jina-embeddings-v3-ort

jina-embeddings-v3-ort (ONNX Model Mirror)

This repository provides the official ONNX export files of jinaai/jina-embeddings-v3 (570M parameters, multilingual XLM-RoBERTa + 5 task-specific LoRA adapters) optimized and verified for standalone Rust inference via the jina-embeddings-v3-ort crate.

Attribution & Licensing: This repository is a companion ONNX mirror of the official model developed and published by Jina AI. The underlying model weights remain strictly licensed under the original CC-BY-NC-4.0 license by Jina AI. Full credit belongs to the Jina AI team.

---

Model Details

• Base Architecture: XLM-RoBERTa with 570M parameters.
• Output Embedding Dimension: 1024 float32 (mean-pooled, L2-normalized).
• Supported Languages: 89+ languages including Persian (Farsi), English, Arabic, German, French, Spanish, Russian, Chinese, Japanese, etc.
• Dynamic Task Adapters (LoRA): Selectable via task_id input tensor during inference.

LoRA Task Adaptation Map

---

Numerical Parity & Verification

This ONNX model has been comprehensively verified against the official PyTorch AutoModel and Python reference implementation:

• PyTorch vs. ONNX Parity: Cosine similarity $> 0.9999997$ (Max absolute error $< 4.51 \times 10^{-7}$)
• Pure-Rust vs. Ground Truth Parity: 500/500 vector tests passed (min_cosine = 0.99999893, mean_cosine = 1.00000005, max_diff = 4.32e-6)
• Multilingual Cross-Lingual Alignment: Mean cosine similarity $0.8092$ for paired Persian-English project descriptions.

---

Quickstart in Rust

Use the companion Rust crate jina-embeddings-v3-ort:

use std::path::Path;
use jina_embeddings_v3_ort::{JinaEmbedder, JinaTask};

fn main() -> anyhow::Result<()> {
    let model_path = Path::new("onnx/model.onnx");
    let tokenizer_path = Path::new("tokenizer.json");

    let embedder = JinaEmbedder::new(model_path, tokenizer_path)?;

    // Single text embedding (1024-dim, float32, normalized)
    let text = "ربات تلگرام دانلود خودکار ویدیو";
    let embedding = embedder.embed(text, JinaTask::TextMatching)?;
    assert_eq!(embedding.len(), 1024);

    // Batched embedding
    let texts = &["Rust web microservice", "React frontend dashboard"];
    let batch = embedder.embed_batch(texts, JinaTask::RetrievalPassage)?;
    assert_eq!(batch.len(), 2);

    Ok(())
}

---

Files in this Repository

• onnx/model.onnx: ONNX computation graph definition.
• onnx/model.onnx_data: External ONNX weight tensor storage (~2.29 GB).
• tokenizer.json: XLM-RoBERTa tokenizer vocabulary and merge rules.
• tokenizer_config.json: Tokenizer configuration.
• special_tokens_map.json: Special token definitions.
• config.json: Original model configuration with lora_adaptations mapping.