Ashx098/Jina-v5-nano_ONNX-for-vespa

Jina v5 Nano ONNX for Vespa

This is a Jina v5 nano embedding model in ONNX format, optimized for use with Vespa database.

Model Details

Property	Value
Architecture	Jina v5 nano embedding model
Embedding Dimension	768
Input	Token IDs (int64) + Attention Mask
Output	last_hidden_state (needs pooling)

Files

File	Description
model.onnx	ONNX model
tokenizer.json	HuggingFace tokenizer
special_tokens_map.json	Special tokens mapping

Usage with ONNX Runtime

import onnxruntime as ort
from transformers import PreTrainedTokenizerFast
import numpy as np

# Load ONNX model
session = ort.InferenceSession('model.onnx')

# Load tokenizer
tokenizer = PreTrainedTokenizerFast(tokenizer_file='tokenizer.json')

# Generate embedding
text = 'Your search query'
inputs = tokenizer(text, padding=True, truncation=True, max_length=512, return_tensors='np')

outputs = session.run(None, {
    'input_ids': inputs['input_ids'],
    'attention_mask': inputs['attention_mask']
})

# Mean pooling
hidden = outputs[0]
mask = inputs['attention_mask']
mask_expanded = np.expand_dims(mask, -1)
pooled = np.sum(hidden * mask_expanded, axis=1) / np.clip(mask_expanded.sum(axis=1), a_min=1e-9)

print(pooled.shape)  # (1, 768)

Requirements

• onnx
• onnxruntime
• transformers
• numpy