Skip to main content

ripvec_core/encoder/ripvec/
dense.rs

1//! Static encoder: in-process `StaticEmbedModel` reimplementation.
2//!
3//! Port of `~/src/semble/src/semble/index/dense.py`. Wraps
4//! [`StaticEmbedModel`] loaded with `minishlab/potion-code-16M`
5//! (256-dim, L2-normalized). Implements [`VectorEncoder`] for the
6//! `--model ripvec` path. CPU-only; no batching ring buffer.
7//!
8//! ## Why not `model2vec-rs`?
9//!
10//! The previous wave used the upstream `model2vec-rs` crate. Two real
11//! problems pushed us to reimplement (see
12//! `crates/ripvec-core/src/encoder/semble/static_model.rs` for the
13//! full design rationale):
14//!
15//! 1. `model2vec_rs::StaticModel::encode_with_args` runs `pool_ids`
16//!    in a serial inner loop while `tokenizers::encode_batch_fast`
17//!    spawns its own rayon pool. Wrapping that path in our outer
18//!    `par_chunks` produced 60% `__psynch_cvwait` in the linux-corpus
19//!    profile — nested rayon scopes parking on each other. The
20//!    reimplementation does ONE big tokenize plus a `par_iter` over
21//!    `pool_ids` — no nested rayon, no parking.
22//! 2. `model2vec-rs 0.2` pinned `ndarray 0.15`; ripvec-core uses
23//!    `ndarray 0.17`. The two `Array2<f32>` types were not
24//!    interchangeable, forcing a `Vec<Vec<f32>>` shim. Owning the
25//!    load path eliminates the mismatch.
26
27use std::path::{Path, PathBuf};
28use std::sync::Mutex;
29
30use crossbeam_channel::bounded;
31use hf_hub::api::sync::Api;
32use rayon::prelude::*;
33
34use crate::chunk::CodeChunk;
35use crate::embed::SearchConfig;
36use crate::encoder::VectorEncoder;
37use crate::encoder::ripvec::chunking::{DEFAULT_DESIRED_CHUNK_CHARS, chunk_source};
38use crate::encoder::ripvec::static_model::StaticEmbedModel;
39use crate::languages::config_for_extension;
40use crate::profile::Profiler;
41use crate::walk::collect_files_with_options;
42
43/// Encode batch size used by the streaming pipeline. Matches
44/// `StaticEmbedModel`'s internal `BATCH_SIZE` so each emitted batch
45/// is exactly one `encode_batch_fast` call's worth of work.
46const PIPELINE_BATCH_SIZE: usize = 1024;
47
48/// Number of full batches allowed in-flight from chunker to encoder.
49/// Provides enough pipeline depth for the encoder to stay busy while
50/// the chunker fills the next batch; small enough that peak memory
51/// stays bounded.
52const PIPELINE_RING_SIZE: usize = 4;
53
54/// Default model repo identifier for the ripvec path. This is the HF
55/// repo string used as `identity()`; the loader reads files from a
56/// local path passed via `--model-repo`.
57pub const DEFAULT_MODEL_REPO: &str = "minishlab/potion-code-16M";
58
59/// Default hidden dimension for [`DEFAULT_MODEL_REPO`].
60pub const DEFAULT_HIDDEN_DIM: usize = 256;
61
62/// Maximum source file size to read, in bytes (mirrors semble's
63/// `_MAX_FILE_BYTES = 1_000_000` from `index/create.py:16`).
64const MAX_FILE_BYTES: u64 = 1_000_000;
65
66/// CPU-only static encoder.
67///
68/// Owns a loaded [`StaticEmbedModel`] plus identity metadata. The
69/// embedder is constructed by `main.rs::load_pipeline` via
70/// [`StaticEncoder::from_pretrained`], passing either a local path
71/// containing the Model2Vec files or (planned) an HF repo ID.
72pub struct StaticEncoder {
73    model: StaticEmbedModel,
74    model_repo: String,
75    hidden_dim: usize,
76}
77
78impl StaticEncoder {
79    /// Encode a query string into a single embedding row.
80    ///
81    /// Used by `RipvecIndex::search` for hybrid/semantic dispatch.
82    #[must_use]
83    pub fn encode_query(&self, query: &str) -> Vec<f32> {
84        self.model.encode_query(query)
85    }
86
87    /// Load a model by HuggingFace repo ID or local path.
88    ///
89    /// Two acceptance shapes:
90    ///
91    /// 1. **Local path** — if `model_repo` names an existing directory,
92    ///    load directly from it. Used by the parity test fixture path
93    ///    (`/tmp/potion-base-32M`) and any user pre-staging files.
