inference/
models.rs

1//! Model configurations for supported embedding models.
2//!
3//! Supported models:
4//! - **BGE-large** (BAAI/bge-large-en-v1.5): Highest quality, 1024 dimensions (default)
5//! - **MiniLM** (all-MiniLM-L6-v2): Fast, 384 dimensions, good for general use
6//! - **BGE-small** (BAAI/bge-small-en-v1.5): Balanced, 384 dimensions, high quality
7//! - **E5-small** (intfloat/e5-small-v2): Quality-focused, 384 dimensions
8
9use serde::{Deserialize, Serialize};
10
11/// Supported embedding models.
12#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash, Serialize, Deserialize, Default)]
13#[serde(rename_all = "kebab-case")]
14pub enum EmbeddingModel {
15    /// BAAI/bge-large-en-v1.5 - Highest quality, 1024 dimensions (default)
16    /// - Dimensions: 1024
17    /// - Max tokens: 512
18    /// - Speed: Slower than small models, but highest quality
19    #[default]
20    BgeLarge,
21
22    /// all-MiniLM-L6-v2 - Fast and efficient, good for general use
23    /// - Dimensions: 384
24    /// - Max tokens: 256
25    /// - Speed: Fastest
26    MiniLM,
27
28    /// BAAI/bge-small-en-v1.5 - Balanced quality and speed
29    /// - Dimensions: 384
30    /// - Max tokens: 512
31    /// - Speed: Medium
32    BgeSmall,
33
34    /// intfloat/e5-small-v2 - Higher quality embeddings
35    /// - Dimensions: 384
36    /// - Max tokens: 512
37    /// - Speed: Medium
38    E5Small,
39}
40
41impl EmbeddingModel {
42    /// Get the HuggingFace model ID.
43    pub fn model_id(&self) -> &'static str {
44        match self {
45            EmbeddingModel::BgeLarge => "BAAI/bge-large-en-v1.5",
46            EmbeddingModel::MiniLM => "sentence-transformers/all-MiniLM-L6-v2",
47            EmbeddingModel::BgeSmall => "BAAI/bge-small-en-v1.5",
48            EmbeddingModel::E5Small => "intfloat/e5-small-v2",
49        }
50    }
51
52    /// Get the embedding dimension for this model.
53    pub fn dimension(&self) -> usize {
54        match self {
55            EmbeddingModel::BgeLarge => 1024,
56            EmbeddingModel::MiniLM => 384,
57            EmbeddingModel::BgeSmall => 384,
58            EmbeddingModel::E5Small => 384,
59        }
60    }
61
62    /// Get the maximum sequence length (in tokens).
63    pub fn max_seq_length(&self) -> usize {
64        match self {
65            EmbeddingModel::BgeLarge => 512,
66            EmbeddingModel::MiniLM => 256,
67            EmbeddingModel::BgeSmall => 512,
68            EmbeddingModel::E5Small => 512,
69        }
70    }
71
72    /// Get the query prefix for models that require it.
73    /// Some models like E5 require a prefix for queries vs documents.
74    pub fn query_prefix(&self) -> Option<&'static str> {
75        match self {
76            EmbeddingModel::BgeLarge => None,
77            EmbeddingModel::MiniLM => None,
78            EmbeddingModel::BgeSmall => None,
79            EmbeddingModel::E5Small => Some("query: "),
80        }
81    }
82
83    /// Get the document/passage prefix for models that require it.
84    pub fn document_prefix(&self) -> Option<&'static str> {
85        match self {
86            EmbeddingModel::BgeLarge => None,
87            EmbeddingModel::MiniLM => None,
88            EmbeddingModel::BgeSmall => None,
89            EmbeddingModel::E5Small => Some("passage: "),
90        }
91    }
92
93    /// Whether this model uses mean pooling (vs CLS token).
94    pub fn use_mean_pooling(&self) -> bool {
95        match self {
96            EmbeddingModel::BgeLarge => true,
97            EmbeddingModel::MiniLM => true,
98            EmbeddingModel::BgeSmall => true,
99            EmbeddingModel::E5Small => true,
100        }
101    }
102
103    /// Whether embeddings should be normalized.
104    pub fn normalize_embeddings(&self) -> bool {
105        true // All supported models use normalized embeddings
106    }
107
108    /// Get approximate tokens per second on CPU (for estimation).
109    pub fn tokens_per_second_cpu(&self) -> usize {
110        match self {
111            EmbeddingModel::BgeLarge => 1000,
112            EmbeddingModel::MiniLM => 5000,
113            EmbeddingModel::BgeSmall => 3000,
114            EmbeddingModel::E5Small => 3000,
115        }
116    }
117
118    /// Get the HuggingFace repository ID hosting the ONNX INT8 model for this embedding model.
119    ///
120    /// These are Xenova-hosted Optimum ONNX exports — quantized INT8, pre-built, no conversion
121    /// needed. BgeLarge: ~130 MB, MiniLM: 23 MB, BGE-small: 35 MB, E5-small: 35 MB.
