hermes_core/query/

scoring.rs

//! Shared scoring abstractions for text and sparse vector search
//!
//! Provides common traits and utilities for efficient top-k retrieval:
//! - `ScoringIterator`: Common interface for posting list iteration with scoring
//! - `TopKCollector`: Efficient min-heap for maintaining top-k results
//! - `BlockMaxScoreExecutor`: Unified Block-Max MaxScore with conjunction optimization
//! - `BmpExecutor`: Block-at-a-time executor for learned sparse retrieval (12+ terms)
//! - `SparseTermScorer`: ScoringIterator implementation for sparse vectors

use std::cmp::Ordering;
use std::collections::BinaryHeap;
use std::sync::Arc;

use log::{debug, trace};

use crate::DocId;
use crate::structures::BlockSparsePostingList;

/// Common interface for scoring iterators (text terms or sparse dimensions)
///
/// Abstracts the common operations needed for WAND-style top-k retrieval.
pub trait ScoringIterator {
    /// Current document ID (u32::MAX if exhausted)
    fn doc(&self) -> DocId;

    /// Current ordinal for multi-valued fields (default 0)
    fn ordinal(&self) -> u16 {
        0
    }

    /// Advance to next document, returns new doc ID
    fn advance(&mut self) -> DocId;

    /// Seek to first document >= target, returns new doc ID
    fn seek(&mut self, target: DocId) -> DocId;

    /// Check if iterator is exhausted
    fn is_exhausted(&self) -> bool {
        self.doc() == u32::MAX
    }

    /// Score contribution for current document
    fn score(&self) -> f32;

    /// Maximum possible score for this term/dimension (global upper bound)
    fn max_score(&self) -> f32;

    /// Current block's maximum score upper bound (for block-level pruning)
    fn current_block_max_score(&self) -> f32;

    /// Skip to the next block, returning the first doc_id in the new block.
    /// Used for block-max WAND optimization when current block can't beat threshold.
    /// Default implementation just advances (no block-level skipping).
    fn skip_to_next_block(&mut self) -> DocId {
        self.advance()
    }
}

/// Entry for top-k min-heap
#[derive(Clone, Copy)]
pub struct HeapEntry {
    pub doc_id: DocId,
    pub score: f32,
    pub ordinal: u16,
}

impl PartialEq for HeapEntry {
    fn eq(&self, other: &Self) -> bool {
        self.score == other.score && self.doc_id == other.doc_id
    }
}

impl Eq for HeapEntry {}

impl Ord for HeapEntry {
    fn cmp(&self, other: &Self) -> Ordering {
        // Min-heap: lower scores come first (to be evicted)
        other
            .score
            .partial_cmp(&self.score)
            .unwrap_or(Ordering::Equal)
            .then_with(|| self.doc_id.cmp(&other.doc_id))
    }
}

impl PartialOrd for HeapEntry {
    fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
        Some(self.cmp(other))
    }
}

/// Efficient top-k collector using min-heap
///
/// Maintains the k highest-scoring documents using a min-heap where the
/// lowest score is at the top for O(1) threshold lookup and O(log k) eviction.
/// No deduplication - caller must ensure each doc_id is inserted only once.
pub struct ScoreCollector {
    /// Min-heap of top-k entries (lowest score at top for eviction)
    heap: BinaryHeap<HeapEntry>,
    pub k: usize,
}

impl ScoreCollector {
    /// Create a new collector for top-k results
    pub fn new(k: usize) -> Self {
        // Cap capacity to avoid allocation overflow for very large k
        let capacity = k.saturating_add(1).min(1_000_000);
        Self {
            heap: BinaryHeap::with_capacity(capacity),
            k,
        }
    }

    /// Current score threshold (minimum score to enter top-k)
    #[inline]
    pub fn threshold(&self) -> f32 {
        if self.heap.len() >= self.k {
            self.heap.peek().map(|e| e.score).unwrap_or(0.0)
        } else {
            0.0
        }
    }

    /// Insert a document score. Returns true if inserted in top-k.
    /// Caller must ensure each doc_id is inserted only once.
    #[inline]
    pub fn insert(&mut self, doc_id: DocId, score: f32) -> bool {
        self.insert_with_ordinal(doc_id, score, 0)
    }

