| license: mit | |
| language: | |
| - multilingual | |
| library_name: onnxruntime | |
| tags: | |
| - sentence-similarity | |
| - feature-extraction | |
| - onnx | |
| - int8 | |
| base_model: intfloat/multilingual-e5-base | |
| # Leonis Embedder — multilingual-e5-base (int8 ONNX) | |
| On-device query embedder for the **Leonis AI** maritime assistant (Android). | |
| This is `intfloat/multilingual-e5-base` exported to ONNX and dynamically | |
| quantized to int8, paired with a SentencePiece tokenizer baked into an | |
| `onnxruntime-extensions` graph so the model takes a raw `"query: …"` string. | |
| It must match the model the desktop app indexes passages with, so query and | |
| passage vectors share one 768-dim space. | |
| | File | Size | Purpose | | |
| |------|------|---------| | |
| | `model_int8.onnx` | ~278 MB | e5-base int8 encoder → `last_hidden_state` | | |
| | `tokenizer.onnx` | ~5 MB | SentencePiece tokenizer (string → token ids) | | |
| **Pipeline:** `"query: " + text` → `tokenizer.onnx` → `model_int8.onnx` → | |
| mean-pool (attention mask) → L2-normalize → 768-dim unit vector. | |
| Verified parity vs the desktop sentence-transformers embedder: cosine | |
| 0.974–0.983 (full int8 chain), 1.000 for the fp32 export. | |
| License: MIT (inherits from the base model). | |