94    /// 2. **HuggingFace repo ID** — otherwise treat as `org/repo`,
95    ///    download `config.json` / `tokenizer.json` / `model.safetensors`
96    ///    via `hf-hub` into `~/.cache/huggingface/hub/`, and load from
97    ///    there. Matches `load_classic_cpu` / `load_modernbert_cpu`'s
98    ///    behaviour so the user-facing API is consistent: bare `--model
99    ///    ripvec` with no `--model-repo` flag works.
100    ///
101    /// # Errors
102    ///
103    /// Propagates the underlying I/O, download, or parse error if the
104    /// files cannot be obtained or the safetensors layout is
105    /// unrecognized.
106    pub fn from_pretrained(model_repo: &str) -> crate::Result<Self> {
107        let resolved = Self::resolve_model_dir(model_repo)?;
108        let model = StaticEmbedModel::from_path(&resolved, Some(true))
109            .map_err(|e| crate::Error::Other(anyhow::anyhow!("static model load failed: {e}")))?;
110        let hidden_dim = model.hidden_dim();
111        Ok(Self {
112            model,
113            model_repo: model_repo.to_string(),
114            hidden_dim,
115        })
116    }
117
118    /// Resolve `model_repo` to a directory containing the model files.
119    ///
120    /// If `model_repo` is an existing local directory, returns it as-is.
121    /// Otherwise downloads via `hf-hub` and returns the cache directory.
122    fn resolve_model_dir(model_repo: &str) -> crate::Result<PathBuf> {
123        let local = Path::new(model_repo);
124        if local.is_dir() {
125            return Ok(local.to_path_buf());
126        }
127
128        // HuggingFace repo path. Download the three required files and
129        // return the directory `hf-hub` cached them into. All files
130        // land in the same snapshot directory.
131        let api = Api::new().map_err(|e| crate::Error::Download(e.to_string()))?;
132        let repo = api.model(model_repo.to_string());
133        let _ = repo
134            .get("config.json")
135            .map_err(|e| crate::Error::Download(e.to_string()))?;
136        let _ = repo
137            .get("tokenizer.json")
138            .map_err(|e| crate::Error::Download(e.to_string()))?;
139        let weights_path = repo
140            .get("model.safetensors")
141            .map_err(|e| crate::Error::Download(e.to_string()))?;
142        // hf-hub returns the file path; the snapshot directory is its parent.
143        weights_path
144            .parent()
145            .map(std::path::Path::to_path_buf)
146            .ok_or_else(|| {
147                crate::Error::Other(anyhow::anyhow!(
148                    "hf-hub returned root path for {model_repo}; cannot resolve snapshot dir"
149                ))
150            })
151    }
152}
153
154impl VectorEncoder for StaticEncoder {
155    /// Three-stage bounded-queue pipeline:
156    ///
157    /// 1. **Chunk producer** — rayon `par_iter` over the file list. Each
158    ///    file is read, parsed by tree-sitter (or line-merged on
159    ///    fallback), and emitted as `(CodeChunk, String)` pairs into a
160    ///    bounded channel of capacity `PIPELINE_BATCH_SIZE * 8`.
161    /// 2. **Batch accumulator** — a single scoped thread drains the
162    ///    chunk channel, packs `PIPELINE_BATCH_SIZE` pairs per batch,
163    ///    and forwards into a bounded channel of capacity
164    ///    `PIPELINE_RING_SIZE`.
165    /// 3. **Encode worker** — a single scoped thread receives batches
166    ///    and calls `StaticEmbedModel::encode_batch`, whose internal
167    ///    `par_iter` lights up rayon for the pool_ids kernel.
168    ///
169    /// Why this shape:
170    ///
171    /// - The previous "chunk all, then embed all" implementation held
172    ///   the entire `Vec<String>` of chunk contents in memory between
173    ///   phases. On the linux corpus that was ~400 MB peak. The
174    ///   bounded queues cap in-flight memory at
175    ///   `PIPELINE_BATCH_SIZE * 8 + PIPELINE_RING_SIZE * PIPELINE_BATCH_SIZE`
176    ///   chunks regardless of corpus size — under 15 MB.
177    /// - The chunk phase (13s on linux) is hidden inside the embed
178    ///   phase (70s) instead of serializing before it. Pre-pipeline
179    ///   profile showed user-time at 394s on 82s wall = 4.8x
180    ///   parallelism on 12 cores; pipeline lets idle cores chew on
181    ///   chunking while embed runs.
182    /// - Mirrors `embed::embed_all_streaming`'s shape so the two
183    ///   pipelines (BERT + semble) share architectural conventions.
