README.md · inf-retriever-v1-pro

1

---

2

license: apache-2.0

3

language:

4

- en

5

base_model: Qwen/Qwen2.5-7B-Instruct

6

tags:

7

- retrieval

8

- query-rewriting

9

- reinforcement-learning

10

---

11

12

<h1 align="center">inf-retriever-v1-pro</h1>

13

14

<p align="center">

15

    <a href="https://brightbenchmark.github.io/"><img src="https://img.shields.io/badge/BRIGHT_Benchmark-Rank_1st-8A2BE2" alt="Rank"></a>

16

    <a href="https://huggingface.co/infly/inf-retrieve-v1-pro"><img src="https://img.shields.io/badge/🤗%20Hugging%20Face-INF--Retrieve--v1--pro-blue" alt="Hugging Face"></a>

17

    <a href="https://opensource.org/licenses/Apache-2.0"><img src="https://img.shields.io/badge/License-Apache--2.0-green.svg" alt="License"></a>

18

</p>

19

20

## 📖 Overview

21

22

**inf-retriever-v1-pro** is a specialized retrieval component of the **INF-X-Retriever** framework, designed to distill the core retrieval intent from complex, verbose, or reasoning-intensive queries. Built upon the **inf-retriever-v1** foundation and further trained to serve as the retriever within a RAG (retrieval-augmented generation) system, it transforms raw user queries into concise, search-optimized queries for dense retrieval systems.

23

24

In our experiments, a single canonical query-writing prompt was applied across all datasets to ensure consistency and reproducibility.

25

```bash

26

task = 'Given a web search query, retrieve relevant passages that answer the query'

27

```

28

29

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](https://brightbenchmark.github.io/) (as of Dec 17, 2025).

30

31

For more details on the full framework, please visit the [INF-X-Retriever Repository](https://github.com/yaoyichen/INF-X-Retriever).

32

33

---

34

35

### Requirements

36

37

```bash

38

transformers==4.51.0

39

```

40

41

### Usage

42

43

### Sentence Transformers

44

```python

45

from sentence_transformers import SentenceTransformer

46

47

model = SentenceTransformer("infly/inf-retriever-v1", trust_remote_code=True)

48

# In case you want to reduce the maximum length:

49

model.max_seq_length = 8192

50

51

queries = [

52

"how much protein should a female eat",

53

"summit define",

54

]

55

documents = [

56

    "As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day. But, as you can see from this chart, you'll need to increase that if you're expecting or training for a marathon. Check out the chart below to see how much protein you should be eating each day.",

57

    "Definition of summit for English Language Learners. : 1  the highest point of a mountain : the top of a mountain. : 2  the highest level. : 3  a meeting or series of meetings between the leaders of two or more governments.",

58

]

59

60

query_embeddings = model.encode(queries, prompt_name="query")

61

document_embeddings = model.encode(documents)

62

63

scores = (query_embeddings @ document_embeddings.T) * 100

64

print(scores.tolist())

65

# [[91.46116638183594, 76.9832992553711], [70.7034683227539, 87.15817260742188]]

66

```

67

68

### Transformers

69

```python

70

import torch

71

import torch.nn.functional as F

72

73

from torch import Tensor

74

from transformers import AutoTokenizer, AutoModel

75

76

77

def last_token_pool(last_hidden_states: Tensor,

78

attention_mask: Tensor) -> Tensor:

79

left_padding = (attention_mask[:, -1].sum() == attention_mask.shape[0])

80

if left_padding:

81

return last_hidden_states[:, -1]

82

else:

83

sequence_lengths = attention_mask.sum(dim=1) - 1

84

batch_size = last_hidden_states.shape[0]

85

return last_hidden_states[torch.arange(batch_size, device=last_hidden_states.device), sequence_lengths]

86

87

88

def get_detailed_instruct(task_description: str, query: str) -> str:

89

return f'Instruct: {task_description}\nQuery: {query}'

90

91

92

# Each query must come with a one-sentence instruction that describes the task

93

task = 'Given a web search query, retrieve relevant passages that answer the query'

94

queries = [

95

get_detailed_instruct(task, 'how much protein should a female eat'),

96

get_detailed_instruct(task, 'summit define')

97

]

98

# No need to add instruction for retrieval documents

99

documents = [

100

    "As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day. But, as you can see from this chart, you'll need to increase that if you're expecting or training for a marathon. Check out the chart below to see how much protein you should be eating each day.",

101

    "Definition of summit for English Language Learners. : 1  the highest point of a mountain : the top of a mountain. : 2  the highest level. : 3  a meeting or series of meetings between the leaders of two or more governments."

102

]

103

input_texts = queries + documents

104

105

tokenizer = AutoTokenizer.from_pretrained('infly/inf-retriever-v1', trust_remote_code=True)

106

model = AutoModel.from_pretrained('infly/inf-retriever-v1', trust_remote_code=True)

107

108

max_length = 8192

109

110

# Tokenize the input texts

111

batch_dict = tokenizer(input_texts, max_length=max_length, padding=True, truncation=True, return_tensors='pt')

112

outputs = model(**batch_dict)

113

embeddings = last_token_pool(outputs.last_hidden_state, batch_dict['attention_mask'])

114

115

# normalize embeddings

116

embeddings = F.normalize(embeddings, p=2, dim=1)

117

scores = (embeddings[:2] @ embeddings[2:].T) * 100

118

print(scores.tolist())

119

# [[91.46114349365234, 76.98332214355469], [70.7035140991211, 87.158203125]]

120