122    pub fn onnx_repo_id(&self) -> &'static str {
123        match self {
124            EmbeddingModel::BgeLarge => "Xenova/bge-large-en-v1.5",
125            EmbeddingModel::MiniLM => "Xenova/all-MiniLM-L6-v2",
126            EmbeddingModel::BgeSmall => "Xenova/bge-small-en-v1.5",
127            EmbeddingModel::E5Small => "Xenova/e5-small-v2",
128        }
129    }
130
131    /// Get the ONNX model filename for CPU inference (INT8 quantized).
132    pub fn onnx_filename(&self) -> &'static str {
133        "onnx/model_quantized.onnx"
134    }
135
136    /// Get the ONNX model filename for GPU (CUDA EP) inference.
137    ///
138    /// Returns the FP32 model (`onnx/model.onnx`) instead of INT8. The INT8 quantized
139    /// model has 336 Memcpy CPU↔GPU round-trips caused by ORT falling back to CPU EP
140    /// for every unsupported INT8 op — making CUDA 24× slower than pure CPU inference.
141    /// The FP32 model contains no unsupported ops and runs entirely on-device.
142    pub fn onnx_filename_gpu(&self) -> &'static str {
143        "onnx/model.onnx"
144    }
145
146    /// List all available models.
147    pub fn all() -> &'static [EmbeddingModel] {
148        &[
149            EmbeddingModel::BgeLarge,
150            EmbeddingModel::MiniLM,
151            EmbeddingModel::BgeSmall,
152            EmbeddingModel::E5Small,
153        ]
154    }
155
156    /// Parse model from string (case-insensitive).
157    pub fn parse(s: &str) -> Option<Self> {
158        match s.to_lowercase().as_str() {
159            "bge-large" | "bge-large-en" | "bge-large-en-v1.5" => Some(EmbeddingModel::BgeLarge),
160            "minilm" | "all-minilm-l6-v2" | "mini-lm" => Some(EmbeddingModel::MiniLM),
161            "bge-small" | "bge" | "bge-small-en" => Some(EmbeddingModel::BgeSmall),
162            "e5-small" | "e5" | "e5-small-v2" => Some(EmbeddingModel::E5Small),
163            _ => None,
164        }
165    }
166}
167
168impl std::fmt::Display for EmbeddingModel {
169    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
170        match self {
171            EmbeddingModel::BgeLarge => write!(f, "bge-large-en-v1.5"),
172            EmbeddingModel::MiniLM => write!(f, "all-MiniLM-L6-v2"),
173            EmbeddingModel::BgeSmall => write!(f, "bge-small-en-v1.5"),
174            EmbeddingModel::E5Small => write!(f, "e5-small-v2"),
175        }
176    }
177}
178
179/// Configuration for model loading and inference.
180#[derive(Debug, Clone, Serialize, Deserialize)]
181pub struct ModelConfig {
182    /// The embedding model to use.
183    pub model: EmbeddingModel,
184
185    /// Custom cache directory for model files.
186    /// If None, uses HuggingFace default cache.
187    pub cache_dir: Option<String>,
188
189    /// Maximum batch size for inference.
190    pub max_batch_size: usize,
191
192    /// Whether to use GPU acceleration if available.
193    pub use_gpu: bool,
194
195    /// Number of threads for CPU inference.
196    pub num_threads: Option<usize>,
197
198    /// Number of parallel ONNX sessions in the session pool.
199    ///
200    /// Each session holds its own ORT context. Pool members serve batches
201    /// concurrently via `spawn_blocking`, eliminating Mutex head-of-line
202    /// blocking when multiple callers embed text simultaneously.
203    /// Defaults to 4; override with `DAKERA_ONNX_POOL_SIZE` env var at startup.
204    pub session_pool_size: usize,
205}
206
207impl Default for ModelConfig {
208    fn default() -> Self {
209        // DAK-5746: pool=4 restored. PR#488 regressed LME ingest: pool=1 serializes all
210        // ONNX calls onto session[0]. With 4 concurrent HTTP requests × 7 sub-batches each,
211        // pool=1 produces ~28 serial ONNX calls vs pool=4's 7 parallel chains — ~4× throughput
212        // regression measured at 2761ms/50-text batch on prod. OOM root causes (unbounded HNSW,
213        // RocksDB cache) fixed by PR#488 other changes; pool=4 × BGE-Large INT8 ≈ 1.6GB fits
214        // comfortably on the 8GB server. pool_size 4→1 downgrade was the wrong OOM fix.
215        let pool_size = std::env::var("DAKERA_ONNX_POOL_SIZE")
216            .ok()
217            .and_then(|v| v.parse::<usize>().ok())
218            .filter(|&n| n >= 1)
219            .unwrap_or(4);
220        // DAK-5716: no length-sorting (PR#476 proved sorted batching regresses INT8
221        // quantization quality). DAK-5953: default raised 8→32 — amortises per-call ONNX
222        // overhead 4× with no quality impact (size-only change, not order). Bench sets
223        // DAKERA_ONNX_BATCH_SIZE=128; 32 is a safe default for CPU-only deployments.
