memd 1.9.0

Local memory CLI for AI agents with hybrid 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
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
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
use super::onnx_runtime;
use ndarray::{Array2, ArrayViewD, Axis};
use ort::session::builder::GraphOptimizationLevel;
use ort::session::Session;
use ort::value::TensorRef;
use parking_lot::Mutex;
use std::fs;
use std::io;
use std::panic::{catch_unwind, AssertUnwindSafe};
use std::path::{Path, PathBuf};
use std::sync::{Mutex as StdMutex, OnceLock};
use tokenizers::{EncodeInput, PaddingParams, PaddingStrategy, Tokenizer, TruncationParams};

const MODEL_URL: &str =
    "https://huggingface.co/Xenova/ms-marco-MiniLM-L-6-v2/resolve/main/onnx/model.onnx";
const TOKENIZER_URL: &str =
    "https://huggingface.co/Xenova/ms-marco-MiniLM-L-6-v2/resolve/main/tokenizer.json";
const MODEL_FILENAME: &str = "ms-marco-minilm-l6-v2.onnx";
const TOKENIZER_FILENAME: &str = "ms-marco-minilm-l6-v2-tokenizer.json";
const MODEL_MIN_BYTES: u64 = 1_000_000;
const TOKENIZER_MIN_BYTES: u64 = 100_000;
const DEFAULT_MAX_LENGTH: usize = 256;

static SCORER: OnceLock<Result<OnnxCrossEncoderScorer, String>> = OnceLock::new();
static PANIC_HOOK_LOCK: OnceLock<StdMutex<()>> = OnceLock::new();

pub(crate) fn is_available() -> bool {
    catch_onnx_panic(|| get_scorer().is_ok()).unwrap_or(false)
}

pub(crate) fn score_pairs(query: &str, docs: &[String]) -> Result<Vec<f32>, String> {
    catch_onnx_panic(|| {
        let scorer = get_scorer()?;
        scorer.score_pairs(query, docs)
    })?
}

fn get_scorer() -> Result<&'static OnnxCrossEncoderScorer, String> {
    SCORER
        .get_or_init(OnnxCrossEncoderScorer::new)
        .as_ref()
        .map_err(|err| err.clone())
}

fn catch_onnx_panic<T, F>(f: F) -> Result<T, String>
where
    F: FnOnce() -> T,
{
    let lock = PANIC_HOOK_LOCK.get_or_init(|| StdMutex::new(()));
    let _guard = lock
        .lock()
        .map_err(|_| "cross-encoder panic hook mutex poisoned".to_string())?;
    let old_hook = std::panic::take_hook();
    std::panic::set_hook(Box::new(|_| {}));
    let result = catch_unwind(AssertUnwindSafe(f));
    std::panic::set_hook(old_hook);
    result.map_err(|_| "cross-encoder scorer panicked during ONNX initialization".to_string())
}

struct OnnxCrossEncoderScorer {
    session: Mutex<Session>,
    tokenizer: Mutex<Tokenizer>,
    has_token_type_ids: bool,
    max_length: usize,
}

impl OnnxCrossEncoderScorer {
    fn new() -> Result<Self, String> {
        if std::env::var("MEMD_CROSS_ENCODER_DISABLE").ok().as_deref() == Some("1") {
            return Err(
                "cross-encoder scorer disabled via MEMD_CROSS_ENCODER_DISABLE=1".to_string(),
            );
        }

        let cache_root = cache_dir()?;
        let ort_dylib_path = onnx_runtime::ensure_dylib(&cache_root)?;
        let model_path = resolve_model_path(&cache_root)?;
        let tokenizer_path = resolve_tokenizer_path(&cache_root)?;
        let session = Session::builder()
            .map_err(|err| format!("create ONNX session builder: {err}"))?
            .with_optimization_level(GraphOptimizationLevel::Level3)
            .map_err(|err| format!("set ONNX optimization level: {err}"))?
            .with_intra_threads(2)
            .map_err(|err| format!("set ONNX intra threads: {err}"))?
            .commit_from_file(&model_path)
            .map_err(|err| format!("load ONNX model '{}': {err}", model_path.display()))?;
        let has_token_type_ids = session
            .inputs()
            .iter()
            .any(|input| input.name() == "token_type_ids");

