laurus 0.12.0

Unified search library for lexical, vector, and semantic retrieval
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
207
208
209
210
211
212
213
214
215
216
217
218
219

//! Shared writer / reader helpers for the HNSW PQ segment payload
//! (Issue #481 Stage 3, parallel to
//! [`crate::vector::index::quantized_io`] for the Scalar8Bit variant).
//!
//! # On-disk layout for the PQ region
//!
//! ```text
//! [ VectorSegmentHeader        (LVS1 + PQ params + codebook) ]
//! repeat num_vectors times:
//!   [ doc_id            u64 LE  8 bytes ]
//!   [ field ref: v3+ = field_id u16 (per-segment dictionary,     ]
//!   [   Issue #633); v1/v2 = name_len u32 + UTF-8 name           ]
//!   [ codes                   m bytes (per sub-vector centroid index) ]
//! ```
//!
//! Unlike the Stage 1 layout, PQ records have no per-vector metadata
//! block — the codebook covers all the affine corrections SQ needs to
//! reconstruct distances.

use std::io::{Read, Write};

use crate::error::Result;
use crate::vector::core::quantization::{PqParams, QuantizationMethod, VectorQuantizer, pq_decode};
use crate::vector::core::vector::Vector;
use crate::vector::index::pq_codebook::SharedPqCodebook;

/// Bytes consumed by the per-vector codes (everything after the
/// field_name string ends). Equal to `m`.
///
/// Currently only used by the in-module roundtrip test; the writer
/// (Phase 3b) inlines `codes.len()` and the reader inlines
/// `params.m as usize` since both already carry that information.
#[inline]
#[cfg(test)]
pub(super) const fn pq_record_payload_size(m: u16) -> usize {
    m as usize
}

/// Encode `vectors` against a PQ codebook, either training a fresh one
/// or reusing a pre-trained [`SharedPqCodebook`] (Issue #631).
///
/// Returns `(params, codebook, codes)` where `codes` is one
/// `Vec<u8>` of length `params.m` per input vector, in the same order
/// as the input slice. The caller pairs each entry back with its
/// `(doc_id, field_name)` triple before writing.
///
/// # Arguments
///
/// * `vectors` - The vectors to encode.
/// * `dim` - Original vector dimension.
/// * `subvector_count` - PQ `m` (must divide `dim`).
/// * `shared` - When `Some`, skip training entirely and encode against
///   this pre-trained codebook via [`VectorQuantizer::from_pq_codebook`]
///   (Issue #631 — training a fresh codebook on every segment/merge
///   costs multiple seconds and recurs at every tier of the
///   segment-per-commit merge hierarchy). When `None`, train a fresh
///   codebook on `vectors` as before.
///
/// # Errors
///
/// * Forwards [`SharedPqCodebook::validate_for`]'s error if `shared`'s
///   geometry does not match `dim`/`subvector_count`.
/// * Otherwise forwards any error from [`VectorQuantizer::train`] /
///   [`VectorQuantizer::quantize`] (e.g. empty input, dimension
///   mismatch, non-finite values, `dim % m != 0`).
pub(super) fn quantize_segment_pq(
    vectors: &[Vector],
    dim: usize,
    subvector_count: usize,
    shared: Option<&SharedPqCodebook>,
) -> Result<(PqParams, Vec<f32>, Vec<Vec<u8>>)> {
    let quantizer = match shared {
        Some(cb) => {
            cb.validate_for(dim, subvector_count, 256)?;
            VectorQuantizer::from_pq_codebook(dim, cb.params, cb.codebook.clone())?
        }
        None => {
            let mut quantizer = VectorQuantizer::new(
                QuantizationMethod::ProductQuantization { subvector_count },
                dim,
            );
            quantizer.train(vectors)?;
            quantizer
        }
    };
    let (params, codebook_slice) = quantizer
        .pq_state()
        .expect("PQ quantizer is trained or built from an existing codebook, so pq_state is set");
    let params = *params;
    let codebook: Vec<f32> = codebook_slice.to_vec();
    let codes: Vec<Vec<u8>> = vectors
        .iter()
        .map(|v| quantizer.quantize(v).map(|(c, _)| c))
        .collect::<Result<_>>()?;
    Ok((params, codebook, codes))
}

/// Write the PQ codes tail of one vector record.
///
/// The caller is responsible for writing the preceding `doc_id` /
/// `field_name_len` / `field_name` fields and the per-segment
/// [`crate::vector::index::format::VectorSegmentHeader`].
pub(super) fn write_pq_record<W: Write>(output: &mut W, codes: &[u8]) -> Result<()> {
    output.write_all(codes)?;
    Ok(())
}