224        let max_batch_size = std::env::var("DAKERA_ONNX_BATCH_SIZE")
225            .ok()
226            .and_then(|v| v.parse::<usize>().ok())
227            .filter(|&n| n >= 1)
228            .unwrap_or(32);
229        Self {
230            model: EmbeddingModel::default(),
231            cache_dir: None,
232            max_batch_size,
233            use_gpu: false,
234            num_threads: None,
235            session_pool_size: pool_size,
236        }
237    }
238}
239
240impl ModelConfig {
241    /// Create a new config with the specified model.
242    pub fn new(model: EmbeddingModel) -> Self {
243        Self {
244            model,
245            ..Default::default()
246        }
247    }
248
249    /// Set the cache directory.
250    pub fn with_cache_dir(mut self, dir: impl Into<String>) -> Self {
251        self.cache_dir = Some(dir.into());
252        self
253    }
254
255    /// Set the maximum batch size.
256    pub fn with_max_batch_size(mut self, size: usize) -> Self {
257        self.max_batch_size = size;
258        self
259    }
260
261    /// Enable GPU acceleration.
262    pub fn with_gpu(mut self, use_gpu: bool) -> Self {
263        self.use_gpu = use_gpu;
264        self
265    }
266
267    /// Set the number of CPU threads.
268    pub fn with_num_threads(mut self, threads: usize) -> Self {
269        self.num_threads = Some(threads);
270        self
271    }
272
273    /// Set the number of parallel ONNX sessions in the pool.
274    pub fn with_session_pool_size(mut self, size: usize) -> Self {
275        self.session_pool_size = size.max(1);
276        self
277    }
278}
279
280#[cfg(test)]
281mod tests {
282    use super::*;
283
284    #[test]
285    fn test_model_ids() {
286        assert_eq!(
287            EmbeddingModel::BgeLarge.model_id(),
288            "BAAI/bge-large-en-v1.5"
289        );
290        assert_eq!(
291            EmbeddingModel::MiniLM.model_id(),
292            "sentence-transformers/all-MiniLM-L6-v2"
293        );
294        assert_eq!(
295            EmbeddingModel::BgeSmall.model_id(),
296            "BAAI/bge-small-en-v1.5"
297        );
298        assert_eq!(EmbeddingModel::E5Small.model_id(), "intfloat/e5-small-v2");
299    }
300
301    #[test]
302    fn test_dimensions() {
303        assert_eq!(EmbeddingModel::BgeLarge.dimension(), 1024);
304        assert_eq!(EmbeddingModel::MiniLM.dimension(), 384);
305        assert_eq!(EmbeddingModel::BgeSmall.dimension(), 384);
306        assert_eq!(EmbeddingModel::E5Small.dimension(), 384);
307        // Verify each model reports its own dimension
308        for model in EmbeddingModel::all() {
309            assert!(model.dimension() > 0);
310        }
311    }
312
313    #[test]
314    fn test_from_str() {
315        assert_eq!(
316            EmbeddingModel::parse("bge-large"),
317            Some(EmbeddingModel::BgeLarge)
318        );
319        assert_eq!(
320            EmbeddingModel::parse("minilm"),
321            Some(EmbeddingModel::MiniLM)
322        );
323        assert_eq!(
324            EmbeddingModel::parse("BGE-SMALL"),
325            Some(EmbeddingModel::BgeSmall)
326        );
327        assert_eq!(EmbeddingModel::parse("e5"), Some(EmbeddingModel::E5Small));
328        assert_eq!(EmbeddingModel::parse("unknown"), None);
329    }
330
331    #[test]
332    fn test_e5_prefixes() {
333        assert_eq!(EmbeddingModel::E5Small.query_prefix(), Some("query: "));
334        assert_eq!(EmbeddingModel::E5Small.document_prefix(), Some("passage: "));
335        assert_eq!(EmbeddingModel::MiniLM.query_prefix(), None);
336    }
337
338    #[test]
339    fn test_onnx_filenames() {
340        // INT8 model for CPU — all models use the same quantized file
341        for model in EmbeddingModel::all() {
342            assert_eq!(model.onnx_filename(), "onnx/model_quantized.onnx");
343        }
344        // FP32 model for GPU — no Memcpy fallback ops
345        for model in EmbeddingModel::all() {
346            assert_eq!(model.onnx_filename_gpu(), "onnx/model.onnx");
347        }
348        // Sanity: GPU and CPU filenames are distinct
349        assert_ne!(
350            EmbeddingModel::BgeLarge.onnx_filename(),
351            EmbeddingModel::BgeLarge.onnx_filename_gpu()
352        );
353    }
354}
inference/models.rs

inference/
models.rs
