Struct BmpExecutor 

pub struct BmpExecutor<'a> { /* private fields */ }

Block-Max Pruning (BMP) executor for learned sparse retrieval

Processes blocks in score-descending order using a priority queue. Best for queries with many terms (20+), like SPLADE expansions. Uses document accumulators (FxHashMap) instead of per-term iterators.

Memory-efficient: Only skip entries (block metadata) are kept in memory. Actual block data is loaded on-demand via mmap range reads during execution.

Reference: Mallia et al., “Faster Learned Sparse Retrieval with Block-Max Pruning” (SIGIR 2024)

Implementations

impl<'a> BmpExecutor<'a>

pub fn new( sparse_index: &'a SparseIndex, query_terms: Vec<(u32, f32)>, k: usize, heap_factor: f32, ) -> Self

Create a new BMP executor with lazy block loading

query_terms should contain only dimensions that exist in the index. Block metadata (skip entries) is read from the sparse index directly.

pub async fn execute(self) -> Result<Vec<ScoredDoc>>

Execute BMP and return top-k results

Builds the priority queue from skip entries (already in memory), then loads blocks on-demand via mmap range reads as they are visited.