    /// Insert a document score with ordinal. Returns true if inserted in top-k.
    /// Caller must ensure each doc_id is inserted only once.
    #[inline]
    pub fn insert_with_ordinal(&mut self, doc_id: DocId, score: f32, ordinal: u16) -> bool {
        if self.heap.len() < self.k {
            self.heap.push(HeapEntry {
                doc_id,
                score,
                ordinal,
            });
            true
        } else if score > self.threshold() {
            self.heap.push(HeapEntry {
                doc_id,
                score,
                ordinal,
            });
            self.heap.pop(); // Remove lowest
            true
        } else {
            false
        }
    }

    /// Check if a score could potentially enter top-k
    #[inline]
    pub fn would_enter(&self, score: f32) -> bool {
        self.heap.len() < self.k || score > self.threshold()
    }

    /// Get number of documents collected so far
    #[inline]
    pub fn len(&self) -> usize {
        self.heap.len()
    }

    /// Check if collector is empty
    #[inline]
    pub fn is_empty(&self) -> bool {
        self.heap.is_empty()
    }

    /// Convert to sorted top-k results (descending by score)
    pub fn into_sorted_results(self) -> Vec<(DocId, f32, u16)> {
        let mut results: Vec<_> = self
            .heap
            .into_vec()
            .into_iter()
            .map(|e| (e.doc_id, e.score, e.ordinal))
            .collect();

        // Sort by score descending, then doc_id ascending
        results.sort_by(|a, b| {
            b.1.partial_cmp(&a.1)
                .unwrap_or(Ordering::Equal)
                .then_with(|| a.0.cmp(&b.0))
        });

        results
    }
}

/// Search result from WAND execution
#[derive(Debug, Clone, Copy)]
pub struct ScoredDoc {
    pub doc_id: DocId,
    pub score: f32,
    /// Ordinal for multi-valued fields (which vector in the field matched)
    pub ordinal: u16,
}

/// Unified Block-Max MaxScore executor for top-k retrieval
///
/// Combines three optimizations from the literature into one executor:
/// 1. **MaxScore partitioning** (Turtle & Flood 1995): terms split into essential
///    (must check) and non-essential (only scored if candidate is promising)
/// 2. **Block-max pruning** (Ding & Suel 2011): skip blocks where per-block
///    upper bounds can't beat the current threshold
/// 3. **Conjunction optimization** (Lucene/Grand 2023): progressively intersect
///    essential terms as threshold rises, skipping docs that lack enough terms
///
/// Works with any type implementing `ScoringIterator` (text or sparse).
/// Replaces separate WAND and MaxScore executors with better performance
/// across all query lengths.
pub struct BlockMaxScoreExecutor<S: ScoringIterator> {
    /// Scorers sorted by max_score ascending (non-essential first)
    scorers: Vec<S>,
    /// Cumulative max_score prefix sums: prefix_sums[i] = sum(max_score[0..=i])
    prefix_sums: Vec<f32>,
    /// Top-k collector
    collector: ScoreCollector,
    /// Heap factor for approximate search (SEISMIC-style)
    /// - 1.0 = exact search (default)
    /// - 0.8 = approximate, faster with minor recall loss
    heap_factor: f32,
}

/// Backwards-compatible alias for `BlockMaxScoreExecutor`
pub type WandExecutor<S> = BlockMaxScoreExecutor<S>;

impl<S: ScoringIterator> BlockMaxScoreExecutor<S> {
    /// Create a new executor with exact search (heap_factor = 1.0)
    pub fn new(scorers: Vec<S>, k: usize) -> Self {
        Self::with_heap_factor(scorers, k, 1.0)
    }

    /// Create a new 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 with_heap_factor(mut scorers: Vec<S>, k: usize, heap_factor: f32) -> Self {
        // Sort scorers by max_score ascending (non-essential terms first)
        scorers.sort_by(|a, b| {
            a.max_score()
                .partial_cmp(&b.max_score())
                .unwrap_or(Ordering::Equal)
        });

        // Compute prefix sums of max_scores
        let mut prefix_sums = Vec::with_capacity(scorers.len());
        let mut cumsum = 0.0f32;
        for s in &scorers {
            cumsum += s.max_score();
            prefix_sums.push(cumsum);
        }

        debug!(
            "Creating BlockMaxScoreExecutor: num_scorers={}, k={}, total_upper={:.4}, heap_factor={:.2}",
            scorers.len(),
            k,
            cumsum,
            heap_factor
        );

        Self {
            scorers,
            prefix_sums,
            collector: ScoreCollector::new(k),
            heap_factor: heap_factor.clamp(0.0, 1.0),
        }
    }

