---
language: multilingual
license: other
datasets:
  - jinaai/jina-vdr
pipeline_tag: feature-extraction
tags:
  - embeddings
  - multilingual-embeddings
  - multimodal-embeddings
  - text-to-image
  - sentence-transformers
  - sentence-similarity
  - visual-document-retrieval
---

# Custom Embedding Model

This is a custom embedding model based on the Jina Embeddings V4 architecture, specially adapted for embedding tasks involving text, images, and visual documents.

## Model Description

The model supports:

- **Multimodal Embeddings**: Generate unified embeddings for text and images
- **Multilingual Support**: Works across 30+ languages
- **Task-specific Modes**: Optimized for retrieval, text-matching, and code tasks
- **Flexible Dimensions**: Dense embeddings that can be truncated with minimal performance loss

## Usage

### Text Embeddings

```python
from custom_st import Transformer

# Initialize the model
model = Transformer(
    model_name_or_path="path/to/model",
    model_args={"default_task": "retrieval", "trust_remote_code": True},
    trust_remote_code=True
)

# Encode text
texts = ["Your text here", "Another text example"]
features = model.tokenize(texts)
outputs = model.forward(features, task="retrieval")
embeddings = outputs["sentence_embedding"]
```

### Image Embeddings

```python
from PIL import Image
from custom_st import Transformer

# Initialize the model
model = Transformer(
    model_name_or_path="path/to/model",
    model_args={"default_task": "retrieval", "trust_remote_code": True},
    trust_remote_code=True
)

# Load images
images = [Image.open("image1.jpg"), Image.open("image2.jpg")]
# Or use URLs
image_urls = ["http://example.com/image1.jpg", "http://example.com/image2.jpg"]

# Encode images
features = model.tokenize(images)  # or model.tokenize(image_urls)
outputs = model.forward(features, task="retrieval")
embeddings = outputs["sentence_embedding"]
```

## Requirements

- Python 3.8+
- PyTorch 2.0+
- Transformers 4.30+
- PEFT 0.4+
- Pillow 9.0+

## License

This model is available under the same terms as the original model it's based on. Please refer to the license information in the repository for details.