| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | # ruff: noqa: E501 |
| 4 | |
| 5 | """Example Python client for multimodal rerank API which is compatible with |
| 6 | Jina and Cohere https://jina.ai/reranker |
| 7 | |
| 8 | Run `vllm serve <model> --runner pooling` to start up the server in vLLM. |
| 9 | e.g. |
| 10 | vllm serve jinaai/jina-reranker-m0 --runner pooling |
| 11 | |
| 12 | vllm serve Qwen/Qwen3-VL-Reranker-2B \ |
| 13 | --runner pooling \ |
| 14 | --max-model-len 4096 \ |
| 15 | --hf_overrides '{"architectures": ["Qwen3VLForSequenceClassification"],"classifier_from_token": ["no", "yes"],"is_original_qwen3_reranker": true}' \ |
| 16 | --chat-template examples/pooling/score/template/qwen3_vl_reranker.jinja |
| 17 | """ |
| 18 | |
| 19 | import argparse |
| 20 | import pprint |
| 21 | |
| 22 | import requests |
| 23 | |
| 24 | from vllm.multimodal.utils import encode_image_url, fetch_image |
| 25 | |
| 26 | query = "A woman playing with her dog on a beach at sunset." |
| 27 | document = ( |
| 28 | "A woman shares a joyful moment with her golden retriever on a sun-drenched beach at sunset, " |
| 29 | "as the dog offers its paw in a heartwarming display of companionship and trust." |
| 30 | ) |
| 31 | image_url = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen-VL/assets/demo.jpeg" |
| 32 | video_url = "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/draw.mp4" |
| 33 | documents = [ |
| 34 | { |
| 35 | "type": "text", |
| 36 | "text": document, |
| 37 | }, |
| 38 | { |
| 39 | "type": "image_url", |
| 40 | "image_url": {"url": image_url}, |
| 41 | }, |
| 42 | { |
| 43 | "type": "image_url", |
| 44 | "image_url": {"url": encode_image_url(fetch_image(image_url))}, |
| 45 | }, |
| 46 | { |
| 47 | "type": "video_url", |
| 48 | "video_url": {"url": video_url}, |
| 49 | }, |
| 50 | ] |
| 51 | |
| 52 | |
| 53 | def parse_args(): |
| 54 | parser = argparse.ArgumentParser() |
| 55 | parser.add_argument("--host", type=str, default="localhost") |
| 56 | parser.add_argument("--port", type=int, default=8000) |
| 57 | return parser.parse_args() |
| 58 | |
| 59 | |
| 60 | def main(args): |
| 61 | base_url = f"http://{args.host}:{args.port}" |
| 62 | models_url = base_url + "/v1/models" |
| 63 | rerank_url = base_url + "/rerank" |
| 64 | |
| 65 | response = requests.get(models_url) |
| 66 | model = response.json()["data"][0]["id"] |
| 67 | |
| 68 | print("Query: string & Document: list of string") |
| 69 | prompt = {"model": model, "query": query, "documents": [document]} |
| 70 | response = requests.post(rerank_url, json=prompt) |
| 71 | pprint.pprint(response.json()) |
| 72 | |
| 73 | print("Query: string & Document: text") |
| 74 | prompt = {"model": model, "query": query, "documents": {"content": [documents[0]]}} |
| 75 | response = requests.post(rerank_url, json=prompt) |
| 76 | pprint.pprint(response.json()) |
| 77 | |
| 78 | print("Query: string & Document: image url") |
| 79 | prompt = { |
| 80 | "model": model, |
| 81 | "query": query, |
| 82 | "documents": {"content": [documents[1]]}, |
| 83 | } |
| 84 | response = requests.post(rerank_url, json=prompt) |
| 85 | pprint.pprint(response.json()) |
| 86 | |
| 87 | print("Query: string & Document: image base64") |
| 88 | prompt = { |
| 89 | "model": model, |
| 90 | "query": query, |
| 91 | "documents": {"content": [documents[2]]}, |
| 92 | } |
| 93 | response = requests.post(rerank_url, json=prompt) |
| 94 | pprint.pprint(response.json()) |
| 95 | |
| 96 | print("Query: string & Document: video url") |
| 97 | prompt = { |
| 98 | "model": model, |
| 99 | "query": query, |
| 100 | "documents": {"content": [documents[3]]}, |
| 101 | } |
| 102 | response = requests.post(rerank_url, json=prompt) |
| 103 | pprint.pprint(response.json()) |
| 104 | |
| 105 | print("Query: string & Document: text + image url") |
| 106 | prompt = { |
| 107 | "model": model, |
| 108 | "query": query, |
| 109 | "documents": {"content": [documents[0], documents[1]]}, |
| 110 | } |
| 111 | response = requests.post(rerank_url, json=prompt) |
| 112 | pprint.pprint(response.json()) |
| 113 | |
| 114 | print("Query: string & Document: list") |
| 115 | prompt = { |
| 116 | "model": model, |
| 117 | "query": query, |
| 118 | "documents": [ |
| 119 | document, |
| 120 | {"content": [documents[0]]}, |
| 121 | {"content": [documents[1]]}, |
| 122 | {"content": [documents[0], documents[1]]}, |
| 123 | ], |
| 124 | } |
| 125 | response = requests.post(rerank_url, json=prompt) |
| 126 | pprint.pprint(response.json()) |
| 127 | |
| 128 | |
| 129 | if __name__ == "__main__": |
| 130 | args = parse_args() |
| 131 | main(args) |