llm-kernel 0.25.0

# Supported Embedding Models

llm-kernel supports **46 embedding models** across four backends.

## ONNX Models (fastembed-rs)

44 models via ONNX Runtime. Enable with `embedding-fastembed` feature.
No API key required — models download from HuggingFace on first use.

| Model | Dim | Quantized | Description |
|-------|-----|-----------|-------------|
| **BGE / BAAI** | | | |
| `BGESmallENV15` | 384 | | v1.5 release of the fast and default English model |
| `BGESmallENV15Q` | 384 | ✅ | Quantized v1.5 release of the fast and default English model |
| `BGEBaseENV15` | 768 | | v1.5 release of the base English model |
| `BGEBaseENV15Q` | 768 | ✅ | Quantized v1.5 release of the base English model |
| `BGELargeENV15` | 1024 | | v1.5 release of the large English model |
| `BGELargeENV15Q` | 1024 | ✅ | Quantized v1.5 release of the large English model |
| `BGESmallZHV15` | 512 | | v1.5 release of the small Chinese model |
| `BGELargeZHV15` | 1024 | | v1.5 release of the large Chinese model |
| `BGEM3` | 1024 | | Multilingual M3 model with 8192 context length, 100+ languages |
| **Sentence Transformers** | | | |
| `AllMiniLML6V2` | 384 | | Sentence Transformer model, MiniLM-L6-v2 |
| `AllMiniLML6V2Q` | 384 | ✅ | Quantized Sentence Transformer model, MiniLM-L6-v2 |
| `AllMiniLML12V2` | 384 | | Sentence Transformer model, MiniLM-L12-v2 |
| `AllMiniLML12V2Q` | 384 | ✅ | Quantized Sentence Transformer model, MiniLM-L12-v2 |
| `AllMpnetBaseV2` | 768 | | Sentence Transformer model, mpnet-base-v2 |
| **Nomic** | | | |
| `NomicEmbedTextV1` | 768 | | 8192 context length english model |
| `NomicEmbedTextV15` | 768 | | v1.5 release of the 8192 context length english model |
| `NomicEmbedTextV15Q` | 768 | ✅ | Quantized v1.5 release of the 8192 context length english model |
| **Paraphrase** | | | |
| `ParaphraseMLMiniLML12V2` | 384 | | Multi-lingual model |
| `ParaphraseMLMiniLML12V2Q` | 384 | ✅ | Quantized multi-lingual model |
| `ParaphraseMLMpnetBaseV2` | 768 | | Sentence-transformers model for clustering or semantic search |
| **ModernBERT** | | | |
| `ModernBertEmbedLarge` | 1024 | | Large model of ModernBert Text Embeddings |
| **E5 Multilingual** | | | |
| `MultilingualE5Small` | 384 | | Small model of multilingual E5 Text Embeddings |
| `MultilingualE5Base` | 768 | | Base model of multilingual E5 Text Embeddings |
| `MultilingualE5Large` | 1024 | | Large model of multilingual E5 Text Embeddings |
| **Mixedbread** | | | |
| `MxbaiEmbedLargeV1` | 1024 | | Large English embedding model from MixedBreed.ai |
| `MxbaiEmbedLargeV1Q` | 1024 | ✅ | Quantized large English embedding model from MixedBreed.ai |
| **GTE (Alibaba)** | | | |
| `GTEBaseENV15` | 768 | | Base multilingual embedding model from Alibaba |
| `GTEBaseENV15Q` | 768 | ✅ | Quantized base multilingual embedding model from Alibaba |
| `GTELargeENV15` | 1024 | | Large multilingual embedding model from Alibaba |
| `GTELargeENV15Q` | 1024 | ✅ | Quantized large multilingual embedding model from Alibaba |
| **CLIP** | | | |
| `ClipVitB32` | 512 | | CLIP text encoder based on ViT-B/32 (image+text) |
| **Jina** | | | |
| `JinaEmbeddingsV2BaseCode` | 768 | | Jina embeddings v2 base code |
| `JinaEmbeddingsV2BaseEN` | 768 | | Jina embeddings v2 base English |
| **Gemma** | | | |
| `EmbeddingGemma300M` | 768 | | EmbeddingGemma 300M parameter model from Google |
| **Snowflake Arctic** | | | |
| `SnowflakeArcticEmbedXS` | 384 | | Snowflake Arctic embed model, xs |
| `SnowflakeArcticEmbedXSQ` | 384 | ✅ | Quantized Snowflake Arctic embed model, xs |
| `SnowflakeArcticEmbedS` | 384 | | Snowflake Arctic embed model, small |
| `SnowflakeArcticEmbedSQ` | 384 | ✅ | Quantized Snowflake Arctic embed model, small |
| `SnowflakeArcticEmbedM` | 768 | | Snowflake Arctic embed model, medium |
| `SnowflakeArcticEmbedMQ` | 768 | ✅ | Quantized Snowflake Arctic embed model, medium |
| `SnowflakeArcticEmbedMLong` | 768 | | Snowflake Arctic embed model, medium with 2048 context |
| `SnowflakeArcticEmbedMLongQ` | 768 | ✅ | Quantized Snowflake Arctic embed model, medium with 2048 context |
| `SnowflakeArcticEmbedL` | 1024 | | Snowflake Arctic embed model, large |
| `SnowflakeArcticEmbedLQ` | 1024 | ✅ | Quantized Snowflake Arctic embed model, large |

