cxpak 1.0.1

Spends CPU cycles so you don't spend tokens. The LLM gets a briefing packet instead of a flashlight in a dark room.
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
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252

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

/// Local inference provider using candle + all-MiniLM-L6-v2 (SafeTensors).
///
/// Model files are cached in `~/.cxpak/models/all-MiniLM-L6-v2/`.
pub struct LocalEmbeddingProvider {
    model: candle_transformers::models::bert::BertModel,
    tokenizer: tokenizers::Tokenizer,
    device: candle_core::Device,
    dims: usize,
}

const HF_BASE: &str = "https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2/resolve/main";
const MODEL_DIMS: usize = 384;
const CACHE_SUBDIR: &str = "all-MiniLM-L6-v2";

impl LocalEmbeddingProvider {
    /// Build the provider, downloading the model weights if necessary.
    pub fn new() -> Result<Self, String> {
        let cache_dir = model_cache_dir()?;
        ensure_model_files(&cache_dir)?;

        let device = candle_core::Device::Cpu;

        let tokenizer = tokenizers::Tokenizer::from_file(cache_dir.join("tokenizer.json"))
            .map_err(|e| format!("tokenizer load error: {e}"))?;

        // Load SafeTensors weights via buffered loader (no unsafe).
        let safetensors_bytes = std::fs::read(cache_dir.join("model.safetensors"))
            .map_err(|e| format!("model read error: {e}"))?;
        let vb = candle_nn::VarBuilder::from_buffered_safetensors(
            safetensors_bytes,
            candle_core::DType::F32,
            &device,
        )
        .map_err(|e| format!("varbuilder error: {e}"))?;

        let config_file = std::fs::File::open(cache_dir.join("config.json"))
            .map_err(|e| format!("config open error: {e}"))?;
        let config: candle_transformers::models::bert::Config =
            serde_json::from_reader(config_file).map_err(|e| format!("config parse error: {e}"))?;

        let model = candle_transformers::models::bert::BertModel::load(vb, &config)
            .map_err(|e| format!("model load error: {e}"))?;

        Ok(Self {
            model,
            tokenizer,
            device,
            dims: MODEL_DIMS,
        })
    }

    /// Embed a single text. Returns a normalized 384-dim vector.
    pub fn embed(&self, text: &str) -> Result<Vec<f32>, String> {
        let batch = self.embed_batch(&[text])?;
        batch
            .into_iter()
            .next()
            .ok_or_else(|| "empty batch result".to_string())
    }

    /// Embed a batch of texts. Returns one normalized vector per input.
    pub fn embed_batch(&self, texts: &[&str]) -> Result<Vec<Vec<f32>>, String> {
        use candle_core::Tensor;

        let encodings = self
            .tokenizer
            .encode_batch(texts.to_vec(), true)
            .map_err(|e| format!("tokenize error: {e}"))?;

        // Pad all sequences to the same length.
        let max_len = encodings.iter().map(|e| e.len()).max().unwrap_or(0);
        if max_len == 0 {
            return Ok(texts.iter().map(|_| vec![0.0f32; self.dims]).collect());
        }

        let n = texts.len();

        let mut input_ids_data = Vec::with_capacity(n * max_len);
        let mut attention_mask_data = Vec::with_capacity(n * max_len);
        let mut token_type_ids_data = Vec::with_capacity(n * max_len);

        for enc in &encodings {
            let ids = enc.get_ids();
            let mask = enc.get_attention_mask();
            let ttids = enc.get_type_ids();

            input_ids_data.extend(ids.iter().map(|&x| x as i64));
            attention_mask_data.extend(mask.iter().map(|&x| x as i64));
            token_type_ids_data.extend(ttids.iter().map(|&x| x as i64));

            // Pad to max_len.
            let pad = max_len - ids.len();
            for _ in 0..pad {
                input_ids_data.push(0);
                attention_mask_data.push(0);
                token_type_ids_data.push(0);
            }
        }

        let input_ids = Tensor::from_vec(input_ids_data, (n, max_len), &self.device)
            .map_err(|e| format!("tensor error: {e}"))?;
        let attention_mask = Tensor::from_vec(attention_mask_data, (n, max_len), &self.device)
            .map_err(|e| format!("tensor error: {e}"))?;
        let token_type_ids = Tensor::from_vec(token_type_ids_data, (n, max_len), &self.device)
            .map_err(|e| format!("tensor error: {e}"))?;

        let output = self
            .model
            .forward(&input_ids, &token_type_ids, Some(&attention_mask))
            .map_err(|e| format!("model forward error: {e}"))?;