```

121

122

---

123

124

## Performance

125

126

**INF-X-Retriever** achieves state-of-the-art results on the [BRIGHT Benchmark](https://brightbenchmark.github.io/) (as of Dec 20, 2025).

127

128

The **BRIGHT** (Benchmark for Reasoning-Intensive Grounded HT) is a rigorous text retrieval benchmark designed to evaluate the capability of retrieval models in handling questions that require intensive reasoning and cross-document synthesis. Collected from real-world sources such as StackExchange, competitive programming platforms, and mathematical competitions, it comprises complex queries spanning diverse domains like mathematics, coding, biology, economics, and robotics.

129

130

### Short document

131

132

#### Overall & Category Performance

133

134

135

|:---|:---:|:---:|:---:|:---:|

136

| **INF-X-Retriever** | **63.4** | **68.3** | **55.3** | **57.7** |

137

| DIVER (v3) | 46.8 | 51.8 | 39.9 | 39.7 |

138

| BGE-Reasoner-0928 | 46.4 | 52.0 | 35.3 | 40.7 |

139

| LATTICE | 42.1 | 51.6 | 26.9 | 30.0 |

140

| ReasonRank | 40.8 | 46.9 | 27.6 | 35.5 |

141

| XDR2 | 40.3 | 47.1 | 28.5 | 32.1 |

142

143

#### Detailed Results Across 12 Datasets

144

145

| Model | Avg | Bio. | Earth. | Econ. | Psy. | Rob. | Stack. | Sus. | Leet. | Pony | AoPS | TheoQ. | TheoT. |

146

| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |

147

| **INF-X-Retriever** | **63.4** | **79.8** | **70.9** | **69.9** | **73.3** | **57.7** | **64.3** | **61.9** | **56.1** | **54.5** | **51.9** | **53.1** | **67.9** |

148

| DIVER (v3) | 46.8 | 66.0 | 63.7 | 42.4 | 55.0 | 40.6 | 44.7 | 50.4 | 32.5 | 47.3 | 17.2 | 46.4 | 55.6 |

149

| BGE-Reasoner-0928 | 46.4 | 68.5 | 66.4 | 40.6 | 53.1 | 43.2 | 44.1 | 47.8 | 29.0 | 41.6 | 17.2 | 46.5 | 58.4 |

150

| LATTICE | 42.1 | 64.4 | 62.4 | 45.4 | 57.4 | 47.6 | 37.6 | 46.4 | 19.9 | 34.0 | 12.0 | 30.1 | 47.8 |

151

| ReasonRank | 40.8 | 62.7 | 55.5 | 36.7 | 54.6 | 35.7 | 38.0 | 44.8 | 29.5 | 25.6 | 14.4 | 42.0 | 50.1 |

152

| XDR2 | 40.3 | 63.1 | 55.4 | 38.5 | 52.9 | 37.1 | 38.2 | 44.6 | 21.9 | 35.0 | 15.7 | 34.4 | 46.2 |

153

154

### Long document

155

156

#### Detailed Results Across 8 Datasets

157

158

| Model | Avg | Bio. | Earth. | Econ. | Pony | Psy. | Rob. | Stack. | Sus. |

159

| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |

160

| **INF-X-Retriever** | **54.6** | **73.2** | **59.6** | **69.3** | **12.1** | **74.3** | **55.9** | **27.8** | **64.8** |

161

| inf-retriever-v1-pro | 30.5 | 44.1 | 42.2 | 31.4 | 0.4 | 43.1 | 20.8 | 21.4 | 41.0 |

162

163

164

---

165

166

## 🖊️ Citation

167

168

If you find this model useful, please consider citing our work:

169

170

```bibtex

171

@misc{inf-x-retriever-2025,

172

title = {INF-X-Retriever},

173

    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},

174

year = {2025},

175

url = {https://yaoyichen.github.io/INF-X-Retriever},

176

publisher = {GitHub repository}

177

}

178

```

179

180

---

181

182

## 📬 Contact

183

184

Email: [eason.yyc@inftech.ai](mailto:eason.yyc@inftech.ai)

185

186