Struct IvfFlatIndex 

pub struct IvfFlatIndex { /* private fields */ }

An IVF-Flat index. All methods take &self; internal mutable state is guarded by a single RwLock, so the index is Send + Sync and safe to share behind an Arc like the usearch backend.

Implementations

impl IvfFlatIndex

pub fn new(dims: usize, metric: Metric, nlist: usize, nprobe: usize) -> Self

pub fn len(&self) -> usize

pub fn is_empty(&self) -> bool

pub fn is_trained(&self) -> bool

pub const fn metric(&self) -> Metric

pub const fn nlist(&self) -> usize

pub const fn nprobe(&self) -> usize

pub fn centroid_count(&self) -> usize

The number of trained centroids (0 until IvfFlatIndex::train runs).

pub fn memory_bytes(&self) -> usize

Approximate resident bytes: stored vectors + norms + centroids + bookkeeping.

pub fn upsert(&self, id: u64, v: &[f32]) -> Result<bool, String>

Insert or replace the vector for id. Returns true if it replaced an existing entry.

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

Remove the vector for id. Returns true if it existed.

pub fn search( &self, query: &[f32], top_k: usize, nprobe_override: Option<usize>, ) -> Result<Vec<(u64, f32)>, String>

Return the top_k nearest ids with their distances (ascending). Uses nprobe_override cells when given, else the configured default; falls back to an exact scan when untrained.

pub fn train(&self) -> Result<(), String>

(Re)train centroids over the currently stored vectors via k-means, then rebuild the inverted lists. Cheap no-op when empty. The centroid count is min(nlist, n).

pub fn bulk_load( &self, items: impl IntoIterator<Item = (u64, Vec<f32>)>, ) -> Result<(), String>

Replace the entire contents in one shot (used for the initial bulk load). Does not train; call IvfFlatIndex::train afterwards.