Struct VectorDb 

pub struct VectorDb { /* private fields */ }

Vector database for storing and querying embedding vectors.

VectorDb provides a high-level API for ingesting vectors with metadata. It handles internal details like ID allocation, centroid assignment, and metadata index maintenance automatically.

Implementations

impl VectorDb

pub async fn open(config: Config) -> Result<Self>

Open or create a vector database with the given configuration and centroids.

If the database already exists (centroids are already stored), the provided centroids are ignored and the stored centroids are used instead.

If the database is new, the provided centroids are written to storage and used to build the HNSW index.

Arguments

• config - Database configuration
• centroids - Initial centroids to use if database is new

Configuration Compatibility

If the database already exists, the configuration must be compatible:

• dimensions must match exactly
• distance_metric must match exactly

Other configuration options (like flush_interval) can be changed on subsequent opens.

pub async fn open_with_storage( config: Config, builder: StorageBuilder, ) -> Result<Self>

pub async fn open_with_centroids( config: Config, centroids: Vec<Vec<f32>>, builder: StorageBuilder, ) -> Result<Self>

pub async fn write(&self, vectors: Vec<Vector>) -> Result<()>

Write vectors to the database.

This is the primary write method. It accepts a batch of vectors and returns when the data has been accepted for ingestion (but not necessarily flushed to durable storage).

Atomicity

This operation is atomic: either all vectors in the batch are accepted, or none are. This matches the behavior of TimeSeriesDb::write().

Upsert Semantics

Writing a vector with an ID that already exists performs an upsert: the old vector is deleted and replaced with the new one. The system allocates a new internal ID for the updated vector and marks the old internal ID as deleted. This ensures index structures are updated correctly without expensive read-modify-write cycles.

Validation

The following validations are performed:

• Vector dimensions must match Config::dimensions
• Attribute names must be defined in Config::metadata_fields (if specified)
• Attribute types must match the schema

pub async fn write_timeout( &self, vectors: Vec<Vector>, timeout: Duration, ) -> Result<()>

Write vectors to the database with a timeout.

This is the primary write method. It accepts a batch of vectors and returns when the data has been accepted for ingestion (but not necessarily flushed to durable storage).

The write may time out if the db is busy compacting or indexing.

Atomicity

This operation is atomic: either all vectors in the batch are accepted, or none are. This matches the behavior of TimeSeriesDb::write().

Upsert Semantics

Writing a vector with an ID that already exists performs an upsert: the old vector is deleted and replaced with the new one. The system allocates a new internal ID for the updated vector and marks the old internal ID as deleted. This ensures index structures are updated correctly without expensive read-modify-write cycles.

Validation

The following validations are performed:

• Vector dimensions must match Config::dimensions
• Attribute names must be defined in Config::metadata_fields (if specified)
• Attribute types must match the schema

pub async fn flush(&self) -> Result<()>

Force flush all pending data to durable storage.

Flushes the in-memory delta to the storage memtable, then persists to durable storage. After this returns, data is both readable and durable.

Atomic Flush

The flush operation is atomic:

1. All pending writes are frozen into an immutable delta
2. RecordOps are applied in one batch via storage.apply()
3. The snapshot is updated for queries
4. Data is flushed to durable storage

This ensures ID dictionary updates, deletes, and new records are all applied together, maintaining consistency.

pub async fn close(self) -> Result<()>

Closes the vector database, flushing any pending data and releasing resources.

All written data is flushed to durable storage before the database is closed. For SlateDB-backed storage, this also releases the database fence.

pub fn num_centroids(&self) -> usize

pub async fn search_exact_nprobe( &self, query: &Query, nprobe: usize, ) -> Result<Vec<SearchResult>>

Search using brute-force centroid lookup (for diagnostics).

pub async fn snapshot(&self) -> Box<dyn VectorDbRead>

Trait Implementations

impl VectorDbRead for VectorDb

fn search<'life0, 'life1, 'async_trait>( &'life0 self, query: &'life1 Query, ) -> Pin<Box<dyn Future<Output = Result<Vec<SearchResult>>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait,

Search for k-nearest neighbors to a query vector.

fn search_with_nprobe<'life0, 'life1, 'async_trait>( &'life0 self, query: &'life1 Query, nprobe: usize, ) -> Pin<Box<dyn Future<Output = Result<Vec<SearchResult>>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait,

fn get<'life0, 'life1, 'async_trait>( &'life0 self, id: &'life1 str, ) -> Pin<Box<dyn Future<Output = Result<Option<Vector>>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait,

Retrieve a vector record by its external ID.