    /// Find partition point: [0..partition) = non-essential, [partition..n) = essential
    /// Non-essential terms have cumulative max_score <= threshold
    #[inline]
    fn find_partition(&self) -> usize {
        let threshold = self.collector.threshold() * self.heap_factor;
        self.prefix_sums
            .iter()
            .position(|&sum| sum > threshold)
            .unwrap_or(self.scorers.len())
    }

    /// Execute Block-Max MaxScore and return top-k results
    ///
    /// Algorithm:
    /// 1. Partition terms into essential/non-essential based on max_score
    /// 2. Find min_doc across essential scorers
    /// 3. Conjunction check: skip if not enough essential terms present
    /// 4. Block-max check: skip if block upper bounds can't beat threshold
    /// 5. Score essential scorers, check if non-essential scoring is needed
    /// 6. Score non-essential scorers, group by ordinal, insert results
    pub fn execute(mut self) -> Vec<ScoredDoc> {
        if self.scorers.is_empty() {
            debug!("BlockMaxScoreExecutor: no scorers, returning empty results");
            return Vec::new();
        }

        let n = self.scorers.len();
        let mut docs_scored = 0u64;
        let mut docs_skipped = 0u64;
        let mut blocks_skipped = 0u64;
        let mut conjunction_skipped = 0u64;

        // Pre-allocate scratch buffers outside the loop
        let mut ordinal_scores: Vec<(u16, f32)> = Vec::with_capacity(n * 2);

        loop {
            let partition = self.find_partition();

            // If all terms are non-essential, we're done
            if partition >= n {
                debug!("BlockMaxScore: all terms non-essential, early termination");
                break;
            }

            // Find minimum doc_id across essential scorers [partition..n)
            let mut min_doc = u32::MAX;
            for i in partition..n {
                let doc = self.scorers[i].doc();
                if doc < min_doc {
                    min_doc = doc;
                }
            }

            if min_doc == u32::MAX {
                break; // All essential scorers exhausted
            }

            let non_essential_upper = if partition > 0 {
                self.prefix_sums[partition - 1]
            } else {
                0.0
            };
            let adjusted_threshold = self.collector.threshold() * self.heap_factor;

            // --- Conjunction optimization (Lucene-style) ---
            // Check if enough essential terms are present at min_doc.
            // Sum max_scores of essential terms AT min_doc. If that plus
            // non-essential upper can't beat threshold, skip this doc.
            if self.collector.len() >= self.collector.k {
                let present_upper: f32 = (partition..n)
                    .filter(|&i| self.scorers[i].doc() == min_doc)
                    .map(|i| self.scorers[i].max_score())
                    .sum();

                if present_upper + non_essential_upper <= adjusted_threshold {
                    // Not enough essential terms present - advance past min_doc
                    for i in partition..n {
                        if self.scorers[i].doc() == min_doc {
                            self.scorers[i].advance();
                        }
                    }
                    conjunction_skipped += 1;
                    continue;
                }
            }

            // --- Block-max pruning ---
            // Sum block-max scores for essential scorers at min_doc.
            // If block-max sum + non-essential upper can't beat threshold, skip blocks.
            if self.collector.len() >= self.collector.k {
                let block_max_sum: f32 = (partition..n)
                    .filter(|&i| self.scorers[i].doc() == min_doc)
                    .map(|i| self.scorers[i].current_block_max_score())
                    .sum();

                if block_max_sum + non_essential_upper <= adjusted_threshold {
                    for i in partition..n {
                        if self.scorers[i].doc() == min_doc {
                            self.scorers[i].skip_to_next_block();
                        }
                    }
                    blocks_skipped += 1;
                    continue;
                }
            }

            // --- Score essential scorers ---
            // Drain all entries for min_doc from each essential scorer
            ordinal_scores.clear();

            for i in partition..n {
                if self.scorers[i].doc() == min_doc {
                    while self.scorers[i].doc() == min_doc {
                        ordinal_scores.push((self.scorers[i].ordinal(), self.scorers[i].score()));
                        self.scorers[i].advance();
                    }
                }
            }

            // Check if essential score + non-essential upper could beat threshold
            let essential_total: f32 = ordinal_scores.iter().map(|(_, s)| *s).sum();

            if self.collector.len() >= self.collector.k
                && essential_total + non_essential_upper <= adjusted_threshold
            {
                docs_skipped += 1;
                continue;
            }

            // --- Score non-essential scorers ---
            for i in 0..partition {
                let doc = self.scorers[i].seek(min_doc);
                if doc == min_doc {
                    while self.scorers[i].doc() == min_doc {
                        ordinal_scores.push((self.scorers[i].ordinal(), self.scorers[i].score()));
                        self.scorers[i].advance();
                    }
                }
            }