/// Read the PQ codes tail of one vector record and reconstruct it
/// back to a `Vec<f32>` of length `params.original_dim()` via the
/// codebook. Used by load paths that keep the in-memory representation
/// as f32 (the HNSW writer's `load` path on rebuild).
pub(super) fn read_dequantized_pq_vector<R: Read>(
    input: &mut R,
    params: PqParams,
    codebook: &[f32],
) -> Result<Vec<f32>> {
    let mut codes = vec![0u8; params.m as usize];
    input.read_exact(&mut codes)?;
    Ok(pq_decode(&codes, params, codebook))
}

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

    fn vec_of(values: &[f32]) -> Vector {
        Vector::new(values.to_vec())
    }

    #[test]
    fn quantize_segment_pq_returns_params_codebook_codes() {
        let vectors = vec![
            vec_of(&[10.0, 10.0, 20.0, 20.0]),
            vec_of(&[-10.0, -10.0, -20.0, -20.0]),
            vec_of(&[10.1, 10.1, 20.1, 20.1]),
        ];
        let (params, codebook, codes) = quantize_segment_pq(&vectors, 4, 2, None).unwrap();
        assert_eq!(params.m, 2);
        assert_eq!(params.k, 256);
        assert_eq!(params.sub_dim, 2);
        assert_eq!(codebook.len(), params.codebook_len());
        assert_eq!(codes.len(), 3);
        for c in &codes {
            assert_eq!(c.len(), 2);
        }
    }

    #[test]
    fn write_then_read_pq_roundtrips_to_codebook() {
        let vectors = vec![
            vec_of(&[10.0, 10.0, 20.0, 20.0]),
            vec_of(&[-10.0, -10.0, -20.0, -20.0]),
        ];
        let (params, codebook, codes) = quantize_segment_pq(&vectors, 4, 2, None).unwrap();

        let mut buf = Vec::new();
        for c in &codes {
            write_pq_record(&mut buf, c).unwrap();
        }
        assert_eq!(buf.len(), codes.len() * pq_record_payload_size(params.m));

        let mut cursor = Cursor::new(&buf);
        for (i, original) in vectors.iter().enumerate() {
            let recovered = read_dequantized_pq_vector(&mut cursor, params, &codebook).unwrap();
            assert_eq!(recovered.len(), 4);
            // PQ is lossy; just require that the reconstruction is on
            // the correct side of zero for this well-separated fixture.
            assert_eq!(
                recovered[0].signum(),
                original.data[0].signum(),
                "vector {i} sign flip on first coord"
            );
        }
    }

    #[test]
    fn payload_size_equals_m() {
        assert_eq!(pq_record_payload_size(0), 0);
        assert_eq!(pq_record_payload_size(16), 16);
    }

    /// Issue #631: encoding against a pre-trained [`SharedPqCodebook`]
    /// must produce byte-identical output to training a fresh codebook
    /// on the exact same corpus -- the strongest cheap correctness
    /// check available, since `pq_train_codebook` is deterministic
    /// (seeded k-means).
    #[test]
    fn shared_codebook_produces_byte_identical_codes_to_fresh_training() {
        let mut state: u64 = 0x1234_5678_9ABC_DEF0;
        let vectors: Vec<Vector> = (0..300)
            .map(|_| {
                let data: Vec<f32> = (0..32)
                    .map(|_| {
                        state = state
                            .wrapping_mul(6_364_136_223_846_793_005)
                            .wrapping_add(1_442_695_040_888_963_407);
                        ((state >> 33) as f32 / u32::MAX as f32) * 2.0 - 1.0
                    })
                    .collect();
                Vector::new(data)
            })
            .collect();

        let (trained_params, trained_codebook, trained_codes) =
            quantize_segment_pq(&vectors, 32, 4, None).unwrap();

        let shared = crate::vector::index::pq_codebook::SharedPqCodebook {
            params: trained_params,
            codebook: trained_codebook.clone(),
        };
        let (shared_params, shared_codebook, shared_codes) =
            quantize_segment_pq(&vectors, 32, 4, Some(&shared)).unwrap();

        assert_eq!(shared_params, trained_params);
        assert_eq!(shared_codebook, trained_codebook);
        assert_eq!(shared_codes, trained_codes);
    }
}