        // Mean-pool over token dimension (dim 1), excluding padding tokens.
        let mask_f32 = attention_mask
            .to_dtype(candle_core::DType::F32)
            .map_err(|e| format!("dtype error: {e}"))?;
        // mask_f32: (n, seq_len), output: (n, seq_len, hidden)
        let mask_expanded = mask_f32
            .unsqueeze(2)
            .map_err(|e| format!("unsqueeze error: {e}"))?;
        let masked = (output * mask_expanded).map_err(|e| format!("mul error: {e}"))?;
        let summed = masked.sum(1).map_err(|e| format!("sum error: {e}"))?;
        let counts = mask_f32
            .sum(1)
            .map_err(|e| format!("sum mask error: {e}"))?
            .unsqueeze(1)
            .map_err(|e| format!("unsqueeze error: {e}"))?;
        let mean = (summed / counts).map_err(|e| format!("div error: {e}"))?;

        // L2-normalize each row.
        let mean_data: Vec<f32> = mean
            .flatten_all()
            .map_err(|e| format!("flatten error: {e}"))?
            .to_vec1()
            .map_err(|e| format!("to_vec1 error: {e}"))?;

        let mut result = Vec::with_capacity(n);
        for i in 0..n {
            let row = &mean_data[i * self.dims..(i + 1) * self.dims];
            result.push(l2_normalize(row));
        }

        Ok(result)
    }

    /// Dimensionality of the produced embeddings (384).
    pub fn dimensions(&self) -> usize {
        self.dims
    }
}

// ---------------------------------------------------------------------------
// Internal helpers
// ---------------------------------------------------------------------------

fn model_cache_dir() -> Result<PathBuf, String> {
    #[allow(deprecated)]
    let home = std::env::home_dir().ok_or_else(|| "cannot find home directory".to_string())?;
    let dir = home.join(".cxpak").join("models").join(CACHE_SUBDIR);
    std::fs::create_dir_all(&dir).map_err(|e| format!("create dirs error: {e}"))?;
    Ok(dir)
}

fn ensure_model_files(dir: &Path) -> Result<(), String> {
    let files = ["model.safetensors", "config.json", "tokenizer.json"];

    for name in files {
        let dest = dir.join(name);
        if dest.exists() {
            continue;
        }
        let url = format!("{HF_BASE}/{name}");
        download_file(&url, &dest)?;
    }
    Ok(())
}

fn download_file(url: &str, dest: &Path) -> Result<(), String> {
    let response =
        reqwest::blocking::get(url).map_err(|e| format!("download error for {url}: {e}"))?;

    if !response.status().is_success() {
        return Err(format!("HTTP {} downloading {url}", response.status()));
    }

    let bytes = response
        .bytes()
        .map_err(|e| format!("read bytes error: {e}"))?;

    std::fs::write(dest, &bytes).map_err(|e| format!("write error: {e}"))
}

fn l2_normalize(v: &[f32]) -> Vec<f32> {
    let norm: f32 = v.iter().map(|x| x * x).sum::<f32>().sqrt();
    if norm == 0.0 {
        return v.to_vec();
    }
    v.iter().map(|x| x / norm).collect()
}

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

    #[test]
    #[ignore = "requires network to download model"]
    fn test_local_provider_single_embed() {
        let provider = LocalEmbeddingProvider::new().expect("should construct");
        let vec = provider
            .embed("fn hello() { println!(\"hello\"); }")
            .unwrap();
        assert_eq!(vec.len(), 384);
        let norm: f32 = vec.iter().map(|x| x * x).sum::<f32>().sqrt();
        assert!((norm - 1.0).abs() < 1e-3, "norm={norm}");
    }

    #[test]
    #[ignore = "requires network to download model"]
    fn test_local_provider_batch_embed() {
        let provider = LocalEmbeddingProvider::new().expect("should construct");
        let texts = vec!["fn foo() {}", "struct Bar {}"];
        let vecs = provider.embed_batch(&texts).unwrap();
        assert_eq!(vecs.len(), 2);
        assert_eq!(vecs[0].len(), 384);
        assert_eq!(vecs[1].len(), 384);
    }

    #[test]
    #[ignore = "requires network to download model"]
    fn test_local_provider_dimensions() {
        let provider = LocalEmbeddingProvider::new().expect("should construct");
        assert_eq!(provider.dimensions(), 384);
    }

    #[test]
    fn test_l2_normalize_unit_vector() {
        let v = vec![3.0f32, 4.0, 0.0];
        let n = l2_normalize(&v);
        let norm: f32 = n.iter().map(|x| x * x).sum::<f32>().sqrt();
        assert!((norm - 1.0).abs() < 1e-6, "norm={norm}");
        assert!((n[0] - 0.6).abs() < 1e-6);
        assert!((n[1] - 0.8).abs() < 1e-6);
    }

    #[test]
    fn test_l2_normalize_zero_vector() {
        let v = vec![0.0f32, 0.0, 0.0];
        let n = l2_normalize(&v);
        assert_eq!(n, vec![0.0, 0.0, 0.0]);
    }
}