            // --- Group by ordinal and insert ---
            ordinal_scores.sort_unstable_by_key(|(ord, _)| *ord);
            let mut j = 0;
            while j < ordinal_scores.len() {
                let current_ord = ordinal_scores[j].0;
                let mut score = 0.0f32;
                while j < ordinal_scores.len() && ordinal_scores[j].0 == current_ord {
                    score += ordinal_scores[j].1;
                    j += 1;
                }

                trace!(
                    "Doc {}: ordinal={}, score={:.4}, threshold={:.4}",
                    min_doc, current_ord, score, adjusted_threshold
                );

                if self
                    .collector
                    .insert_with_ordinal(min_doc, score, current_ord)
                {
                    docs_scored += 1;
                } else {
                    docs_skipped += 1;
                }
            }
        }

        let results: Vec<ScoredDoc> = self
            .collector
            .into_sorted_results()
            .into_iter()
            .map(|(doc_id, score, ordinal)| ScoredDoc {
                doc_id,
                score,
                ordinal,
            })
            .collect();

        debug!(
            "BlockMaxScoreExecutor completed: scored={}, skipped={}, blocks_skipped={}, conjunction_skipped={}, returned={}, top_score={:.4}",
            docs_scored,
            docs_skipped,
            blocks_skipped,
            conjunction_skipped,
            results.len(),
            results.first().map(|r| r.score).unwrap_or(0.0)
        );

        results
    }
}

/// Scorer for full-text terms using WAND optimization
///
/// Wraps a `BlockPostingList` with BM25 parameters to implement `ScoringIterator`.
/// Enables MaxScore pruning for efficient top-k retrieval in OR queries.
pub struct TextTermScorer {
    /// Iterator over the posting list (owned)
    iter: crate::structures::BlockPostingIterator<'static>,
    /// IDF component for BM25
    idf: f32,
    /// Average field length for BM25 normalization
    avg_field_len: f32,
    /// Pre-computed max score (using max_tf from posting list)
    max_score: f32,
}

impl TextTermScorer {
    /// Create a new text term scorer with BM25 parameters
    pub fn new(
        posting_list: crate::structures::BlockPostingList,
        idf: f32,
        avg_field_len: f32,
    ) -> Self {
        // Compute max score using actual max_tf from posting list
        let max_tf = posting_list.max_tf() as f32;
        let doc_count = posting_list.doc_count();
        let max_score = super::bm25_upper_bound(max_tf.max(1.0), idf);

        debug!(
            "Created TextTermScorer: doc_count={}, max_tf={:.0}, idf={:.4}, avg_field_len={:.2}, max_score={:.4}",
            doc_count, max_tf, idf, avg_field_len, max_score
        );

        Self {
            iter: posting_list.into_iterator(),
            idf,
            avg_field_len,
            max_score,
        }
    }
}

impl ScoringIterator for TextTermScorer {
    #[inline]
    fn doc(&self) -> DocId {
        self.iter.doc()
    }

    #[inline]
    fn advance(&mut self) -> DocId {
        self.iter.advance()
    }

    #[inline]
    fn seek(&mut self, target: DocId) -> DocId {
        self.iter.seek(target)
    }

    #[inline]
    fn score(&self) -> f32 {
        let tf = self.iter.term_freq() as f32;
        // Use tf as proxy for doc length (common approximation when field lengths aren't stored)
        super::bm25_score(tf, self.idf, tf, self.avg_field_len)
    }

    #[inline]
    fn max_score(&self) -> f32 {
        self.max_score
    }

    #[inline]
    fn current_block_max_score(&self) -> f32 {
        // Use per-block max_tf for tighter Block-Max WAND bounds
        let block_max_tf = self.iter.current_block_max_tf() as f32;
        super::bm25_upper_bound(block_max_tf.max(1.0), self.idf)
    }

    #[inline]
    fn skip_to_next_block(&mut self) -> DocId {
        self.iter.skip_to_next_block()
    }
}

