Struct Graph 

pub struct Graph { /* private fields */ }

A built proximity graph, before any encoding: over normalized vectors (angular) or the vectors as given (Euclidean). Holds a full-precision copy of the data, so it is a build-time object; the indexes keep only their encodings.

Implementations

impl Graph

pub fn build_with_training_queries( data: &[f32], dim: usize, params: &BuildParams, training_queries: Option<&[f32]>, ) -> Result<Self, WarmError>

Builds a graph with optional construction queries in the graph’s input coordinates. Queries must be independent of evaluation data. The first 32768 rows guide additional projected edges; all supplied rows are validated. None preserves the legacy build. MIPS callers must augment base rows and append zero to queries before this call.

pub fn build_owned_with_training_queries( data: Vec<f32>, dim: usize, params: &BuildParams, training_queries: Option<&[f32]>, ) -> Result<Self, WarmError>

Owned-data equivalent of Graph::build_with_training_queries.

pub fn apply_training_queries_with( &mut self, queries: &[f32], limit: usize, neighbors: usize, projected_degree: usize, ) -> Result<ProjectionStats, WarmError>

Applies the first limit construction queries, with the exact neighbours scored per query and the projected degree as arguments; each is capped by the graph size. The neighbours are found by exact scoring of every base row, so the cost grows with limit * n.

pub fn project_training_neighbors( &mut self, queries: &[f32], query_neighbor_ids: &[u32], neighbors_per_query: usize, projected_degree: usize, ) -> Result<ProjectionStats, WarmError>

Projects neighbour lists the caller found for construction queries (for example by an approximate search of an index encoded from this graph), neighbors_per_query ids per query, nearest first. queries are in graph coordinates and only validated; the lists must come from construction queries, never evaluation data.

impl Graph

Benchmark support: the graph structure (not the vectors) on disk, so that search experiments on one graph need not rebuild it.

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

Writes the header (n, dim, degree, build_ef, alpha, seed, entry, data fingerprint, quantized_search, metric), the adjacency and the upper layers; the vectors are not stored.

pub fn load( data: &[f32], dim: usize, path: &Path, params: &BuildParams, ) -> Result<Self>

Reads a graph written by Graph::save for the same data (prepared here exactly as Graph::build does) and the same params. Any mismatch in the vector count, dimension, degree, build_ef, alpha, seed, quantized search, metric or the data fingerprint is an error, so a stale cache never stands in for a build; delete the file to rebuild.

impl Graph

pub fn metric(&self) -> Metric

The metric the graph was built for.

pub fn structure_hash(&self) -> u64

Stable 64-bit hash of the structure (adjacency, upper layers, entry): equal values mean a bit-identical graph, across builds, binaries and cache round trips.

pub fn build( data: &[f32], dim: usize, params: &BuildParams, ) -> Result<Self, WarmError>

Build over data (row-major, dim floats per vector). Angular rows are normalized; zero rows stay zero. Uses the rayon pool; the result depends only on the data and params (code_bits is not used).

pub fn build_owned( data: Vec<f32>, dim: usize, params: &BuildParams, ) -> Result<Self, WarmError>

Graph::build over an owned matrix, prepared in place: the caller’s copy is the one the graph keeps, so the vectors are resident once during construction instead of twice (the build’s own copy next to the caller’s).