184    fn embed_root(
185        &self,
186        root: &Path,
187        cfg: &SearchConfig,
188        profiler: &Profiler,
189    ) -> crate::Result<(Vec<CodeChunk>, Vec<Vec<f32>>)> {
190        // Phase 1: walk (still serial-to-pipeline because we need the
191        // full file list to par_iter over; the walk itself is rayon).
192        let walk_options = cfg.walk_options();
193        let file_paths = {
194            let _guard = profiler.phase("walk");
195            collect_files_with_options(root, &walk_options)
196        };
197        if file_paths.is_empty() {
198            return Ok((Vec::new(), Vec::new()));
199        }
200
201        // Bounded channels. See module constants for the rationale on
202        // PIPELINE_BATCH_SIZE and PIPELINE_RING_SIZE.
203        let (chunk_tx, chunk_rx) = bounded::<(CodeChunk, String)>(PIPELINE_BATCH_SIZE * 8);
204        let (batch_tx, batch_rx) = bounded::<Vec<(CodeChunk, String)>>(PIPELINE_RING_SIZE);
205
206        // The encoder stage writes ordered output behind a Mutex. Order
207        // across files isn't meaningful (RipvecIndex doesn't rely on
208        // chunk order), only the chunk[i] <-> embedding[i] pairing
209        // matters — which we preserve trivially by pushing in lockstep.
210        let output: Mutex<Vec<(CodeChunk, Vec<f32>)>> = Mutex::new(Vec::new());
211        let model = &self.model;
212
213        // Stage 1 runs on a DEDICATED rayon thread pool. If we used
214        // the global pool, Stage 1's par_iter workers would park on
215        // full `chunk_tx.send()` calls, and Stage 3's
216        // `encode_batch` → `pool_ids` par_iter would have no rayon
217        // workers available (they're all parked). That's a classic
218        // nested-rayon deadlock — observed in profiling as PID stuck
219        // at 0% CPU with 16 parked threads.
220        //
221        // Half the cores for chunking, half remain in the global pool
222        // for the encode worker's pool_ids. The chunk phase (tree-
223        // sitter + I/O bound) doesn't need full parallelism to
224        // pipeline cleanly behind embed.
225        let num_cores = rayon::current_num_threads().max(2);
226        let chunk_threads = (num_cores / 2).max(1);
227        let chunk_pool = rayon::ThreadPoolBuilder::new()
228            .num_threads(chunk_threads)
229            .thread_name(|i| format!("semble-chunk-{i}"))
230            .build()
231            .map_err(|e| crate::Error::Other(anyhow::anyhow!("chunk thread pool build: {e}")))?;
232
233        let _phase_guard = profiler.phase("pipeline");
234        std::thread::scope(|scope| {
235            // Stage 1: chunk producer on the dedicated pool.
236            let chunk_tx_owned = chunk_tx;
237            scope.spawn(move || {
238                chunk_pool.install(|| {
239                    file_paths.par_iter().for_each(|full| {
240                        let (chunks, contents) = chunk_one_file(root, full);
241                        for (chunk, content) in chunks.into_iter().zip(contents) {
242                            if chunk_tx_owned.send((chunk, content)).is_err() {
243                                return;
244                            }
245                        }
246                    });
247                });
248                // chunk_tx_owned drops here, closing the channel.
249            });
250
251            // Stage 2: batch accumulator.
252            let batch_tx_owned = batch_tx;
253            scope.spawn(move || {
254                let mut buf: Vec<(CodeChunk, String)> = Vec::with_capacity(PIPELINE_BATCH_SIZE);
255                for pair in chunk_rx {
256                    buf.push(pair);
257                    if buf.len() >= PIPELINE_BATCH_SIZE {
258                        let batch =
259                            std::mem::replace(&mut buf, Vec::with_capacity(PIPELINE_BATCH_SIZE));
260                        if batch_tx_owned.send(batch).is_err() {
261                            return;
262                        }
263                    }
264                }
265                if !buf.is_empty() {
266                    let _ = batch_tx_owned.send(buf);
267                }
268                // batch_tx_owned drops here, closing the channel.
269            });
270
271            // Stage 3: encode worker.
