Skip to main content

hermes_core/structures/postings/sparse/
config.rs

1//! Configuration types for sparse vector posting lists
2
3use serde::{Deserialize, Serialize};
4
5/// Sparse vector index format
6///
7/// Determines the on-disk layout and query execution strategy:
8/// - **MaxScore**: Per-dimension variable-size blocks (DAAT — document-at-a-time).
9///   Default, optimal for general sparse retrieval with block-max pruning.
10/// - **Bmp**: Fixed doc_id range blocks (BAAT — block-at-a-time).
11///   Based on Mallia, Suel & Tonellotto (SIGIR 2024). Divides the document
12///   space into fixed-size blocks and processes them in decreasing upper-bound
13///   order, enabling aggressive early termination.
14#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Serialize, Deserialize)]
15pub enum SparseFormat {
16    /// Per-dimension variable-size blocks (existing format, DAAT MaxScore)
17    #[default]
18    MaxScore,
19    /// Fixed doc_id range blocks (BMP, BAAT block-at-a-time)
20    Bmp,
21}
22
23impl SparseFormat {
24    fn is_default(&self) -> bool {
25        *self == Self::MaxScore
26    }
27}
28
29/// Size of the index (term/dimension ID) in sparse vectors
30#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Serialize, Deserialize)]
31#[repr(u8)]
32pub enum IndexSize {
33    /// 16-bit index (0-65535), ideal for SPLADE vocabularies
34    U16 = 0,
35    /// 32-bit index (0-4B), for large vocabularies
36    #[default]
37    U32 = 1,
38}
39
40impl IndexSize {
41    /// Bytes per index
42    pub fn bytes(&self) -> usize {
43        match self {
44            IndexSize::U16 => 2,
45            IndexSize::U32 => 4,
46        }
47    }
48
49    /// Maximum value representable
50    pub fn max_value(&self) -> u32 {
51        match self {
52            IndexSize::U16 => u16::MAX as u32,
53            IndexSize::U32 => u32::MAX,
54        }
55    }
56
57    pub(crate) fn from_u8(v: u8) -> Option<Self> {
58        match v {
59            0 => Some(IndexSize::U16),
60            1 => Some(IndexSize::U32),
61            _ => None,
62        }
63    }
64}
65
66/// Quantization format for sparse vector weights
67///
68/// Research-validated compression/effectiveness trade-offs (Pati, 2025):
69/// - **UInt8**: 4x compression, ~1-2% nDCG@10 loss (RECOMMENDED for production)
70/// - **Float16**: 2x compression, <1% nDCG@10 loss
71/// - **Float32**: No compression, baseline effectiveness
72/// - **UInt4**: 8x compression, ~3-5% nDCG@10 loss (experimental)
73#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Serialize, Deserialize)]
74#[repr(u8)]
75pub enum WeightQuantization {
76    /// Full 32-bit float precision
77    #[default]
78    Float32 = 0,
79    /// 16-bit float (half precision) - 2x compression, <1% effectiveness loss
80    Float16 = 1,
81    /// 8-bit unsigned integer with scale factor - 4x compression, ~1-2% effectiveness loss (RECOMMENDED)
82    UInt8 = 2,
83    /// 4-bit unsigned integer with scale factor (packed, 2 per byte) - 8x compression, ~3-5% effectiveness loss
84    UInt4 = 3,
85}
86
87impl WeightQuantization {
88    /// Bytes per weight (approximate for UInt4)
89    pub fn bytes_per_weight(&self) -> f32 {
90        match self {
91            WeightQuantization::Float32 => 4.0,
92            WeightQuantization::Float16 => 2.0,
93            WeightQuantization::UInt8 => 1.0,
94            WeightQuantization::UInt4 => 0.5,
95        }
96    }
97
98    pub(crate) fn from_u8(v: u8) -> Option<Self> {
99        match v {
100            0 => Some(WeightQuantization::Float32),
101            1 => Some(WeightQuantization::Float16),
102            2 => Some(WeightQuantization::UInt8),
103            3 => Some(WeightQuantization::UInt4),
104            _ => None,
105        }
106    }
107}
108
109/// Query-time weighting strategy for sparse vector queries
110#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Serialize, Deserialize)]
111#[serde(rename_all = "snake_case")]
112pub enum QueryWeighting {
113    /// All terms get weight 1.0
114    #[default]
115    One,
116    /// Terms weighted by IDF (inverse document frequency) from global index statistics
117    /// Uses ln(N/df) where N = total docs, df = docs containing dimension
118    Idf,
119    /// Terms weighted by pre-computed IDF from model's idf.json file
120    /// Loaded from HuggingFace model repo. No fallback to global stats.
