holographic-memory 0.6.4

Local-first search engine and agent memory: fast approximate nearest-neighbor and BM25 full-text search, encrypted stores, provable deletion, built on vector-symbolic architectures.
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

// Copyright 2024-2026 WritersLogic Contributors
// SPDX-License-Identifier: Apache-2.0

//! Random-weight model packs for tests: a 2-layer, 32-wide BERT with a 25-word vocabulary. They
//! exercise the whole pipeline (pack, tokenizer, inference, store integration) without real
//! weights or network access. Not for production use.

use std::path::{Path, PathBuf};

use super::pack::{self, Arch, Meta, Pooling, Quantization, Source};
use super::sha256_hex;

pub const WORDS: &[&str] = &[
    "the", "cat", "sat", "on", "mat", "dog", "ran", "in", "park", "bird", "flew", "over", "tree",
    "fish", "swam", "river", "sun", "moon", "star", "night", "day",
];

pub fn tiny_tokenizer() -> Vec<u8> {
    let vocab: serde_json::Map<String, serde_json::Value> = ["[PAD]", "[UNK]", "[CLS]", "[SEP]"]
        .iter()
        .chain(WORDS)
        .enumerate()
        .map(|(i, t)| ((*t).to_string(), serde_json::json!(i)))
        .collect();
    serde_json::to_vec(&serde_json::json!({
        "version": "1.0",
        "truncation": null,
        "padding": null,
        "added_tokens": [],
        "normalizer": {
            "type": "BertNormalizer",
            "clean_text": true,
            "handle_chinese_chars": true,
            "strip_accents": null,
            "lowercase": true
        },
        "pre_tokenizer": { "type": "BertPreTokenizer" },
        "post_processor": { "type": "BertProcessing", "sep": ["[SEP]", 3], "cls": ["[CLS]", 2] },
        "decoder": null,
        "model": {
            "type": "WordPiece",
            "unk_token": "[UNK]",
            "continuing_subword_prefix": "##",
            "max_input_chars_per_word": 100,
            "vocab": vocab
        }
    }))
    .expect("tokenizer json")
}

pub(crate) struct Lcg(pub(crate) u64);
impl Lcg {
    pub(crate) fn next(&mut self) -> f32 {
        self.0 = self
            .0
            .wrapping_mul(6364136223846793005)
            .wrapping_add(1442695040888963407);
        ((self.0 >> 40) as f32 / (1u64 << 24) as f32) * 2.0 - 1.0
    }
}

pub fn tiny_arch() -> Arch {
    Arch {
        hidden: 32,
        layers: 2,
        heads: 4,
        intermediate: 64,
        vocab: 4 + WORDS.len(),
        max_positions: 64,
        type_vocab: 2,
        layer_norm_eps: 1e-12,
        position_offset: 0,
    }
}

pub fn tiny_pack_bytes(quant: Quantization, pooling: Pooling) -> Vec<u8> {
    tiny_pack_bytes_with_group(quant, pooling, 16)
}

/// Like [`tiny_pack_bytes`] with a chosen int8 group (0 or a non-divisor means one scale per row).
pub fn tiny_pack_bytes_with_group(quant: Quantization, pooling: Pooling, group: usize) -> Vec<u8> {
    let a = tiny_arch();
    let mut rng = Lcg(7);
    let mut src = Vec::new();
    let mut add = |name: String, shape: Vec<usize>, matrix: bool, scale: f32, ln: Option<f32>| {
        let n: usize = shape.iter().product();
        let values = (0..n)
            .map(|_| ln.map_or(rng.next() * scale, |c| c + rng.next() * 0.05))
            .collect();
        src.push(Source {
            name,
            shape,
            values,
            matrix,
        });
    };
    let h = a.hidden;
    add("word_emb".into(), vec![a.vocab, h], true, 0.5, None);
    add("pos_emb".into(), vec![a.max_positions, h], false, 0.2, None);
    add("type_emb".into(), vec![a.type_vocab, h], false, 0.2, None);
    add("emb_ln.w".into(), vec![h], false, 0.0, Some(1.0));
    add("emb_ln.b".into(), vec![h], false, 0.0, Some(0.0));
    for l in 0..a.layers {
        let n = |s: &str| format!("l{l}.{s}");
        add(n("qkv.w"), vec![3 * h, h], true, 0.3, None);
        add(n("qkv.b"), vec![3 * h], false, 0.05, None);
        add(n("ao.w"), vec![h, h], true, 0.3, None);
        add(n("ao.b"), vec![h], false, 0.05, None);
        add(n("aln.w"), vec![h], false, 0.0, Some(1.0));
        add(n("aln.b"), vec![h], false, 0.0, Some(0.0));
        add(n("f1.w"), vec![a.intermediate, h], true, 0.3, None);
        add(n("f1.b"), vec![a.intermediate], false, 0.05, None);
        add(n("f2.w"), vec![h, a.intermediate], true, 0.3, None);
        add(n("f2.b"), vec![h], false, 0.05, None);
        add(n("fln.w"), vec![h], false, 0.0, Some(1.0));
        add(n("fln.b"), vec![h], false, 0.0, Some(0.0));
    }
    let meta = Meta {
        model: "test/tiny".into(),
        revision: "0".into(),
        license: "Apache-2.0".into(),
        pooling,
        normalize: true,
        query_prefix: String::new(),
        passage_prefix: String::new(),
        max_length: 32,
    };
    pack::build(meta, a, quant, group, &tiny_tokenizer(), &src).unwrap()
}

pub fn write_tiny(dir: &Path, quant: Quantization, pooling: Pooling) -> (PathBuf, String) {
    let bytes = tiny_pack_bytes(quant, pooling);
    let path = dir.join(format!("tiny-{}.hmsemb", quant.name()));
    std::fs::write(&path, &bytes).unwrap();
    (path, sha256_hex(&bytes))
}