272            scope.spawn(|| {
273                for batch in batch_rx {
274                    if batch.is_empty() {
275                        continue;
276                    }
277                    let mut chunks = Vec::with_capacity(batch.len());
278                    let mut texts: Vec<String> = Vec::with_capacity(batch.len());
279                    for (chunk, text) in batch {
280                        chunks.push(chunk);
281                        texts.push(text);
282                    }
283                    let text_refs: Vec<&str> = texts.iter().map(String::as_str).collect();
284                    let embeddings = model.encode_batch(&text_refs);
285                    debug_assert_eq!(embeddings.len(), chunks.len());
286                    let mut out = output.lock().expect("output mutex poisoned");
287                    for (chunk, emb) in chunks.into_iter().zip(embeddings) {
288                        out.push((chunk, emb));
289                    }
290                }
291            });
292        });
293
294        let collected = output.into_inner().expect("output mutex poisoned");
295        let mut chunks_out = Vec::with_capacity(collected.len());
296        let mut embs_out = Vec::with_capacity(collected.len());
297        for (chunk, emb) in collected {
298            chunks_out.push(chunk);
299            embs_out.push(emb);
300        }
301        Ok((chunks_out, embs_out))
302    }
303
304    fn hidden_dim(&self) -> usize {
305        self.hidden_dim
306    }
307
308    fn identity(&self) -> &str {
309        &self.model_repo
310    }
311}
312
313/// Chunk one file. Returns `(file_chunks, file_contents)` — empty
314/// when the file is too large, can't be read, or has no chunks.
315fn chunk_one_file(root: &Path, full: &Path) -> (Vec<CodeChunk>, Vec<String>) {
316    match std::fs::metadata(full) {
317        Ok(meta) if meta.len() > MAX_FILE_BYTES => return (Vec::new(), Vec::new()),
318        Err(_) => return (Vec::new(), Vec::new()),
319        _ => {}
320    }
321    let Ok(source) = std::fs::read_to_string(full) else {
322        return (Vec::new(), Vec::new());
323    };
324
325    let ext = full
326        .extension()
327        .and_then(|e| e.to_str())
328        .unwrap_or_default();
329    let lang_cfg = config_for_extension(ext);
330    let language = lang_cfg.as_ref().map(|c| &c.language);
331
332    let rel_path = full
333        .strip_prefix(root)
334        .unwrap_or(full)
335        .display()
336        .to_string();
337
338    let boundaries = chunk_source(&source, language, DEFAULT_DESIRED_CHUNK_CHARS);
339    let mut chunks = Vec::with_capacity(boundaries.len());
340    let mut contents = Vec::with_capacity(boundaries.len());
341    for b in boundaries {
342        let text = b.content(&source).to_string();
343        if text.trim().is_empty() {
344            continue;
345        }
346        contents.push(text.clone());
347        chunks.push(CodeChunk {
348            file_path: rel_path.clone(),
349            name: String::new(),
350            kind: String::new(),
351            start_line: b.start_line,
352            end_line: b.end_line,
353            content: text.clone(),
354            enriched_content: text,
355        });
356    }
357    (chunks, contents)
358}
359
360#[cfg(test)]
361mod tests {
362    use super::*;
363    use crate::encoder::VectorEncoder;
364
365    /// `StaticEncoder` implements `VectorEncoder` + Send + Sync.
366    /// Compile-time check (`test:static-encoder-implements-vector-encoder`).
367    #[test]
368    fn static_encoder_implements_vector_encoder() {
369        fn assert_trait_object<T: VectorEncoder + Send + Sync>() {}
370        assert_trait_object::<StaticEncoder>();
371    }
372
373    /// `from_pretrained` returns the right hidden_dim from a probe encode.
374    /// Ignored by default because it requires a model download (~16 MB).
375    ///
376    /// Corresponds to acceptance `test:static-encoder-hidden-dim-256` and
377    /// `test:static-encoder-loads-potion-code-16m` and
378    /// `test:static-encoder-output-is-l2-normalized`.
379    #[test]
380    #[ignore = "requires local model files at RIPVEC_SEMBLE_MODEL_PATH"]
381    fn static_encoder_loads_potion_code_16m() {
382        let Ok(path) = std::env::var("RIPVEC_SEMBLE_MODEL_PATH") else {
383            eprintln!("RIPVEC_SEMBLE_MODEL_PATH not set; skipping");
384            return;
385        };
386        let enc = StaticEncoder::from_pretrained(&path).expect("model load should succeed");
387        assert_eq!(enc.hidden_dim(), DEFAULT_HIDDEN_DIM);
388        // identity() reflects what the caller passed (typically the
389        // local path under test).
390        assert_eq!(enc.identity(), path);
391
392        // Verify L2-normalized output via the public encode_query path.
393        let row = enc.encode_query("hello world");
394        let norm: f32 = row.iter().map(|x| x * x).sum::<f32>().sqrt();
395        assert!(
396            (norm - 1.0).abs() < 1e-3,
397            "expected L2-normalized output; got norm={norm}"
398        );
399    }
400}