        let mut tokenizer = Tokenizer::from_file(&tokenizer_path)
            .map_err(|err| format!("load tokenizer '{}': {err}", tokenizer_path.display()))?;
        tokenizer
            .with_truncation(Some(TruncationParams {
                max_length: DEFAULT_MAX_LENGTH,
                ..Default::default()
            }))
            .map_err(|err| format!("configure tokenizer truncation: {err}"))?;
        tokenizer.with_padding(Some(PaddingParams {
            strategy: PaddingStrategy::BatchLongest,
            ..Default::default()
        }));

        tracing::info!(
            model = %model_path.display(),
            tokenizer = %tokenizer_path.display(),
            onnxruntime = %ort_dylib_path.display(),
            has_token_type_ids,
            "initialized ONNX cross-encoder scorer"
        );

        Ok(Self {
            session: Mutex::new(session),
            tokenizer: Mutex::new(tokenizer),
            has_token_type_ids,
            max_length: DEFAULT_MAX_LENGTH,
        })
    }

    fn score_pairs(&self, query: &str, docs: &[String]) -> Result<Vec<f32>, String> {
        if query.trim().is_empty() || docs.is_empty() {
            return Ok(vec![0.0; docs.len()]);
        }

        let inputs: Vec<EncodeInput> = docs
            .iter()
            .map(|doc| EncodeInput::Dual(query.into(), doc.as_str().into()))
            .collect();
        let encodings = self
            .tokenizer
            .lock()
            .encode_batch(inputs, true)
            .map_err(|err| format!("tokenize query/document pairs: {err}"))?;
        let (input_ids, attention_mask, token_type_ids) = self.to_tensors(&encodings)?;
        self.run_model(input_ids, attention_mask, token_type_ids)
    }

    fn to_tensors(
        &self,
        encodings: &[tokenizers::Encoding],
    ) -> Result<(Array2<i64>, Array2<i64>, Array2<i64>), String> {
        let max_len = encodings
            .iter()
            .map(|enc| enc.get_ids().len())
            .max()
            .unwrap_or(0)
            .min(self.max_length);
        if max_len == 0 {
            return Err("tokenizer produced zero-length input".to_string());
        }

        let batch = encodings.len();
        let mut input_ids = Array2::<i64>::zeros((batch, max_len));
        let mut attention_mask = Array2::<i64>::zeros((batch, max_len));
        let mut token_type_ids = Array2::<i64>::zeros((batch, max_len));
        for (i, enc) in encodings.iter().enumerate() {
            let ids = enc.get_ids();
            let mask = enc.get_attention_mask();
            let types = enc.get_type_ids();
            for j in 0..max_len {
                input_ids[[i, j]] = ids.get(j).copied().unwrap_or(0) as i64;
                attention_mask[[i, j]] = mask.get(j).copied().unwrap_or(0) as i64;
                token_type_ids[[i, j]] = types.get(j).copied().unwrap_or(0) as i64;
            }
        }

        Ok((input_ids, attention_mask, token_type_ids))
    }

    fn run_model(
        &self,
        input_ids: Array2<i64>,
        attention_mask: Array2<i64>,
        token_type_ids: Array2<i64>,
    ) -> Result<Vec<f32>, String> {
        let input_ids_tensor = TensorRef::from_array_view(input_ids.view())
            .map_err(|err| format!("create input_ids tensor: {err}"))?;
        let attention_mask_tensor = TensorRef::from_array_view(attention_mask.view())
            .map_err(|err| format!("create attention_mask tensor: {err}"))?;

        let mut session = self.session.lock();
        let outputs = if self.has_token_type_ids {
            let token_type_ids_tensor = TensorRef::from_array_view(token_type_ids.view())
                .map_err(|err| format!("create token_type_ids tensor: {err}"))?;
            session.run(ort::inputs![
                "input_ids" => input_ids_tensor,
                "attention_mask" => attention_mask_tensor,
                "token_type_ids" => token_type_ids_tensor,
            ])
        } else {
            session.run(ort::inputs![
                "input_ids" => input_ids_tensor,
                "attention_mask" => attention_mask_tensor,
            ])
        }
        .map_err(|err| format!("cross-encoder inference failed: {err}"))?;

        let logits = if let Some(output) = outputs.get("logits") {
            output
                .try_extract_array::<f32>()
                .map_err(|err| format!("extract logits output: {err}"))?
                .into_owned()
        } else {
            outputs[0]
                .try_extract_array::<f32>()
                .map_err(|err| format!("extract default output tensor: {err}"))?
                .into_owned()
        };
        Ok(logits_to_scores(logits.view()))
    }
}

