Struct IVFIndex 

pub struct IVFIndex {
    pub centroids: Vec<f32>,
    pub num_clusters: usize,
    pub dim: usize,
    pub inverted_lists: Vec<Vec<usize>>,
}

An inverted file index that partitions vectors into clusters.

Fields

centroids: Vec<f32>

Cluster centroids: [num_clusters][dim] stored flat.

num_clusters: usize

Number of clusters.

dim: usize

Vector dimension.

inverted_lists: Vec<Vec<usize>>

Inverted lists: for each cluster, the indices of vectors assigned to it.

Implementations

impl IVFIndex

pub fn train(vectors: &[Vec<f32>], dim: usize, num_clusters: usize) -> Self

Train an IVF index using k-means clustering.

pub fn assign(&self, vector: &[f32]) -> usize

Assign a vector to its nearest cluster.

pub fn search( &self, query: &[f32], vectors: &[Vec<f32>], norms: &[f32], tombstones: &[u8], nprobe: usize, k: usize, ) -> Vec<(usize, f32)>

Search using IVF: probe the nprobe nearest clusters and return top-k results by cosine similarity.

pub fn is_balanced(&self) -> bool

Check if clusters are reasonably balanced (no cluster has more than 3x the average size).

pub fn to_hdf5_data(&self) -> (&[f32], Vec<i64>, Vec<i64>, [i64; 2])

Serialize for HDF5 storage. Returns (centroids, inverted_list_offsets, inverted_list_data, metadata).

pub fn from_hdf5_data( centroids: Vec<f32>, offsets: &[i64], data: &[i64], metadata: [i64; 2], ) -> Self

Reconstruct from HDF5 data.