121    IdfFile,
122}
123
124/// Query-time configuration for sparse vectors
125///
126/// Research-validated query optimization strategies:
127/// - **weight_threshold (0.01-0.05)**: Drop query dimensions with weight below threshold
128///   - Filters low-IDF tokens that add latency without improving relevance
129/// - **max_query_dims (10-20)**: Process only top-k dimensions by weight
130///   - 30-50% latency reduction with <2% nDCG loss (Qiao et al., 2023)
131/// - **heap_factor (0.8)**: Skip blocks with low max score contribution
132///   - ~20% speedup with minor recall loss (SEISMIC-style)
133#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
134pub struct SparseQueryConfig {
135    /// HuggingFace tokenizer path/name for query-time tokenization
136    /// Example: "Alibaba-NLP/gte-Qwen2-1.5B-instruct"
137    #[serde(default, skip_serializing_if = "Option::is_none")]
138    pub tokenizer: Option<String>,
139    /// Weighting strategy for tokenized query terms
140    #[serde(default)]
141    pub weighting: QueryWeighting,
142    /// Heap factor for approximate search (SEISMIC-style optimization)
143    /// A block is skipped if its max possible score < heap_factor * threshold
144    ///
145    /// Research recommendation:
146    /// - 1.0 = exact search (default)
147    /// - 0.8 = approximate, ~20% faster with minor recall loss (RECOMMENDED for production)
148    /// - 0.5 = very approximate, much faster but higher recall loss
149    #[serde(default = "default_heap_factor")]
150    pub heap_factor: f32,
151    /// Minimum weight for query dimensions (query-time pruning)
152    /// Dimensions with abs(weight) below this threshold are dropped before search.
153    /// Useful for filtering low-IDF tokens that add latency without improving relevance.
154    ///
155    /// - 0.0 = no filtering (default)
156    /// - 0.01-0.05 = recommended for SPLADE/learned sparse models
157    #[serde(default)]
158    pub weight_threshold: f32,
159    /// Maximum number of query dimensions to process (query pruning)
160    /// Processes only the top-k dimensions by weight
161    ///
162    /// Research recommendation (Multiple papers 2022-2024):
163    /// - None = process all dimensions (default, exact)
164    /// - Some(10-20) = process top 10-20 dimensions only (RECOMMENDED for SPLADE)
165    ///   - 30-50% latency reduction
166    ///   - <2% nDCG@10 loss
167    #[serde(default, skip_serializing_if = "Option::is_none")]
168    pub max_query_dims: Option<usize>,
169    /// Fraction of query dimensions to keep (0.0-1.0), same semantics as
170    /// indexing-time `pruning`: sort by abs(weight) descending,
171    /// keep top fraction. None or 1.0 = no pruning.
172    #[serde(default, skip_serializing_if = "Option::is_none")]
173    pub pruning: Option<f32>,
174    /// Minimum number of query dimensions before pruning and weight_threshold
175    /// filtering are applied. Protects short queries from losing most signal.
176    ///
177    /// Default: 4. Set to 0 to always apply pruning/filtering.
178    #[serde(default = "default_min_terms")]
179    pub min_query_dims: usize,
180    /// Maximum number of superblocks to visit (LSP/0 gamma cap).
181    /// 0 = unlimited (default). Only applies to BMP format.
