Struct WandExecutor 

pub struct WandExecutor<S: ScoringIterator> { /* private fields */ }

Generic MaxScore WAND executor for top-k retrieval

Works with any type implementing ScoringIterator. Implements:

• WAND pivot-based pruning: skip documents that can’t beat threshold
• Block-max WAND: skip blocks that can’t beat threshold
• Efficient top-k collection

Implementations

impl<S: ScoringIterator> WandExecutor<S>

pub fn new(scorers: Vec<S>, k: usize) -> Self

Create a new WAND executor with exact search (heap_factor = 1.0)

pub fn with_heap_factor(scorers: Vec<S>, k: usize, heap_factor: f32) -> Self

Create a new WAND executor with approximate search

heap_factor controls the trade-off between speed and recall:

• 1.0 = exact search
• 0.8 = ~20% faster, minor recall loss
• 0.5 = much faster, noticeable recall loss

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

Execute WAND and return top-k results

Implements the WAND (Weak AND) algorithm with pivot-based pruning:

1. Maintain iterators sorted by current docID (using sorted vector)
2. Find pivot: first term where cumulative upper bounds > threshold
3. If all iterators at pivot docID, fully score; otherwise skip to pivot
4. Insert into collector and advance

Reference: Broder et al., “Efficient Query Evaluation using a Two-Level Retrieval Process” (CIKM 2003)

Note: For small number of terms (typical queries), a sorted vector with insertion sort is faster than a heap due to better cache locality. The vector stays mostly sorted, so insertion sort is ~O(n) amortized.