### Query/Document Prefixes

Some models require text prefixes for optimal results:

| Model family | Query prefix | Document prefix |
|-------------|-------------|----------------|
| E5 (`MultilingualE5*`) | `query: ` | `passage: ` |
| Snowflake Arctic | `Represent this sentence for searching relevant passages: ` | _(none)_ |

Prefixes are automatically prepended by `FastembedProvider`.

## Candle Models (pure Rust)

No ONNX Runtime — uses candle-nn for GPU/CPU inference.

### Qwen3

Enable with `embedding-fastembed-qwen3` feature.

| HuggingFace repo | Params | Description |
|-----------------|--------|-------------|
| `Qwen/Qwen3-Embedding-0.6B` | 600M | Lightweight Qwen3 embedding |
| `Qwen/Qwen3-Embedding-8B` | 8B | Full-size Qwen3 embedding |
| `Qwen/Qwen3-VL-Embedding-2B` | 2B | Qwen3 VL embedding (text-only mode) |

```rust
use llm_kernel::embedding::{Qwen3Provider, EmbeddingProvider};

let provider = Qwen3Provider::new("Qwen/Qwen3-Embedding-0.6B")?;
let result = provider.embed("hello world")?;
```

### Nomic V2 MoE

Enable with `embedding-fastembed-nomic-moe` feature.

| HuggingFace repo | Active params | Total params | Dim |
|-----------------|--------------|-------------|-----|
| `nomic-ai/nomic-embed-text-v2-moe` | 305M | 475M | 768 |

8 experts with top-2 routing, hidden_size=768.

```rust
use llm_kernel::embedding::{NomicMoeProvider, EmbeddingProvider};

let provider = NomicMoeProvider::new()?;
let result = provider.embed("hello world")?;
```

## MLX Models (Apple Silicon native)

Pure-Rust MLX inference on the Apple Silicon GPU via unified memory — the
**batch-throughput** path that complements candle-Metal (which wins on
single-embed latency). Enable with `embedding-mlx` (macOS / aarch64 only).

`MlxEmbeddingProvider` takes a catalog `EmbeddingModel` and covers the
**vanilla-BERT** subset of the fastembed catalog — 13 base models (21 catalog
variants, counting `*Q` aliases that share weights with their non-quantized
twin). All are CLS- or mean-pooled and L2-normalised, matching the reference
`sentence-transformers` output.

| Catalog variant | Weight repo (`mlx_repo`) | Dim | Pooling | Query prefix |
|-----------------|--------------------------|-----|---------|--------------|
| `BGESmallENV15` | `BAAI/bge-small-en-v1.5` | 384 | CLS | BGE instruction |
| `BGEBaseENV15` | `BAAI/bge-base-en-v1.5` | 768 | CLS | BGE instruction |
| `BGELargeENV15` | `BAAI/bge-large-en-v1.5` | 1024 | CLS | BGE instruction |
| `BGESmallZHV15` | `BAAI/bge-small-zh-v1.5` | 512 | CLS | BGE instruction |
| `AllMiniLML6V2` | `sentence-transformers/all-MiniLM-L6-v2` | 384 | mean | — |
| `AllMiniLML12V2` | `sentence-transformers/all-MiniLM-L12-v2` | 384 | mean | — |
| `ParaphraseMLMiniLML12V2` | `sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2` | 384 | mean | — |
| `MultilingualE5Small` | `intfloat/multilingual-e5-small` | 384 | mean | `query: ` / `passage: ` |
| `SnowflakeArcticEmbedXS` | `Snowflake/snowflake-arctic-embed-xs` | 384 | CLS | BGE instruction |
| `SnowflakeArcticEmbedS` | `Snowflake/snowflake-arctic-embed-s` | 384 | CLS | BGE instruction |
| `SnowflakeArcticEmbedM` | `Snowflake/snowflake-arctic-embed-m` | 768 | CLS | BGE instruction |
| `SnowflakeArcticEmbedL` | `Snowflake/snowflake-arctic-embed-l` | 1024 | CLS | BGE instruction |
| `MxbaiEmbedLargeV1` | `mixedbread-ai/mxbai-embed-large-v1` | 1024 | CLS | BGE instruction |