182    #[serde(default)]
183    pub max_superblocks: usize,
184}
185
186fn default_heap_factor() -> f32 {
187    1.0
188}
189
190impl Default for SparseQueryConfig {
191    fn default() -> Self {
192        Self {
193            tokenizer: None,
194            weighting: QueryWeighting::One,
195            heap_factor: 1.0,
196            weight_threshold: 0.0,
197            max_query_dims: None,
198            pruning: None,
199            min_query_dims: 4,
200            max_superblocks: 0,
201        }
202    }
203}
204
205/// Configuration for sparse vector storage
206///
207/// Research-validated optimizations for learned sparse retrieval (SPLADE, uniCOIL, etc.):
208/// - **Weight threshold (0.01-0.05)**: Removes ~30-50% of postings with minimal nDCG impact
209/// - **Posting list pruning (0.1)**: Keeps top 10% per dimension, 50-70% index reduction, <1% nDCG loss
210/// - **Query pruning (top 10-20 dims)**: 30-50% latency reduction, <2% nDCG loss
211/// - **UInt8 quantization**: 4x compression, 1-2% nDCG loss (optimal trade-off)
212#[derive(Debug, Clone, PartialEq, Serialize, Deserialize)]
213pub struct SparseVectorConfig {
214    /// Index format: MaxScore (DAAT) or BMP (BAAT)
215    #[serde(default, skip_serializing_if = "SparseFormat::is_default")]
216    pub format: SparseFormat,
217    /// Size of dimension/term indices
218    pub index_size: IndexSize,
219    /// Quantization for weights (see WeightQuantization docs for trade-offs)
220    pub weight_quantization: WeightQuantization,
221    /// Minimum weight threshold - weights below this value are not indexed
222    ///
223    /// Research recommendation (Guo et al., 2022; SPLADE v2):
224    /// - 0.01-0.05 for SPLADE models removes ~30-50% of postings
225    /// - Minimal impact on nDCG@10 (<1% loss)
226    /// - Major reduction in index size and query latency
227    #[serde(default)]
228    pub weight_threshold: f32,
229    /// Block size for posting lists (must be power of 2, default 128 for SIMD)
230    /// Larger blocks = better compression, smaller blocks = faster seeks.
231    /// Used by MaxScore format only.
232    #[serde(default = "default_block_size")]
233    pub block_size: usize,
234    /// BMP block size: number of consecutive doc_ids per block (must be power of 2).
235    /// Default 64. Only used when format = Bmp.
236    /// Smaller = better pruning granularity, larger = less overhead.
237    #[serde(default = "default_bmp_block_size")]
238    pub bmp_block_size: u32,
239    /// Maximum BMP grid memory in bytes. If the grid (num_dims × num_blocks)
240    /// would exceed this, bmp_block_size is automatically increased to cap memory.
241    /// Default: 256MB. Set to 0 to disable the cap.
242    #[serde(default = "default_max_bmp_grid_bytes")]
243    pub max_bmp_grid_bytes: u64,
244    /// BMP superblock size: number of consecutive blocks grouped for hierarchical
245    /// pruning (Carlson et al., SIGIR 2025). Must be power of 2.
246    /// Default 64. Set to 0 to disable superblock pruning (flat BMP scoring).
247    /// Only used when format = Bmp.
248    #[serde(default = "default_bmp_superblock_size")]
249    pub bmp_superblock_size: u32,
250    /// Static pruning: fraction of postings to keep per inverted list (SEISMIC-style)
251    /// Lists are sorted by weight descending and truncated to top fraction.
252    ///
253    /// Research recommendation (SPLADE v2, Formal et al., 2021):
254    /// - None = keep all postings (default, exact)
255    /// - Some(0.1) = keep top 10% of postings per dimension
256    ///   - 50-70% index size reduction
257    ///   - <1% nDCG@10 loss
258    ///   - Exploits "concentration of importance" in learned representations
259    ///
260    /// Applied only during initial segment build, not during merge.
261    #[serde(default, skip_serializing_if = "Option::is_none")]
262    pub pruning: Option<f32>,
263    /// Query-time configuration (tokenizer, weighting)
264    #[serde(default, skip_serializing_if = "Option::is_none")]
265    pub query_config: Option<SparseQueryConfig>,
266    /// Fixed vocabulary size (number of dimensions) for BMP format.
267    ///
268    /// When set, all BMP segments use the same grid dimensions (rows = dims),
269    /// enabling zero-copy block-copy merge. The grid is indexed by dim_id directly
270    /// (no dim_ids Section C needed).
