Trait VectorBackend

Source

pub trait VectorBackend: Send + Sync {
Show 15 methods    // Required methods
    fn insert(
        &self,
        id: u64,
        key: &str,
        vector: &[f32],
        metadata: Value,
    ) -> VectorResult<()>;
    fn search(&self, query: &[f32], k: usize) -> Vec<SearchResult>;
    fn len(&self) -> usize;
    fn contains(&self, id: u64) -> bool;
    fn remove(&self, id: u64) -> bool;
    fn flush(&self) -> VectorResult<()>;
    fn backend_name(&self) -> &str;

    // Provided methods
    fn is_empty(&self) -> bool { ... }
    fn current_epoch(&self) -> u64 { ... }
    fn insert_with_epoch(
        &self,
        id: u64,
        key: &str,
        vector: &[f32],
        metadata: Value,
        parent_epoch: u64,
    ) -> VectorResult<()> { ... }
    fn soft_delete(&self, id: u64) -> bool { ... }
    fn compact(&self, _older_than_epoch: u64) -> usize { ... }
    fn tombstone_count(&self) -> usize { ... }
    fn max_vectors(&self) -> Option<usize> { ... }
    fn set_max_vectors(&self, _limit: Option<usize>) { ... }
}

Expand description

Unified vector search backend interface.

Implementations must be Send + Sync so they can be shared across async tasks via Arc.

Required Methods§

Source

fn insert( &self, id: u64, key: &str, vector: &[f32], metadata: Value, ) -> VectorResult<()>

Insert a vector with the given numeric ID, string key, and metadata.

If a vector with the same id already exists, it is replaced (upsert semantics).

Source

fn search(&self, query: &[f32], k: usize) -> Vec<SearchResult>

Search for the k nearest vectors to query.

Results are sorted by ascending distance (closest first).

Source

fn len(&self) -> usize

Return the number of vectors currently stored.

Source

fn contains(&self, id: u64) -> bool

Check whether a vector with the given ID exists.

Source

fn remove(&self, id: u64) -> bool

Remove a vector by ID. Returns true if it was present.

Source

fn flush(&self) -> VectorResult<()>

Persist any in-memory state to durable storage.

For purely in-memory backends (HNSW), this is a no-op.

Source

fn backend_name(&self) -> &str

Return a human-readable name for this backend (e.g. “hnsw”, “diskann”, “hybrid”).

Provided Methods§

Source

fn is_empty(&self) -> bool

Return true if the store contains no vectors.

Source

fn current_epoch(&self) -> u64

Return the current monotonic epoch.

The epoch is bumped on every mutating operation (insert, delete, update, soft-delete, compact). Implementations should persist the epoch across restarts via [flush] / load.

Source

fn insert_with_epoch( &self, id: u64, key: &str, vector: &[f32], metadata: Value, parent_epoch: u64, ) -> VectorResult<()>

Insert a vector only if the store’s current epoch matches parent_epoch. If the epoch has advanced past parent_epoch, the call is rejected with VectorError::EpochConflict.

This enables optimistic concurrency control for multi-agent writes: each agent reads the epoch, performs its work, then submits with the epoch it observed.

Source

fn soft_delete(&self, id: u64) -> bool

Mark a vector as tombstoned instead of physically removing it.

Tombstoned vectors are excluded from search results but remain in storage until [compact] is called. Returns true if the vector existed and was not already tombstoned.

Source