Struct LshIndex 

pub struct LshIndex { /* private fields */ }

Approximate nearest-neighbour index backed by LSH.

Insert embeddings during indexing, query during search. The returned candidate IDs should then be exact-scored with cosine similarity.

Implementations

impl LshIndex

pub fn new( dim: usize, num_hyperplanes: usize, num_tables: usize, seed: u64, ) -> Self

Create a new index.

• dim — dimensionality of your embedding vectors
• num_hyperplanes — bits per hash (12–16 is a good starting range)
• num_tables — number of independent hash tables (4–8 typical)
• seed — PRNG seed; use ic_cdk::api::time() on ICP

pub fn insert(&mut self, id: String, embedding: &[f64])

Index an embedding under id.

Call once per embedding at insert time. If a document produces multiple embeddings (title + description + keywords), insert each separately with the same id — the candidate set deduplicates by id anyway.

pub fn query(&self, embedding: &[f64]) -> Vec<String>

Return candidate IDs whose hash matches the query in at least one table.

This is the fast path. The caller is responsible for exact-scoring the candidates with cosine similarity and taking the top-N.

pub fn len(&self) -> usize

Number of distinct IDs in the index.

pub fn is_empty(&self) -> bool

pub fn clear(&mut self)

pub fn search( &self, query: &Embedding, store: &HashMap<String, Embedding>, metric: Option<DistanceMetric>, ) -> Vec<(String, f64)>

Score and rank candidates for query against store, returning (id, score) pairs.

The LSH bucket lookup narrows the field; metric exact-scores the survivors. Pass None to use the default DistanceMetric::Cosine { normalized: false }.

Results are sorted best-first:

• similarity metrics (Cosine, DotProduct) → descending
• distance metrics (Euclidean, Manhattan, Chebyshev, Angular) → ascending

IDs present in the candidate set but absent from store are silently skipped.