271    ///
272    /// Required for BMP format. Typical values:
273    /// - SPLADE/BERT: 30522 or 105879 (WordPiece / Unigram vocabulary)
274    /// - uniCOIL: 30522
275    /// - Custom models: set to vocabulary size
276    ///
277    /// If None, BMP builder derives dims from observed data (V10 behavior).
278    #[serde(default, skip_serializing_if = "Option::is_none")]
279    pub dims: Option<u32>,
280    /// Fixed max weight scale for BMP format.
281    ///
282    /// When set, all BMP segments use the same quantization scale
283    /// (`max_weight_scale = max_weight`), eliminating rescaling during merge.
284    ///
285    /// For SPLADE models: 5.0 (covers typical weight range 0-5).
286    /// If None, BMP builder derives scale from data (V10 behavior).
287    #[serde(default, skip_serializing_if = "Option::is_none")]
288    pub max_weight: Option<f32>,
289    /// Minimum number of postings in a dimension before pruning and
290    /// weight_threshold filtering are applied. Protects dimensions with
291    /// very few postings from losing most of their signal.
292    ///
293    /// Default: 4. Set to 0 to always apply pruning/filtering.
294    #[serde(default = "default_min_terms")]
295    pub min_terms: usize,
296}
297
298fn default_block_size() -> usize {
299    128
300}
301
302fn default_bmp_block_size() -> u32 {
303    64
304}
305
306fn default_max_bmp_grid_bytes() -> u64 {
307    0 // disabled by default — masks eliminate DRAM stalls during scoring
308}
309
310fn default_bmp_superblock_size() -> u32 {
311    64
312}
313
314fn default_min_terms() -> usize {
315    4
316}
317
318impl Default for SparseVectorConfig {
319    fn default() -> Self {
320        Self {
321            format: SparseFormat::MaxScore,
322            index_size: IndexSize::U32,
323            weight_quantization: WeightQuantization::Float32,
324            weight_threshold: 0.0,
325            block_size: 128,
326            bmp_block_size: 64,
327            max_bmp_grid_bytes: 0,
328            bmp_superblock_size: 64,
329            pruning: None,
330            query_config: None,
331
332            dims: None,
333            max_weight: None,
334            min_terms: 4,
335        }
336    }
337}
338
339impl SparseVectorConfig {
340    /// SPLADE-optimized config with research-validated defaults
341    ///
342    /// Optimized for SPLADE, uniCOIL, and similar learned sparse retrieval models.
343    /// Based on research findings from:
344    /// - Pati (2025): UInt8 quantization = 4x compression, 1-2% nDCG loss
345    /// - Formal et al. (2021): SPLADE v2 posting list pruning
346    /// - Qiao et al. (2023): Query dimension pruning and approximate search
347    /// - Guo et al. (2022): Weight thresholding for efficiency
348    ///
349    /// Expected performance vs. full precision baseline:
350    /// - Index size: ~15-25% of original (combined effect of all optimizations)
351    /// - Query latency: 40-60% faster
352    /// - Effectiveness: 2-4% nDCG@10 loss (typically acceptable for production)
353    ///
354    /// Vocabulary: ~30K dimensions (fits in u16)
355    pub fn splade() -> Self {
356        Self {
357            format: SparseFormat::MaxScore,
358            index_size: IndexSize::U16,
359            weight_quantization: WeightQuantization::UInt8,
360            weight_threshold: 0.01, // Remove ~30-50% of low-weight postings
361            block_size: 128,
362            bmp_block_size: 64,
363            max_bmp_grid_bytes: 0,
364            bmp_superblock_size: 64,
365            pruning: Some(0.1), // Keep top 10% per dimension
366            query_config: Some(SparseQueryConfig {
367                tokenizer: None,
368                weighting: QueryWeighting::One,
369                heap_factor: 0.8,         // 20% faster approximate search
370                weight_threshold: 0.01,   // Drop low-IDF query tokens
371                max_query_dims: Some(20), // Process top 20 query dimensions
372                pruning: Some(0.1),       // Keep top 10% of query dims
373                min_query_dims: 4,
374                max_superblocks: 0,
375            }),
376
377            dims: None,
378            max_weight: None,
379            min_terms: 4,
380        }
381    }
382
383    /// SPLADE-optimized config with BMP (Block-Max Pruning) format
384    ///
385    /// Same optimization settings as `splade()` but uses the BMP block-at-a-time
386    /// format (Mallia, Suel & Tonellotto, SIGIR 2024) instead of MaxScore.
