Skip to content

Jina Reranker vLLM Deployment Recipe

This guide contains deployment instructions for the jinaai/jina-reranker-m0 using vLLM. This is a multilingual, multimodal reranker model designed to rank visual documents across multiple languages. It processes both textual and visual content, including pages with mixed text, figures, tables, and various layouts across over 29 languages.

This guide uses 2x NVIDIA T4 GPUs or 2x NVIDIA L4 GPUs to launch this model.

Installation

Install vLLM and required dependencies:

uv pip install vllm

Installing vLLM (For AMD ROCm: MI300x/MI325x/MI355x)

uv pip install vllm --extra-index-url https://wheels.vllm.ai/rocm/0.14.1/rocm700
⚠️ The vLLM wheel for ROCm is compatible with Python 3.12, ROCm 7.0, and glibc >= 2.35. If your environment is incompatible, please use docker flow in vLLM

Online Deployment

Deploy the model as a production-ready API server using vLLM.

1. Deploy Model Server

# https://docs.vllm.ai/en/latest/cli/serve.html
vllm serve jinaai/jina-reranker-m0 \
    --host 0.0.0.0 \
    --port 8000 \
    --tensor_parallel_size 2 \
    --gpu-memory-utilization 0.75 \
    --max_num_seqs 32

1.1 Deploy Model Server on 8xMI300x/MI325x/MI355x

# https://docs.vllm.ai/en/latest/cli/serve.html
export VLLM_ROCM_USE_AITER=1
vllm serve jinaai/jina-reranker-m0 \
    --host 0.0.0.0 \
    --port 8000 \
    --tensor_parallel_size 2 \
    --gpu-memory-utilization 0.75 \
    --max_num_seqs 32
* You can set export VLLM_ROCM_USE_AITER=1 for Better Performance on AMD GPUs. The default is export VLLM_ROCM_USE_AITER=0

2. Rerank API

The Rerank API returns a ranked list of documents ordered by their relevance to the query.

Request Format

curl -X POST http://localhost:8000/v1/rerank \
-H "accept: application/json" \
-H "Content-Type: application/json" \
-d '{
    "model": "jinaai/jina-reranker-m0",
    "query": "What are the health benefits of green tea?",
    "documents": [
        "Green tea contains antioxidants called catechins that may help reduce inflammation and protect cells from damage.",
        "El precio del café ha aumentado un 20% este año debido a problemas en la cadena de suministro.",
        "Studies show that drinking green tea regularly can improve brain function and boost metabolism.",
        "Basketball is one of the most popular sports in the United States.",
        "绿茶富含儿茶素等抗氧化剂，可以降低心脏病风险，还有助于控制体重。",
        "Le thé vert est riche en antioxydants et peut améliorer la fonction cérébrale."
    ],
    "top_n": 3,
    "return_documents": true
}'

Response Format

{
    "id": "rerank-f0a2c978b4fb4d61b0a54fd1c05e335f",
    "model": "jinaai/jina-reranker-m0",
    "usage": {
        "total_tokens": 225
    },
    "results": [
        {
            "index": 4,
            "document": {
                "text": "绿茶富含儿茶素等抗氧化剂，可以降低心脏病风险，还有助于控制体重。",
                "multi_modal": null
            },
            "relevance_score": 0.9823843836784363
        },
        {
            "index": 0,
            "document": {
                "text": "Green tea contains antioxidants called catechins that may help reduce inflammation and protect cells from damage.",
                "multi_modal": null
            },
            "relevance_score": 0.9777672290802002
        },
        {
            "index": 2,
            "document": {
                "text": "Studies show that drinking green tea regularly can improve brain function and boost metabolism.",
                "multi_modal": null
            },
            "relevance_score": 0.9752224683761597
        }
    ]
}

3. Score API

The Score API computes similarity scores between a query and multiple documents without ranking them.

Text-to-Text Scoring

curl -X POST http://localhost:8000/v1/score \
-H "accept: application/json" \
-H "Content-Type: application/json" \
-d '{
    "model": "jinaai/jina-reranker-m0",
    "text_1": [
        "What is the capital of Brazil?",
        "What is the capital of France?"
    ],
    "text_2": [
        "The capital of Brazil is Brasilia.",
        "The capital of France is Paris."
    ]
}'

Request Parameters

  • model: Model identifier (required)
  • text_1: Query text (required)
  • text_2: Document(s) to score against. Can be a single string or an array of strings (required)

Response Format

{
    "id":"score-30d069df61924c4292579640c0d97bcc",
    "object":"list",
    "created":1761686670,
    "model":"jinaai/jina-reranker-m0",
    "data":[
        {
            "index":0,
            "object":"score",
            "score":0.9878721237182617
        },
        {
            "index":1,
            "object":"score",
            "score":0.9879010915756226
        }
    ],
    "usage":{
        "prompt_tokens":47,
        "total_tokens":47,
        "completion_tokens":0,
        "prompt_tokens_details":null
    }
}

Multimodal Scoring

The Score API supports multimodal inputs, allowing you to score text against images or vice versa:

curl -X POST http://localhost:8000/v1/score \
-H "accept: application/json" \
-H "Content-Type: application/json" \
-d '{
    "model": "jinaai/jina-reranker-m0",
    "text_1": "A cat",
    "text_2": {
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url": "cat_img.jpg"
                }
            },
            {
                "type": "image_url",
                "image_url": {
                    "url": "dog_img.jpg"
                }
            }
        ]
    }
}'

Offline Deployment

Use the model directly in your Python code without running a server.

from vllm import LLM

MODEL = "jinaai/jina-reranker-m0"

# Initialize the LLM engine
llm = LLM(
    model=MODEL,
    tensor_parallel_size=2,
    gpu_memory_utilization=0.75,
    max_model_len=1024,
    max_num_seqs=32,
    kv_cache_dtype="fp8",
    dtype="bfloat16",
)

# Prepare query and documents
query = "fast recipes for weeknight dinners"
documents = [
    "A 65-minute pasta with garlic and olive oil.",
    "Slow braised short ribs that cook for 5 hours.",
    "Stir-fry veggies with pre-cooked rice.",
]

# Compute scores
res = llm.score(query, documents)

# Extract and print scores
for item in res:
    print(item.outputs.score)

Resources