In our experiments, a single canonical query-writing prompt was applied across all datasets to ensure consistency and reproducibility.
1QUERY_WRITER_PROMPT = (
2 "For the input query, formulating a concise search query for dense retrieval by distilling the core intent from a complex user prompt and ignoring LLM instructions."
3 "The response should be less than 200 words"
4)
This model is a key enabler for
INF-X-Retriever's state-of-the-art performance, currently holding the
No. 1 position on the
BRIGHT Benchmark (as of Dec 17, 2025).
For more details on the full framework, please visit the
INF-X-Retriever Repository.
1from transformers import AutoModelForCausalLM, AutoTokenizer
2
3# Load model and tokenizer
4model_name = "infly/inf-query-aligner"
5model = AutoModelForCausalLM.from_pretrained(
6 model_name,
7 torch_dtype="auto",
8 device_map="auto"
9)
10tokenizer = AutoTokenizer.from_pretrained(model_name)
11
12# Define input query
13query = "Claim in article about why insects are attracted to light\nIn this article they are addressing the reason insects are attracted to light when they say\nHeat radiation as an attractive component is refuted by the effect of LED lighting, which supplies negligible infrared radiation yet still entraps vast numbers of insects.\nI don't see why attraction to LEDs shows they're not seeking heat. Could they for example be evolutionarily programmed to associate light with heat? So that even though they don't encounter heat near/on the LEDs they still \"expect\" to?"
14
15QUERY_WRITER_PROMPT = (
16 "For the input query, formulating a concise search query for dense retrieval by distilling the core intent from a complex user prompt and ignoring LLM instructions."
17 "The response should be less than 200 words"
18)
19messages = [
20 {
21 "role": "system",
22 "content": "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.",
23 },
24 {
25 "role": "user",
26 "content": (
27 f"{QUERY_WRITER_PROMPT}\n\n"
28 f"**Input Query:**\n{query}\n"
29 f"**Your Output:**\n"
30 ),
31 },
32]
33
34# Apply chat template
35text = tokenizer.apply_chat_template(
36 messages,
37 tokenize=False,
38 add_generation_prompt=True
39)
40model_inputs = tokenizer(
41 [text],
42 truncation=True,
43 max_length=8192,
44 return_tensors="pt"
45).to(model.device)
46
47# Generate rewritten query
48generated_ids = model.generate(
49 **model_inputs,
50 max_new_tokens=512
51)
52generated_ids = [
53 output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
54]
55
56response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
57
58print(response)
INF-X-Retriever achieves state-of-the-art results on the
BRIGHT Benchmark (as of Dec 20, 2025).
1@misc{inf-x-retriever-2025,
2 title = {INF-X-Retriever},
3 author = {Yichen Yao, Jiahe Wan, Yuxin Hong, Mengna Zhang, Junhan Yang, Zhouyu Jiang, Qing Xu, Kuan Lu, Yinghui Xu, Wei Chu, Emma Wang, Yuan Qi},
4 year = {2025},
5 url = {https://yaoyichen.github.io/INF-X-Retriever},
6 publisher = {GitHub repository}
7}