387    /// BMP divides the document space into fixed-size blocks and processes them
388    /// in decreasing upper-bound order, enabling aggressive early termination.
389    pub fn splade_bmp() -> Self {
390        Self {
391            format: SparseFormat::Bmp,
392            index_size: IndexSize::U16,
393            weight_quantization: WeightQuantization::UInt8,
394            weight_threshold: 0.01,
395            block_size: 128,
396            bmp_block_size: 64,
397            max_bmp_grid_bytes: 0,
398            bmp_superblock_size: 64,
399            pruning: Some(0.1),
400            query_config: Some(SparseQueryConfig {
401                tokenizer: None,
402                weighting: QueryWeighting::One,
403                heap_factor: 0.8,
404                weight_threshold: 0.01,
405                max_query_dims: Some(20),
406                pruning: Some(0.1),
407                min_query_dims: 4,
408                max_superblocks: 0,
409            }),
410
411            dims: Some(105879),
412            max_weight: Some(5.0),
413            min_terms: 4,
414        }
415    }
416
417    /// Compact config: Maximum compression (experimental)
418    ///
419    /// Uses aggressive UInt4 quantization for smallest possible index size.
420    /// Expected trade-offs:
421    /// - Index size: ~10-15% of Float32 baseline
422    /// - Effectiveness: ~3-5% nDCG@10 loss
423    ///
424    /// Recommended for: Memory-constrained environments, cache-heavy workloads
425    pub fn compact() -> Self {
426        Self {
427            format: SparseFormat::MaxScore,
428            index_size: IndexSize::U16,
429            weight_quantization: WeightQuantization::UInt4,
430            weight_threshold: 0.02, // Slightly higher threshold for UInt4
431            block_size: 128,
432            bmp_block_size: 64,
433            max_bmp_grid_bytes: 0,
434            bmp_superblock_size: 64,
435            pruning: Some(0.15), // Keep top 15% per dimension
436            query_config: Some(SparseQueryConfig {
437                tokenizer: None,
438                weighting: QueryWeighting::One,
439                heap_factor: 0.7,         // More aggressive approximate search
440                weight_threshold: 0.02,   // Drop low-IDF query tokens
441                max_query_dims: Some(15), // Fewer query dimensions
442                pruning: Some(0.15),      // Keep top 15% of query dims
443                min_query_dims: 4,
444                max_superblocks: 0,
445            }),
446
447            dims: None,
448            max_weight: None,
449            min_terms: 4,
450        }
451    }
452
453    /// Full precision config: No compression, baseline effectiveness
454    ///
455    /// Use for: Research baselines, when effectiveness is critical
456    pub fn full_precision() -> Self {
457        Self {
458            format: SparseFormat::MaxScore,
459            index_size: IndexSize::U32,
460            weight_quantization: WeightQuantization::Float32,
461            weight_threshold: 0.0,
462            block_size: 128,
463            bmp_block_size: 64,
464            max_bmp_grid_bytes: 0,
465            bmp_superblock_size: 64,
466            pruning: None,
467            query_config: None,
468
469            dims: None,
470            max_weight: None,
471            min_terms: 4,
472        }
473    }
474
475    /// Conservative config: Mild optimizations, minimal effectiveness loss
476    ///
477    /// Balances compression and effectiveness with conservative defaults.
