pool size during build. None = default (150). Higher ef_construction → better graph quality, slower build. Recommended values: 100 (fast), 150 (default), 200 (quality), 400 (max quality).

ivf_residual: bool

IVF-PQ residual encoding — train PQ on per-cluster residuals (vec - coarse_centroid). Same bytes/vector, ~2-4pp better recall@10. Only applies when IVF-PQ index is used.

embedding_model: Option<EmbeddingModelInfo>

Optional embedding model metadata. When set:

• Stored as ailake.embedding-model in Iceberg table properties.
• Validated on every write_batch: dim mismatch → hard error; name mismatch → warning.
• Required for migrate_embeddings to track the model transition.

modality: Option<VectorModality>

Modality tag for this vector column (text / image / audio / video). Stored as ailake.modality-<col> in Iceberg properties and Parquet KV metadata. Allows readers to select the correct HNSW by modality without reading data.

Implementations

impl VectorStoragePolicy

pub fn default_f16(column: &str, dim: u32, metric: VectorMetric) -> Self

Trait Implementations

impl Clone for VectorStoragePolicy

fn clone(&self) -> VectorStoragePolicy

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for VectorStoragePolicy

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

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for VectorStoragePolicy

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

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for VectorStoragePolicy

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

where __S: Serializer,

Serialize this value into the given Serde serializer.