Struct PgVectorIndex 

pub struct PgVectorIndex { /* private fields */ }

A pgvector-backed VectorIndex for the PostgreSQL backend.

PostgreSQL stores each memory’s embedding in a pgvector vector column (written by crate::PgStorage) and an HNSW index over it (idx_memories_embedding_hnsw, built with vector_cosine_ops). When constructed with a pool via PgVectorIndex::with_pool, search / filtered_search run a real cosine-distance ANN query (<=>) against that index and return the top-k memory ids + distances — the same (id, distance) shape the USearch backend returns, so recall’s score = 1.0 - distance conversion is identical across backends.

Constructed with PgVectorIndex::new (no pool) the ANN paths return the typed Error::BackendUnsupported rather than a silent empty set — silent-empty would make semantic / auto (hybrid) / graph / domain_scoped recall look like it legitimately “found nothing”, the most dangerous failure mode for a memory database.

Runtime (v0.5.18)

VectorIndex::search / filtered_search are async, so this backend .awaits its sqlx query directly on the caller’s ambient Tokio runtime. There is no block_on bridge — semantic recall works from inside the server/CLI #[tokio::main] runtime (any flavor, single- or multi-threaded) without the “Cannot start a runtime from within a runtime” panic or deadlock the old synchronous bridge risked. Integration tests run under #[tokio::test] / #[tokio::test(flavor = "multi_thread")] alike.

add / remove are intentional no-ops: the embedding is maintained by PostgreSQL on the vector column (via PgStorage::insert_memory), not by an in-process index. len() tracks an approximate element count for is_empty() callers.

Implementations

impl PgVectorIndex

pub fn new() -> Self

Create a pgvector index wrapper without a pool. The ANN search paths return Error::BackendUnsupported (fail loud, never silent-empty). Prefer PgVectorIndex::with_pool for a wired backend.

pub fn with_pool(pool: PgPool, dimensions: usize) -> Self

Create a pgvector index that runs real ANN search against pool.

dimensions must match the vector(dim) column width the schema was migrated with (the same value passed to PgStorage::connect).

Trait Implementations

impl Default for PgVectorIndex

fn default() -> Self

Returns the “default value” for a type.

impl VectorIndex for PgVectorIndex

fn add(&self, _id: Uuid, _vector: &[f32]) -> Result<()>

fn remove(&self, _id: Uuid) -> Result<()>

fn search<'life0, 'life1, 'async_trait>( &'life0 self, query: &'life1 [f32], limit: usize, ) -> Pin<Box<dyn Future<Output = Result<Vec<(Uuid, f32)>>> + Send + 'async_trait>>

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

fn filtered_search<'life0, 'life1, 'life2, 'async_trait>( &'life0 self, query: &'life1 [f32], limit: usize, filter: &'life2 (dyn Fn(Uuid) -> bool + Send + Sync), ) -> Pin<Box<dyn Future<Output = Result<Vec<(Uuid, f32)>>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait,

fn save(&self, _path: &Path) -> Result<()>

fn load(&self, _path: &Path) -> Result<()>

fn len(&self) -> usize

fn is_empty(&self) -> bool