478    /// Expected trade-offs:
479    /// - Index size: ~40-50% of Float32 baseline
480    /// - Query latency: ~20-30% faster
481    /// - Effectiveness: <1% nDCG@10 loss
482    ///
483    /// Recommended for: Production deployments prioritizing effectiveness
484    pub fn conservative() -> Self {
485        Self {
486            format: SparseFormat::MaxScore,
487            index_size: IndexSize::U32,
488            weight_quantization: WeightQuantization::Float16,
489            weight_threshold: 0.005, // Minimal pruning
490            block_size: 128,
491            bmp_block_size: 64,
492            max_bmp_grid_bytes: 0,
493            bmp_superblock_size: 64,
494            pruning: None, // No posting list pruning
495            query_config: Some(SparseQueryConfig {
496                tokenizer: None,
497                weighting: QueryWeighting::One,
498                heap_factor: 0.9,         // Nearly exact search
499                weight_threshold: 0.005,  // Minimal query pruning
500                max_query_dims: Some(50), // Process more dimensions
501                pruning: None,            // No fraction-based pruning
502                min_query_dims: 4,
503                max_superblocks: 0,
504            }),
505
506            dims: None,
507            max_weight: None,
508            min_terms: 4,
509        }
510    }
511
512    /// Set weight threshold (builder pattern)
513    pub fn with_weight_threshold(mut self, threshold: f32) -> Self {
514        self.weight_threshold = threshold;
515        self
516    }
517
518    /// Set posting list pruning fraction (builder pattern)
519    /// e.g., 0.1 = keep top 10% of postings per dimension
520    pub fn with_pruning(mut self, fraction: f32) -> Self {
521        self.pruning = Some(fraction.clamp(0.0, 1.0));
522        self
523    }
524
525    /// Bytes per entry (index + weight)
526    pub fn bytes_per_entry(&self) -> f32 {
527        self.index_size.bytes() as f32 + self.weight_quantization.bytes_per_weight()
528    }
529
530    /// Serialize config to a single byte.
531    ///
532    /// Layout: bits 7-4 = IndexSize, bit 3 = format (0=MaxScore, 1=BMP), bits 2-0 = WeightQuantization
533    pub fn to_byte(&self) -> u8 {
534        let format_bit = if self.format == SparseFormat::Bmp {
535            0x08
536        } else {
537            0
538        };
539        ((self.index_size as u8) << 4) | format_bit | (self.weight_quantization as u8)
540    }
541
542    /// Deserialize config from a single byte.
543    ///
544    /// Note: weight_threshold, block_size, bmp_block_size, and query_config are not
545    /// serialized in the byte — they come from the schema.
546    pub fn from_byte(b: u8) -> Option<Self> {
547        let index_size = IndexSize::from_u8((b >> 4) & 0x03)?;
548        let format = if b & 0x08 != 0 {
549            SparseFormat::Bmp
550        } else {
551            SparseFormat::MaxScore
552        };
553        let weight_quantization = WeightQuantization::from_u8(b & 0x07)?;
554        Some(Self {
555            format,
556            index_size,
557            weight_quantization,
558            weight_threshold: 0.0,
559            block_size: 128,
560            bmp_block_size: 64,
561            max_bmp_grid_bytes: 0,
562            bmp_superblock_size: 64,
563            pruning: None,
564            query_config: None,
565
566            dims: None,
567            max_weight: None,
568            min_terms: 4,
569        })
570    }
571
572    /// Set block size (builder pattern)
573    /// Must be power of 2, recommended: 64, 128, 256
574    pub fn with_block_size(mut self, size: usize) -> Self {
575        self.block_size = size.next_power_of_two();
576        self
577    }
578
579    /// Set query configuration (builder pattern)
580    pub fn with_query_config(mut self, config: SparseQueryConfig) -> Self {
581        self.query_config = Some(config);
582        self
583    }
584}
585
586/// A sparse vector entry: (dimension_id, weight)
587#[derive(Debug, Clone, Copy, PartialEq)]
588pub struct SparseEntry {
589    pub dim_id: u32,
590    pub weight: f32,
591}
592
593/// Sparse vector representation
594#[derive(Debug, Clone, Default)]
595pub struct SparseVector {
596    pub(super) entries: Vec<SparseEntry>,
597}
598
599impl SparseVector {
600    /// Create a new sparse vector