/// Scorer for sparse vector dimensions
///
/// Wraps a `BlockSparsePostingList` with a query weight to implement `ScoringIterator`.
pub struct SparseTermScorer<'a> {
    /// Iterator over the posting list
    iter: crate::structures::BlockSparsePostingIterator<'a>,
    /// Query weight for this dimension
    query_weight: f32,
    /// Global max score (query_weight * global_max_weight)
    max_score: f32,
}

impl<'a> SparseTermScorer<'a> {
    /// Create a new sparse term scorer
    ///
    /// Note: Assumes positive weights for WAND upper bound calculation.
    /// For negative query weights, uses absolute value to ensure valid upper bound.
    pub fn new(posting_list: &'a BlockSparsePostingList, query_weight: f32) -> Self {
        // Upper bound must account for sign: |query_weight| * max_weight
        // This ensures the bound is valid regardless of weight sign
        let max_score = query_weight.abs() * posting_list.global_max_weight();
        Self {
            iter: posting_list.iterator(),
            query_weight,
            max_score,
        }
    }

    /// Create from Arc reference (for use with shared posting lists)
    pub fn from_arc(posting_list: &'a Arc<BlockSparsePostingList>, query_weight: f32) -> Self {
        Self::new(posting_list.as_ref(), query_weight)
    }
}

impl ScoringIterator for SparseTermScorer<'_> {
    #[inline]
    fn doc(&self) -> DocId {
        self.iter.doc()
    }

    #[inline]
    fn ordinal(&self) -> u16 {
        self.iter.ordinal()
    }

    #[inline]
    fn advance(&mut self) -> DocId {
        self.iter.advance()
    }

    #[inline]
    fn seek(&mut self, target: DocId) -> DocId {
        self.iter.seek(target)
    }

    #[inline]
    fn score(&self) -> f32 {
        // Dot product contribution: query_weight * stored_weight
        self.query_weight * self.iter.weight()
    }

    #[inline]
    fn max_score(&self) -> f32 {
        self.max_score
    }

    #[inline]
    fn current_block_max_score(&self) -> f32 {
        // Use abs() for valid upper bound with negative weights
        self.iter
            .current_block_max_contribution(self.query_weight.abs())
    }

    #[inline]
    fn skip_to_next_block(&mut self) -> DocId {
        self.iter.skip_to_next_block()
    }
}

/// 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.
///
/// Reference: Mallia et al., "Faster Learned Sparse Retrieval with
/// Block-Max Pruning" (SIGIR 2024)
pub struct BmpExecutor {
    /// Posting lists for each query term
    posting_lists: Vec<Arc<BlockSparsePostingList>>,
    /// Query weight for each term
    query_weights: Vec<f32>,
    /// Number of results to return
    k: usize,
    /// Heap factor for approximate search
    heap_factor: f32,
}

/// Entry in the BMP priority queue: (term_index, block_index)
struct BmpBlockEntry {
    /// Upper bound contribution of this block
    contribution: f32,
    /// Index into posting_lists
    term_idx: usize,
    /// Block index within the posting list
    block_idx: usize,
}

impl PartialEq for BmpBlockEntry {
    fn eq(&self, other: &Self) -> bool {
        self.contribution == other.contribution
    }
}

impl Eq for BmpBlockEntry {}

impl Ord for BmpBlockEntry {
    fn cmp(&self, other: &Self) -> Ordering {
        // Max-heap: higher contributions come first
        self.contribution
            .partial_cmp(&other.contribution)
            .unwrap_or(Ordering::Equal)
    }
}

impl PartialOrd for BmpBlockEntry {
    fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
        Some(self.cmp(other))
    }
}

impl BmpExecutor {
    /// Create a new BMP executor
    pub fn new(
        posting_lists: Vec<Arc<BlockSparsePostingList>>,
        query_weights: Vec<f32>,
        k: usize,
        heap_factor: f32,
    ) -> Self {
        Self {
            posting_lists,
            query_weights,
            k,
            heap_factor: heap_factor.clamp(0.0, 1.0),
        }
    }

    /// Execute BMP and return top-k results
    pub fn execute(self) -> Vec<ScoredDoc> {
        use rustc_hash::FxHashMap;

        if self.posting_lists.is_empty() {
            return Vec::new();
        }

        let num_terms = self.posting_lists.len();

        // Build priority queue of all blocks across all terms
        let mut block_queue: BinaryHeap<BmpBlockEntry> = BinaryHeap::new();

        // Track remaining upper bound per term (sum of unprocessed block contributions)
        let mut remaining_max: Vec<f32> = Vec::with_capacity(num_terms);

        for (term_idx, pl) in self.posting_lists.iter().enumerate() {
            let qw = self.query_weights[term_idx].abs();
            let mut term_remaining = 0.0f32;

            for block_idx in 0..pl.num_blocks() {
                if let Some(block_max_weight) = pl.block_max_weight(block_idx) {
                    let contribution = qw * block_max_weight;
                    term_remaining += contribution;
                    block_queue.push(BmpBlockEntry {
                        contribution,
                        term_idx,
                        block_idx,
                    });
                }
            }
            remaining_max.push(term_remaining);
        }