/// Convert a raw logits tensor from the cross-encoder into a `[0, 1]`
/// relevance score per input pair.
///
/// Supported shapes (last-dim `C` is the number of classes):
/// - `C == 1` — regression head: sigmoid(logit).
/// - `C == 2` — classification head with `[not_relevant, relevant]`:
///   softmax and take the relevance probability. The previous
///   implementation read column 0 (the *not-relevant* score), which
///   silently inverted ranking for any 2-class export.
/// - other — unsupported; warn once per call and fall back to
///   sigmoid(column 0) so the pipeline stays available.
fn logits_to_scores(logits: ArrayViewD<'_, f32>) -> Vec<f32> {
    if logits.ndim() == 1 {
        return logits.iter().copied().map(sigmoid).collect();
    }
    if logits.ndim() >= 2 {
        let last_dim = *logits.shape().last().unwrap_or(&0);
        let rows = logits.shape()[0];
        match last_dim {
            0 => vec![0.0; rows],
            1 => (0..rows)
                .map(|row| {
                    let v = logits
                        .index_axis(Axis(0), row)
                        .iter()
                        .copied()
                        .next()
                        .unwrap_or(0.0);
                    sigmoid(v)
                })
                .collect(),
            2 => (0..rows)
                .map(|row| {
                    let view = logits.index_axis(Axis(0), row);
                    let values: Vec<f32> = view.iter().copied().collect();
                    // Binary classification heads conventionally emit
                    // [not_relevant, relevant]. Softmax the pair and
                    // return the relevance probability.
                    softmax_relevant(values[0], values[1])
                })
                .collect(),
            other => {
                tracing::warn!(
                    last_dim = other,
                    ndim = logits.ndim(),
                    "unrecognised cross-encoder output shape; falling back to sigmoid(column 0)"
                );
                (0..rows)
                    .map(|row| {
                        let v = logits
                            .index_axis(Axis(0), row)
                            .iter()
                            .copied()
                            .next()
                            .unwrap_or(0.0);
                        sigmoid(v)
                    })
                    .collect()
            }
        }
    } else {
        vec![0.0]
    }
}

fn sigmoid(value: f32) -> f32 {
    1.0 / (1.0 + (-value).exp())
}

/// Softmax for a 2-class logits pair, returning the probability of the
/// second class (conventionally "relevant").
fn softmax_relevant(not_relevant: f32, relevant: f32) -> f32 {
    let max = not_relevant.max(relevant);
    let e_not = (not_relevant - max).exp();
    let e_rel = (relevant - max).exp();
    e_rel / (e_not + e_rel)
}

fn resolve_model_path(cache_root: &Path) -> Result<PathBuf, String> {
    if let Ok(path) = std::env::var("MEMD_CROSS_ENCODER_MODEL_PATH") {
        return verify_existing_file(Path::new(&path), MODEL_MIN_BYTES, "model");
    }
    let cache_path = cache_root.join(MODEL_FILENAME);
    ensure_downloaded(&cache_path, MODEL_URL, MODEL_MIN_BYTES, "model")
}

fn resolve_tokenizer_path(cache_root: &Path) -> Result<PathBuf, String> {
    if let Ok(path) = std::env::var("MEMD_CROSS_ENCODER_TOKENIZER_PATH") {
        return verify_existing_file(Path::new(&path), TOKENIZER_MIN_BYTES, "tokenizer");
    }
    let cache_path = cache_root.join(TOKENIZER_FILENAME);
    ensure_downloaded(&cache_path, TOKENIZER_URL, TOKENIZER_MIN_BYTES, "tokenizer")
}

fn cache_dir() -> Result<PathBuf, String> {
    if let Ok(path) = std::env::var("MEMD_CROSS_ENCODER_CACHE_DIR") {
        return Ok(PathBuf::from(path));
    }
    let base = dirs::cache_dir().ok_or_else(|| "resolve cache directory".to_string())?;
    Ok(base.join("memd").join("cross-encoder"))
}

