vtpl 0.1.0

Vector-Threaded Posting Lists — fused n-gram + vector search in a single index pass
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206

use std::collections::{BTreeMap, HashMap};

use serde::{Deserialize, Serialize};

use crate::ngram::trigrams;
use crate::posting::{ChunkId, PostingList, VtplEntry};
use crate::pq::PqCodebook;

#[derive(Debug, Clone)]
pub struct ScoredResult {
    pub chunk_id: ChunkId,
    pub score: f32,
    pub semantic_score: f32,
    pub pattern_score: f32,
}

/// The VTPL index: trigram → posting list with inline PQ codes.
#[derive(Serialize, Deserialize)]
pub struct VtplIndex {
    pub codebook: PqCodebook,
    postings: BTreeMap<String, PostingList>,
    num_chunks: u32,
    df: HashMap<String, u32>,
}

impl VtplIndex {
    pub fn new(codebook: PqCodebook) -> Self {
        Self {
            codebook,
            postings: BTreeMap::new(),
            num_chunks: 0,
            df: HashMap::new(),
        }
    }

    /// Construct from pre-built parts (used by ParallelBuilder).
    pub fn from_parts(
        codebook: PqCodebook,
        postings: BTreeMap<String, PostingList>,
        num_chunks: u32,
        df: HashMap<String, u32>,
    ) -> Self {
        Self { codebook, postings, num_chunks, df }
    }

    /// Insert a chunk. Embedding should be L2-normalized.
    pub fn insert(&mut self, chunk_id: ChunkId, text: &str, embedding: &[f32]) {
        let pq_code = self.codebook.encode(embedding);
        let grams = trigrams(text);

        for gram in &grams {
            self.postings
                .entry(gram.clone())
                .or_default()
                .push(VtplEntry::new(chunk_id, pq_code));

            *self.df.entry(gram.clone()).or_insert(0) += 1;
        }

        self.num_chunks += 1;
    }

    /// Sort posting lists for merge-intersection.
    pub fn finalize(&mut self) {
        for list in self.postings.values_mut() {
            list.sort();
        }
    }

    /// Fused query in a single pass over posting lists.
    ///
    /// `lambda` — weight for semantic: score = λ·cosine + (1-λ)·pattern
    pub fn query(
        &self,
        query_text: &str,
        query_embedding: &[f32],
        lambda: f32,
        top_k: usize,
    ) -> Vec<ScoredResult> {
        let grams = trigrams(query_text);
        let dt = self.codebook.build_distance_table(query_embedding);
        let total_query_grams = grams.len() as f32;

        if total_query_grams == 0.0 {
            return Vec::new();
        }

        let mut accum: HashMap<ChunkId, (f32, f32, u32)> = HashMap::new();

        for gram in &grams {
            if let Some(list) = self.postings.get(gram.as_str()) {
                let idf = idf(self.num_chunks, self.df.get(gram.as_str()).copied().unwrap_or(1));
                for entry in &list.entries {
                    let acc = accum.entry(entry.chunk_id).or_insert_with(|| {
                        let sem = dt.approximate_cosine(&entry.pq_code);
                        (sem, 0.0, 0)
                    });
                    acc.1 += idf;
                    acc.2 += 1;
                }
            }
        }

        let mut results: Vec<ScoredResult> = accum
            .into_iter()
            .map(|(chunk_id, (semantic_score, idf_sum, hit_count))| {
                let pattern_score = (hit_count as f32 / total_query_grams) * (idf_sum / hit_count as f32);
                let score = lambda * semantic_score + (1.0 - lambda) * pattern_score;
                ScoredResult { chunk_id, score, semantic_score, pattern_score }
            })
            .collect();

        results.sort_unstable_by(|a, b| b.score.partial_cmp(&a.score).unwrap());
        results.truncate(top_k);
        results
    }

    /// Text-only query (λ=0).
    pub fn text_query(&self, query_text: &str, top_k: usize) -> Vec<ScoredResult> {
        let grams = trigrams(query_text);
        let total_query_grams = grams.len() as f32;
        if total_query_grams == 0.0 {
            return Vec::new();
        }

        let mut accum: HashMap<ChunkId, (f32, u32)> = HashMap::new();
        for gram in &grams {
            if let Some(list) = self.postings.get(gram.as_str()) {
                let idf = idf(self.num_chunks, self.df.get(gram.as_str()).copied().unwrap_or(1));
                for entry in &list.entries {
                    let acc = accum.entry(entry.chunk_id).or_insert((0.0, 0));
                    acc.0 += idf;
                    acc.1 += 1;
                }
            }
        }

        let mut results: Vec<ScoredResult> = accum
            .into_iter()
            .map(|(chunk_id, (idf_sum, hit_count))| {
                let pattern_score = (hit_count as f32 / total_query_grams) * (idf_sum / hit_count as f32);
                ScoredResult { chunk_id, score: pattern_score, semantic_score: 0.0, pattern_score }
            })
            .collect();

        results.sort_unstable_by(|a, b| b.score.partial_cmp(&a.score).unwrap());
        results.truncate(top_k);
        results
    }

    /// Vector-only query — scan all unique PQ codes.
    pub fn vector_query(&self, query_embedding: &[f32], top_k: usize) -> Vec<ScoredResult> {
        let dt = self.codebook.build_distance_table(query_embedding);
        let mut seen: HashMap<ChunkId, f32> = HashMap::new();

        for list in self.postings.values() {
            for entry in &list.entries {
                seen.entry(entry.chunk_id)
                    .or_insert_with(|| dt.approximate_cosine(&entry.pq_code));
            }
        }

        let mut results: Vec<ScoredResult> = seen
            .into_iter()
            .map(|(chunk_id, sem)| ScoredResult {
                chunk_id, score: sem, semantic_score: sem, pattern_score: 0.0,
            })
            .collect();

        results.sort_unstable_by(|a, b| b.score.partial_cmp(&a.score).unwrap());
        results.truncate(top_k);
        results
    }

    pub fn num_postings(&self) -> usize { self.postings.len() }
    pub fn num_chunks(&self) -> u32 { self.num_chunks }
    pub fn total_entries(&self) -> usize { self.postings.values().map(|p| p.len()).sum() }
    pub fn pq_overhead_bytes(&self) -> usize { self.postings.values().map(|p| p.pq_overhead_bytes()).sum() }

    pub(crate) fn get_posting_list(&self, gram: &str) -> Option<&PostingList> {
        self.postings.get(gram)
    }

    pub(crate) fn get_df(&self, gram: &str) -> u32 {
        self.df.get(gram).copied().unwrap_or(0)
    }

    pub(crate) fn total_docs(&self) -> u32 {
        self.num_chunks
    }
}

impl VtplIndex {
    pub fn serialize(&self) -> Vec<u8> { bincode::serialize(self).expect("serialization failed") }
    pub fn deserialize(bytes: &[u8]) -> Self { bincode::deserialize(bytes).expect("deserialization failed") }
    pub fn save_to_file(&self, path: &str) -> std::io::Result<()> { std::fs::write(path, self.serialize()) }
    pub fn load_from_file(path: &str) -> std::io::Result<Self> { Ok(Self::deserialize(&std::fs::read(path)?)) }
}

/// BM25-style IDF: log((N - df + 0.5) / (df + 0.5) + 1)
#[inline]
pub(crate) fn idf(total_docs: u32, doc_freq: u32) -> f32 {
    let n = total_docs as f32;
    let df = doc_freq as f32;
    ((n - df + 0.5) / (df + 0.5) + 1.0).ln()
}