Struct SparseInvertedIndex 

pub struct SparseInvertedIndex { /* private fields */ }

An in-memory inverted index over sparse vectors (ADR-0045).

Maps each sparse dimension to the documents that have a nonzero weight there, so search accumulates a dot-product score over only the query’s nonzero dimensions. Built and maintained by quiver-embed; never persisted.

Implementations

impl SparseInvertedIndex

pub fn new() -> Self

An empty index.

pub fn len(&self) -> usize

Number of live documents.

pub fn is_empty(&self) -> bool

Whether the index holds no documents.

pub fn contains(&self, ext_id: &str) -> bool

Whether ext_id is currently indexed.

pub fn upsert(&mut self, ext_id: &str, sv: &SparseVector)

Insert or replace ext_id’s sparse vector. Re-upserting an existing id first removes its prior postings, so a dimension it no longer carries does not linger and a changed weight is not double-counted. Duplicate input dimensions are de-duplicated (last weight wins).

pub fn remove(&mut self, ext_id: &str) -> bool

Remove ext_id and free its slot. Returns whether it was present.

pub fn search(&self, query: &SparseVector) -> Vec<(String, f32)>

Score every document that shares a nonzero dimension with query by sparse dot product, and return (ext_id, score) for those with a positive score, sorted by score descending then id ascending (a deterministic, total order). The caller re-checks any payload filter on the ranked ids and truncates to its depth, so low-scored rows never load a payload. Duplicate query dimensions are de-duplicated (last weight wins).

pub fn bm25_search( &self, query_terms: &[u32], k1: f32, b: f32, ) -> Vec<(String, f32)>

Score documents against query_terms (term ids) with Okapi BM25 (ADR-0046), treating each document’s stored weights as term frequencies and using the index’s own corpus statistics — document frequency (posting-list length), document count, and average document length. k1/b are the usual BM25 parameters (BM25_K1, BM25_B). Duplicate query terms count once. Returns (ext_id, score) for documents with a positive score, sorted by score descending then id ascending. Uses the Lucene-style smoothed IDF ln(1 + (N − df + 0.5)/(df + 0.5)), which is always non-negative, so even a term in most of the corpus contributes a small positive amount (no negative scores to clamp).

Trait Implementations

impl Debug for SparseInvertedIndex

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for SparseInvertedIndex

fn default() -> SparseInvertedIndex

Returns the “default value” for a type.