fn ensure_downloaded(
    path: &Path,
    url: &str,
    min_bytes: u64,
    label: &str,
) -> Result<PathBuf, String> {
    if let Ok(existing) = verify_existing_file(path, min_bytes, label) {
        return Ok(existing);
    }
    if let Some(parent) = path.parent() {
        fs::create_dir_all(parent)
            .map_err(|err| format!("create {} cache dir '{}': {err}", label, parent.display()))?;
    }
    tracing::info!(label, url, destination = %path.display(), "downloading cross-encoder asset");
    let mut response = ureq::get(url)
        .call()
        .map_err(|err| format!("download {} from '{}': {err}", label, url))?
        .into_reader();
    let mut file =
        fs::File::create(path).map_err(|err| format!("create '{}': {err}", path.display()))?;
    io::copy(&mut response, &mut file)
        .map_err(|err| format!("write '{}': {err}", path.display()))?;
    verify_existing_file(path, min_bytes, label)
}

fn verify_existing_file(path: &Path, min_bytes: u64, label: &str) -> Result<PathBuf, String> {
    let metadata = fs::metadata(path).map_err(|err| format!("stat '{}': {err}", path.display()))?;
    if metadata.len() < min_bytes {
        return Err(format!(
            "{} file '{}' too small ({} bytes, expected at least {})",
            label,
            path.display(),
            metadata.len(),
            min_bytes
        ));
    }
    Ok(path.to_path_buf())
}

#[cfg(test)]
mod tests {
    use super::{logits_to_scores, score_pairs, softmax_relevant};
    use ndarray::{Array1, Array2};

    #[test]
    #[ignore = "requires real ONNX runtime/model initialization and may download assets"]
    fn smoke_real_onnx_scores_relevant_pair_higher() {
        let docs = vec![
            "const validateEmailFormat = (str) => { return /^\\w+([.-]?\\w+)*@\\w+([.-]?\\w+)*(\\.\\w{2,3})+$/.test(str); };".to_string(),
            "function parseJson(str) { try { return JSON.parse(str); } catch (e) { return null; } }".to_string(),
        ];

        let scores = score_pairs("validate email address format", &docs)
            .expect("real ONNX smoke test should initialize and score");

        assert_eq!(scores.len(), 2);
        assert!(scores.iter().all(|score| score.is_finite()));
        assert!(
            scores[0] > scores[1],
            "expected relevant doc to outrank distractor, got scores {scores:?}"
        );
    }

    /// Regression test: for a 2-class head that emits
    /// `[not_relevant, relevant]`, the previous implementation read the
    /// first column (the *not-relevant* logit) and sigmoided it — which
    /// silently inverted ranking. The fix softmaxes the pair and returns
    /// the relevance probability.
    #[test]
    fn logits_to_scores_handles_two_class_head() {
        // Row 0: strongly "relevant". Row 1: strongly "not relevant".
        let logits = Array2::from_shape_vec((2, 2), vec![-2.0, 3.0, 3.0, -2.0]).unwrap();
        let scores = logits_to_scores(logits.view().into_dyn());

        assert_eq!(scores.len(), 2);
        assert!(
            scores[0] > 0.5,
            "relevant row should produce score > 0.5, got {}",
            scores[0]
        );
        assert!(
            scores[1] < 0.5,
            "not-relevant row should produce score < 0.5, got {}",
            scores[1]
        );
        assert!(
            scores[0] > scores[1],
            "the relevant row must outrank the not-relevant row"
        );
    }

    #[test]
    fn logits_to_scores_handles_regression_head_1d() {
        let logits = Array1::from_vec(vec![2.0, -2.0]);
        let scores = logits_to_scores(logits.view().into_dyn());

        assert_eq!(scores.len(), 2);
        assert!(scores[0] > 0.5);
        assert!(scores[1] < 0.5);
    }

    #[test]
    fn logits_to_scores_handles_regression_head_2d_single_column() {
        let logits = Array2::from_shape_vec((3, 1), vec![3.0, 0.0, -3.0]).unwrap();
        let scores = logits_to_scores(logits.view().into_dyn());

        assert_eq!(scores.len(), 3);
        assert!(scores[0] > scores[1] && scores[1] > scores[2]);
        assert!((scores[1] - 0.5).abs() < 1e-6);
    }

    #[test]
    fn softmax_relevant_is_monotonic_in_the_margin() {
        let low = softmax_relevant(1.0, -1.0);
        let mid = softmax_relevant(0.0, 0.0);
        let high = softmax_relevant(-1.0, 1.0);

        assert!(low < mid && mid < high);
        assert!((mid - 0.5).abs() < 1e-6);
    }
}