Each `*Q` variant (e.g. `BGESmallENV15Q`) maps to the same weight repo — MLX
loads the original float checkpoint, so quantized aliases behave identically.

### Not MLX-supported

Membership was determined by probing every catalog model's original weight repo
(`config.json` + `model.safetensors` header). A model qualifies only if it is
`architectures: ["BertModel"]` with absolute position embeddings, gelu, and the
standard `encoder.layer.N.attention.self.*` tensor layout. Excluded:

| Model(s) | Reason |
|----------|--------|
| `SnowflakeArcticEmbedMLong`, `NomicEmbedTextV1`, `NomicEmbedTextV15` | `NomicBertModel` — fused `attn.Wqkv`, rotary/no position embeddings |
| `GTEBaseENV15`, `GTELargeENV15` | `NewModel` architecture |
| `MultilingualE5Base`, `MultilingualE5Large`, `ParaphraseMLMpnetBaseV2`, `BGEM3` | `XLMRobertaModel` |
| `AllMpnetBaseV2` | `MPNetForMaskedLM` |
| `JinaEmbeddingsV2BaseEN`, `JinaEmbeddingsV2BaseCode` | `JinaBertForMaskedLM` (ALiBi) |
| `ModernBertEmbedLarge` | `ModernBertModel` |
| `EmbeddingGemma300M`, `ClipVitB32` | Not BERT text encoders |
| `BGELargeZHV15` | `BertModel`, but the repo ships only `pytorch_model.bin` — no safetensors |

Calling `MlxEmbeddingProvider::new` with any of these returns an error rather
than failing later at weight load.

```rust
use llm_kernel::embedding::{EmbeddingModel, MlxEmbeddingProvider, EmbeddingProvider};

let provider = MlxEmbeddingProvider::new(EmbeddingModel::BGESmallENV15)?;

// Asymmetric models apply the query prefix on `embed` and none on
// `embed_document` — use the matching call for each side of a search.
let query = provider.embed("What is the capital of France?")?;
let doc = provider.embed_document("The capital of France is Paris.")?;

// Check support before constructing, if the model is chosen at runtime.
assert!(EmbeddingModel::MxbaiEmbedLargeV1.mlx_supported());
assert!(!EmbeddingModel::BGEM3.mlx_supported());
```

Weights load from `model.safetensors` in F32, F16 or BF16 (mxbai ships F16);
any other dtype is an explicit error rather than a silent misread.

## Remote API

| Provider | Feature | Models |
|----------|---------|--------|
| OpenAI | `embedding-openai` | `text-embedding-3-small`, `text-embedding-3-large`, `text-embedding-ada-002` |

```rust
use llm_kernel::embedding::{OpenAIEmbeddingClient, EmbeddingProvider};

let client = OpenAIEmbeddingClient::new("text-embedding-3-small", &api_key)?;
let result = client.embed("hello world")?;
```

## Quick comparison

| Backend | Feature | Offline | Dim range | Notes |
|---------|---------|---------|-----------|-------|
| ONNX (fastembed) | `embedding-fastembed` | ✅ | 384–1024 | 44 models, auto-download |
| Qwen3 (candle) | `embedding-fastembed-qwen3` | ✅ | varies | Pure Rust, GPU support |
| Nomic V2 MoE (candle) | `embedding-fastembed-nomic-moe` | ✅ | 768 | MoE, lightweight |
| Vanilla BERT (MLX) | `embedding-mlx` | ✅ | 384–1024 | 13 models, macOS/aarch64 only, batch throughput |
| OpenAI | `embedding-openai` | ❌ | 1536–3072 | Remote API |