jina-embeddings-text-nano-retrieval

jina-embeddings-v5-text: Task-Targeted Embedding Distillation

Elastic Inference Service | ArXiv | Release Note | Blog

Model Overview

jina-embeddings-v5-text-nano-retrieval is a compact, high-performance text embedding model designed for information retrieval.

It is part of the jina-embeddings-v5-text model family, which also includes jina-embeddings-v5-text-small, for better performance at a bigger size.

Trained using a novel approach that combines distillation with task-specific contrastive losses, jina-embeddings-v5-text-nano-retrieval outperforms existing state-of-the-art models of similar size across diverse embedding benchmarks.

Training and Evaluation

For training details and evaluation results, see our technical report.

Usage

The following Python packages are required:

• transformers>=5.1.0
• torch>=2.8.0
• peft>=0.15.2
• vllm==0.15.1

Optional / Recommended

• flash-attention: Installing flash-attention is recommended for improved inference speed and efficiency, but not mandatory.
• sentence-transformers: If you want to use the model via the sentence-transformers interface, install this package as well.

The fastest way to use v5-text in production. Elastic Inference Service (EIS) provides managed embedding inference with built-in scaling, so you can generate embeddings directly within your Elastic deployment.

PUT _inference/text_embedding/jina-v5
{
  "service": "elastic",
  "service_settings": {
    "model_id": "jina-embeddings-v5-text-nano"
  }
}

See the Elastic Inference Service documentation for setup details.

from sentence_transformers import SentenceTransformer
import torch

model = SentenceTransformer(
    "jinaai/jina-embeddings-v5-text-nano-retrieval",
    trust_remote_code=True,
    model_kwargs={"dtype": torch.bfloat16},  # Recommended for GPUs
    config_kwargs={"_attn_implementation": "flash_attention_2"},  # Recommended but optional
)
# Optional: set truncate_dim in encode() to control embedding size

query = "Which planet is known as the Red Planet?"
documents = [
    "Venus is often called Earth's twin because of its similar size and proximity.",
    "Mars, known for its reddish appearance, is often referred to as the Red Planet.",
    "Jupiter, the largest planet in our solar system, has a prominent red spot.",
    "Saturn, famous for its rings, is sometimes mistaken for the Red Planet.",
]

# Encode query and documents
query_embeddings = model.encode(sentences=query, prompt_name="query")
document_embeddings = model.encode(sentences=documents, prompt_name="document")
print(query_embeddings.shape, document_embeddings.shape)
# (768,) (4, 768)

similarity = model.similarity(query_embeddings, document_embeddings)
print(similarity)
# tensor([[0.5013, 0.7914, 0.6133, 0.5736]])

from vllm import LLM
from vllm.config.pooler import PoolerConfig

# Initialize model
name = "jinaai/jina-embeddings-v5-text-nano-retrieval"
model = LLM(
    model=name,
    dtype="float16",
    runner="pooling",
    trust_remote_code=True,
    pooler_config=PoolerConfig(seq_pooling_type="LAST", normalize=True),
)

# Create text prompts
query = "Overview of climate change impacts on coastal cities"
query_prompt = f"Query: {query}"

document = "The impacts of climate change on coastal cities are significant.."
document_prompt = f"Document: {document}"

# Encode all prompts
prompts = [query_prompt, document_prompt]
outputs = model.encode(prompts, pooling_task="embed")

Since our nano model is based on jinaai/jina-embeddings-v5-text-nano, which is not yet supported by llama.cpp, we provide our own branch of llama.cpp, which implements the necessary changes to support it for now.

To start the OpenAI API compatible HTTP server, run with the respective model version:

llama-server \
  -hf jinaai/jina-embeddings-v5-text-nano-retrieval:F16 \
  --embedding \
  --pooling last \
  --batch-size 8192 \
  --ubatch-size 8192 \
  --ctx-size 8192

Client:

curl -X POST "http://127.0.0.1:8080/v1/embeddings" \
  -H "Content-Type: application/json" \
  -d '{
    "input": [
      "Query: A beautiful sunset over the beach",
      "Query: Un beau coucher de soleil sur la plage",
      "Document: 海滩上美丽的日落",
      "Document: 浜辺に沈む美しい夕日",
      "Document: Golden sunlight melts into the horizon, painting waves in warm amber and rose, while the sky whispers goodnight to the quiet, endless sea."
    ]
  }'

Note: For the retrieval variant, add Query: or Document: prefix in front of your input as shown above.

You can run the ONNX-optimized version of the model locally using Hugging Face's optimum library. Make sure you have the required dependencies installed (e.g., pip install optimum[onnxruntime] transformers torch):

from optimum.onnxruntime import ORTModelForFeatureExtraction
from transformers import AutoTokenizer
import torch

model_id = "jinaai/jina-embeddings-v5-text-nano-retrieval"

# 1. Load tokenizer and ONNX model
# We specify the subfolder 'onnx' where the weights are located
tokenizer = AutoTokenizer.from_pretrained(model_id, trust