601    pub fn new() -> Self {
602        Self {
603            entries: Vec::new(),
604        }
605    }
606
607    /// Create with pre-allocated capacity
608    pub fn with_capacity(capacity: usize) -> Self {
609        Self {
610            entries: Vec::with_capacity(capacity),
611        }
612    }
613
614    /// Create from dimension IDs and weights
615    pub fn from_entries(dim_ids: &[u32], weights: &[f32]) -> Self {
616        assert_eq!(dim_ids.len(), weights.len());
617        let mut entries: Vec<SparseEntry> = dim_ids
618            .iter()
619            .zip(weights.iter())
620            .map(|(&dim_id, &weight)| SparseEntry { dim_id, weight })
621            .collect();
622        // Sort by dimension ID for efficient intersection
623        entries.sort_by_key(|e| e.dim_id);
624        Self { entries }
625    }
626
627    /// Add an entry (must maintain sorted order by dim_id)
628    pub fn push(&mut self, dim_id: u32, weight: f32) {
629        debug_assert!(
630            self.entries.is_empty() || self.entries.last().unwrap().dim_id < dim_id,
631            "Entries must be added in sorted order by dim_id"
632        );
633        self.entries.push(SparseEntry { dim_id, weight });
634    }
635
636    /// Number of non-zero entries
637    pub fn len(&self) -> usize {
638        self.entries.len()
639    }
640
641    /// Check if empty
642    pub fn is_empty(&self) -> bool {
643        self.entries.is_empty()
644    }
645
646    /// Iterate over entries
647    pub fn iter(&self) -> impl Iterator<Item = &SparseEntry> {
648        self.entries.iter()
649    }
650
651    /// Sort by dimension ID (required for posting list encoding)
652    pub fn sort_by_dim(&mut self) {
653        self.entries.sort_by_key(|e| e.dim_id);
654    }
655
656    /// Sort by weight descending
657    pub fn sort_by_weight_desc(&mut self) {
658        self.entries.sort_by(|a, b| {
659            b.weight
660                .partial_cmp(&a.weight)
661                .unwrap_or(std::cmp::Ordering::Equal)
662        });
663    }
664
665    /// Get top-k entries by weight
666    pub fn top_k(&self, k: usize) -> Vec<SparseEntry> {
667        let mut sorted = self.entries.clone();
668        sorted.sort_by(|a, b| {
669            b.weight
670                .partial_cmp(&a.weight)
671                .unwrap_or(std::cmp::Ordering::Equal)
672        });
673        sorted.truncate(k);
674        sorted
675    }
676
677    /// Compute dot product with another sparse vector
678    pub fn dot(&self, other: &SparseVector) -> f32 {
679        let mut result = 0.0f32;
680        let mut i = 0;
681        let mut j = 0;
682
683        while i < self.entries.len() && j < other.entries.len() {
684            let a = &self.entries[i];
685            let b = &other.entries[j];
686
687            match a.dim_id.cmp(&b.dim_id) {
688                std::cmp::Ordering::Less => i += 1,
689                std::cmp::Ordering::Greater => j += 1,
690                std::cmp::Ordering::Equal => {
691                    result += a.weight * b.weight;
692                    i += 1;
693                    j += 1;
694                }
695            }
696        }
697
698        result
699    }
700
701    /// L2 norm squared
702    pub fn norm_squared(&self) -> f32 {
703        self.entries.iter().map(|e| e.weight * e.weight).sum()
704    }
705
706    /// L2 norm
707    pub fn norm(&self) -> f32 {
708        self.norm_squared().sqrt()
709    }
710
711    /// Prune dimensions below a weight threshold
712    pub fn filter_by_weight(&self, min_weight: f32) -> Self {
713        let entries: Vec<SparseEntry> = self
714            .entries
715            .iter()
716            .filter(|e| e.weight.abs() >= min_weight)
717            .cloned()
718            .collect();
719        Self { entries }
720    }
721}
722
723impl From<Vec<(u32, f32)>> for SparseVector {
724    fn from(pairs: Vec<(u32, f32)>) -> Self {
725        Self {
726            entries: pairs
727                .into_iter()
728                .map(|(dim_id, weight)| SparseEntry { dim_id, weight })
729                .collect(),
730        }
731    }
732}
733
734impl From<SparseVector> for Vec<(u32, f32)> {
735    fn from(vec: SparseVector) -> Self {
736        vec.entries
737            .into_iter()
738            .map(|e| (e.dim_id, e.weight))
739            .collect()
740    }
741}