        // Document accumulators: (doc_id, ordinal) -> accumulated_score
        // Using (doc_id, ordinal) as key ensures scores from different ordinals
        // are NOT mixed together for multi-valued sparse vector fields.
        let mut accumulators: FxHashMap<(DocId, u16), f32> = FxHashMap::default();
        let mut collector = ScoreCollector::new(self.k);
        let mut blocks_processed = 0u64;

        // Process blocks in contribution-descending order
        while let Some(entry) = block_queue.pop() {
            // Update remaining max for this term
            remaining_max[entry.term_idx] -= entry.contribution;

            // Early termination: check if total remaining across all terms
            // can beat the current threshold
            let total_remaining: f32 = remaining_max.iter().sum();
            let adjusted_threshold = collector.threshold() * self.heap_factor;
            if collector.len() >= self.k && total_remaining <= adjusted_threshold {
                debug!(
                    "BMP early termination after {} blocks: remaining={:.4} <= threshold={:.4}",
                    blocks_processed, total_remaining, adjusted_threshold
                );
                break;
            }

            // Decode this block and accumulate scores
            let pl = &self.posting_lists[entry.term_idx];
            let block = &pl.blocks[entry.block_idx];
            let doc_ids = block.decode_doc_ids();
            let weights = block.decode_weights();
            let ordinals = block.decode_ordinals();
            let qw = self.query_weights[entry.term_idx];

            for i in 0..block.header.count as usize {
                let score_contribution = qw * weights[i];
                *accumulators.entry((doc_ids[i], ordinals[i])).or_insert(0.0) += score_contribution;
            }

            blocks_processed += 1;
        }

        // Flush accumulators to collector
        for (&(doc_id, ordinal), &score) in &accumulators {
            collector.insert_with_ordinal(doc_id, score, ordinal);
        }

        let results: Vec<ScoredDoc> = collector
            .into_sorted_results()
            .into_iter()
            .map(|(doc_id, score, ordinal)| ScoredDoc {
                doc_id,
                score,
                ordinal,
            })
            .collect();

        debug!(
            "BmpExecutor completed: blocks_processed={}, accumulators={}, returned={}, top_score={:.4}",
            blocks_processed,
            accumulators.len(),
            results.len(),
            results.first().map(|r| r.score).unwrap_or(0.0)
        );

        results
    }
}

#[cfg(test)]
mod tests {
    use super::*;

    #[test]
    fn test_score_collector_basic() {
        let mut collector = ScoreCollector::new(3);

        collector.insert(1, 1.0);
        collector.insert(2, 2.0);
        collector.insert(3, 3.0);
        assert_eq!(collector.threshold(), 1.0);

        collector.insert(4, 4.0);
        assert_eq!(collector.threshold(), 2.0);

        let results = collector.into_sorted_results();
        assert_eq!(results.len(), 3);
        assert_eq!(results[0].0, 4); // Highest score
        assert_eq!(results[1].0, 3);
        assert_eq!(results[2].0, 2);
    }

    #[test]
    fn test_score_collector_threshold() {
        let mut collector = ScoreCollector::new(2);

        collector.insert(1, 5.0);
        collector.insert(2, 3.0);
        assert_eq!(collector.threshold(), 3.0);

        // Should not enter (score too low)
        assert!(!collector.would_enter(2.0));
        assert!(!collector.insert(3, 2.0));

        // Should enter (score high enough)
        assert!(collector.would_enter(4.0));
        assert!(collector.insert(4, 4.0));
        assert_eq!(collector.threshold(), 4.0);
    }

    #[test]
    fn test_heap_entry_ordering() {
        let mut heap = BinaryHeap::new();
        heap.push(HeapEntry {
            doc_id: 1,
            score: 3.0,
            ordinal: 0,
        });
        heap.push(HeapEntry {
            doc_id: 2,
            score: 1.0,
            ordinal: 0,
        });
        heap.push(HeapEntry {
            doc_id: 3,
            score: 2.0,
            ordinal: 0,
        });

        // Min-heap: lowest score should come out first
        assert_eq!(heap.pop().unwrap().score, 1.0);
        assert_eq!(heap.pop().unwrap().score, 2.0);
        assert_eq!(heap.pop().unwrap().score, 3.0);
    }
}