File size: 6,080 Bytes
c037a22
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
05dad97
c037a22
 
 
 
 
 
 
 
 
 
c07cfc8
 
c037a22
 
 
 
 
 
 
 
 
 
 
05dad97
bfb0635
c037a22
 
 
 
 
 
 
 
 
 
 
 
 
05dad97
 
 
 
 
 
 
 
 
 
c037a22
 
 
 
 
 
92cae19
c037a22
 
 
95aad98
92cae19
 
c037a22
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6ad2626
c037a22
 
 
 
 
bfb0635
c037a22
 
 
 
bfb0635
 
 
 
 
 
 
 
 
 
 
 
 
04d3b7e
 
7e0ad60
04d3b7e
 
 
 
 
 
 
 
92cae19
4cca456
04d3b7e
 
984f31c
40bbe34
984f31c
 
 
40bbe34
 
 
 
c037a22
40bbe34
 
984f31c
 
40bbe34
 
984f31c
40bbe34
 
6b089a1
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
#
#  Copyright 2024 The InfiniFlow Authors. All Rights Reserved.
#
#  Licensed under the Apache License, Version 2.0 (the "License");
#  you may not use this file except in compliance with the License.
#  You may obtain a copy of the License at
#
#      http://www.apache.org/licenses/LICENSE-2.0
#
#  Unless required by applicable law or agreed to in writing, software
#  distributed under the License is distributed on an "AS IS" BASIS,
#  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#  See the License for the specific language governing permissions and
#  limitations under the License.
#
import re
import  threading
import requests
import torch
from FlagEmbedding import FlagReranker
from huggingface_hub import snapshot_download
import os
from abc import ABC
import numpy as np
from api.utils.file_utils import get_home_cache_dir
from rag.utils import num_tokens_from_string, truncate

def sigmoid(x):
    return 1 / (1 + np.exp(-x))

class Base(ABC):
    def __init__(self, key, model_name):
        pass

    def similarity(self, query: str, texts: list):
        raise NotImplementedError("Please implement encode method!")


class DefaultRerank(Base):
    _model = None
    _model_lock = threading.Lock()

    def __init__(self, key, model_name, **kwargs):
        """
        If you have trouble downloading HuggingFace models, -_^ this might help!!

        For Linux:
        export HF_ENDPOINT=https://hf-mirror.com

        For Windows:
        Good luck
        ^_-

        """
        if not DefaultRerank._model:
            with DefaultRerank._model_lock:
                if not DefaultRerank._model:
                    try:
                        DefaultRerank._model = FlagReranker(os.path.join(get_home_cache_dir(), re.sub(r"^[a-zA-Z]+/", "", model_name)), use_fp16=torch.cuda.is_available())
                    except Exception as e:
                        model_dir = snapshot_download(repo_id= model_name,
                                                      local_dir=os.path.join(get_home_cache_dir(), re.sub(r"^[a-zA-Z]+/", "", model_name)),
                                                      local_dir_use_symlinks=False)
                        DefaultRerank._model = FlagReranker(model_dir, use_fp16=torch.cuda.is_available())
        self._model = DefaultRerank._model

    def similarity(self, query: str, texts: list):
        pairs = [(query,truncate(t, 2048)) for t in texts]
        token_count = 0
        for _, t in pairs:
            token_count += num_tokens_from_string(t)
        batch_size = 4096
        res = []
        for i in range(0, len(pairs), batch_size):
            scores = self._model.compute_score(pairs[i:i + batch_size], max_length=2048)
            scores = sigmoid(np.array(scores)).tolist()
            if isinstance(scores, float): res.append(scores)
            else:  res.extend(scores)
        return np.array(res), token_count


class JinaRerank(Base):
    def __init__(self, key, model_name="jina-reranker-v1-base-en",
                 base_url="https://api.jina.ai/v1/rerank"):
        self.base_url = "https://api.jina.ai/v1/rerank"
        self.headers = {
            "Content-Type": "application/json",
            "Authorization": f"Bearer {key}"
        }
        self.model_name = model_name

    def similarity(self, query: str, texts: list):
        texts = [truncate(t, 8196) for t in texts]
        data = {
            "model": self.model_name,
            "query": query,
            "documents": texts,
            "top_n": len(texts)
        }
        res = requests.post(self.base_url, headers=self.headers, json=data).json()
        return np.array([d["relevance_score"] for d in res["results"]]), res["usage"]["total_tokens"]


class YoudaoRerank(DefaultRerank):
    _model = None
    _model_lock = threading.Lock()

    def __init__(self, key=None, model_name="maidalun1020/bce-reranker-base_v1", **kwargs):
        from BCEmbedding import RerankerModel
        if not YoudaoRerank._model:
            with YoudaoRerank._model_lock:
                if not YoudaoRerank._model:
                    try:
                        print("LOADING BCE...")
                        YoudaoRerank._model = RerankerModel(model_name_or_path=os.path.join(
                            get_home_cache_dir(),
                            re.sub(r"^[a-zA-Z]+/", "", model_name)))
                    except Exception as e:
                        YoudaoRerank._model = RerankerModel(
                            model_name_or_path=model_name.replace(
                                "maidalun1020", "InfiniFlow"))

        self._model = YoudaoRerank._model
    
    def similarity(self, query: str, texts: list):
        pairs = [(query, truncate(t, self._model.max_length)) for t in texts]
        token_count = 0
        for _, t in pairs:
            token_count += num_tokens_from_string(t)
        batch_size = 32
        res = []
        for i in range(0, len(pairs), batch_size):
            scores = self._model.compute_score(pairs[i:i + batch_size], max_length=self._model.max_length)
            scores = sigmoid(np.array(scores)).tolist()
            if isinstance(scores, float): res.append(scores)
            else: res.extend(scores)
        return np.array(res), token_count
    

class XInferenceRerank(Base):
    def __init__(self, key="xxxxxxx", model_name="", base_url=""):
        self.model_name = model_name
        self.base_url = base_url
        self.headers = {
            "Content-Type": "application/json",
            "accept": "application/json"
        }

    def similarity(self, query: str, texts: list):
        data = {
            "model": self.model_name,
            "query": query,
            "return_documents": "true",
            "return_len": "true",
            "documents": texts
        }
        res = requests.post(self.base_url, headers=self.headers, json=data).json()
        return np.array([d["relevance_score"] for d in res["results"]]), res["meta"]["tokens"]["input_tokens"]+res["meta"]["tokens"]["output_tokens"]