|
1
|
|
|
"""Language model based ensemble backend that combines results from multiple |
|
2
|
|
|
projects.""" |
|
3
|
|
|
|
|
4
|
|
|
from __future__ import annotations |
|
5
|
|
|
|
|
6
|
|
|
import json |
|
7
|
|
|
import os |
|
8
|
|
|
from typing import TYPE_CHECKING, Any |
|
9
|
|
|
|
|
10
|
|
|
import tiktoken |
|
11
|
|
|
from openai import AzureOpenAI, BadRequestError |
|
12
|
|
|
|
|
13
|
|
|
import annif.eval |
|
14
|
|
|
import annif.parallel |
|
15
|
|
|
import annif.util |
|
16
|
|
|
from annif.exception import NotSupportedException |
|
17
|
|
|
from annif.suggestion import SubjectSuggestion, SuggestionBatch |
|
18
|
|
|
|
|
19
|
|
|
from . import backend, ensemble |
|
20
|
|
|
|
|
21
|
|
|
# from openai import AsyncAzureOpenAI |
|
22
|
|
|
|
|
23
|
|
|
|
|
24
|
|
|
if TYPE_CHECKING: |
|
25
|
|
|
from annif.corpus.document import DocumentCorpus |
|
26
|
|
|
|
|
27
|
|
|
|
|
28
|
|
|
class BaseLLMBackend(backend.AnnifBackend): |
|
29
|
|
|
# """Base class for TODO backends""" |
|
30
|
|
|
|
|
31
|
|
|
def initialize(self, parallel: bool = False) -> None: |
|
32
|
|
|
# initialize all the source projects |
|
33
|
|
|
params = self._get_backend_params(None) |
|
34
|
|
|
for project_id, _ in annif.util.parse_sources(params["sources"]): |
|
35
|
|
|
project = self.project.registry.get_project(project_id) |
|
36
|
|
|
project.initialize(parallel) |
|
37
|
|
|
|
|
38
|
|
|
# self.client = AsyncAzureOpenAI( |
|
39
|
|
|
self.client = AzureOpenAI( |
|
40
|
|
|
azure_endpoint=params["endpoint"], |
|
41
|
|
|
api_key=os.getenv("AZURE_OPENAI_KEY"), |
|
42
|
|
|
api_version="2024-02-15-preview", |
|
43
|
|
|
) |
|
44
|
|
|
|
|
45
|
|
|
|
|
46
|
|
|
class LLMEnsembleBackend(BaseLLMBackend, ensemble.BaseEnsembleBackend): |
|
47
|
|
|
# """TODO backend that combines results from multiple projects""" |
|
48
|
|
|
|
|
49
|
|
|
name = "llm_ensemble" |
|
50
|
|
|
|
|
51
|
|
|
system_prompt = """ |
|
52
|
|
|
You will be given text and a list of keywords to describe it. Your task is to |
|
53
|
|
|
score the keywords with a value between 0.0 and 1.0. The score value |
|
54
|
|
|
should depend on how well the keyword represents the text: a perfect |
|
55
|
|
|
keyword should have score 1.0 and completely unrelated keyword score |
|
56
|
|
|
0.0. You must output JSON with keywords as field names and add their scores |
|
57
|
|
|
as field values. |
|
58
|
|
|
There must be the same number of objects in the JSON as there are lines in the |
|
59
|
|
|
intput keyword list; do not skip scoring any keywords. |
|
60
|
|
|
""" |
|
61
|
|
|
# Give zero or very low score to the keywords that do not describe the text. |
|
62
|
|
|
|
|
63
|
|
|
def _train(self, corpus: DocumentCorpus, params: dict[str, Any], jobs: int = 0): |
|
64
|
|
|
raise NotSupportedException("Training LM ensemble backend is not possible.") |
|
65
|
|
|
|
|
66
|
|
|
def _suggest_batch( |
|
67
|
|
|
self, texts: list[str], params: dict[str, Any] |
|
68
|
|
|
) -> SuggestionBatch: |
|
69
|
|
|
sources = annif.util.parse_sources(params["sources"]) |
|
70
|
|
|
batch_by_source = self._suggest_with_sources(texts, sources) |
|
71
|
|
|
# Just add LLM scores to the base suggestions? |
|
72
|
|
|
return self._merge_source_batches(texts, batch_by_source, sources, params) |
|
73
|
|
|
|
|
74
|
|
|
def _merge_source_batches( |
|
75
|
|
|
self, |
|
76
|
|
|
texts: list[str], |
|
77
|
|
|
batch_by_source: dict[str, SuggestionBatch], |
|
78
|
|
|
sources: list[tuple[str, float]], |
|
79
|
|
|
params: dict[str, Any], |
|
80
|
|
|
) -> SuggestionBatch: |
|
81
|
|
|
model = params["model"] |
|
82
|
|
|
# llm_scores_weight = float(params["llm_weight"]) |
|
83
|
|
|
encoding = tiktoken.encoding_for_model(model.rsplit("-", 1)[0]) |
|
84
|
|
|
|
|
85
|
|
|
batches = [batch_by_source[project_id] for project_id, _ in sources] |
|
86
|
|
|
weights = [weight for _, weight in sources] |
|
87
|
|
|
avg_suggestion_batch = SuggestionBatch.from_averaged(batches, weights).filter( |
|
88
|
|
|
limit=int(params["limit"]) # TODO Increase limit |
|
89
|
|
|
) |
|
90
|
|
|
|
|
91
|
|
|
labels_batch = [] |
|
92
|
|
|
for suggestionresult in avg_suggestion_batch: |
|
93
|
|
|
# print(suggestionresult) |
|
94
|
|
|
# for suggestion in suggestionresult: |
|
95
|
|
|
# # print(suggestion) |
|
96
|
|
|
# print(self.project.subjects[suggestion.subject_id].labels["en"]) |
|
97
|
|
|
labels_batch.append( |
|
98
|
|
|
[ |
|
99
|
|
|
self.project.subjects[s.subject_id].labels[ |
|
100
|
|
|
"en" |
|
101
|
|
|
] # TODO: make language selectable |
|
102
|
|
|
for s in suggestionresult |
|
103
|
|
|
] |
|
104
|
|
|
) |
|
105
|
|
|
# print(labels_batch) |
|
106
|
|
|
|
|
107
|
|
|
llm_batch_suggestions = [] |
|
108
|
|
|
for text, labels in zip(texts, labels_batch): |
|
109
|
|
|
print(text) |
|
110
|
|
|
print(labels) |
|
111
|
|
|
|
|
112
|
|
|
prompt = "Here are the keywords:\n" + "\n".join(labels) + "\n" * 3 |
|
113
|
|
|
text = self._truncate_text(text, encoding) |
|
114
|
|
|
prompt += "Here is the text:\n" + text + "\n" |
|
115
|
|
|
|
|
116
|
|
|
response = self._call_llm(prompt, model) |
|
117
|
|
|
print(response) |
|
118
|
|
|
try: |
|
119
|
|
|
llm_result = json.loads(response) |
|
120
|
|
|
except (TypeError, json.decoder.JSONDecodeError) as err: |
|
121
|
|
|
print(err) |
|
122
|
|
|
llm_result = None |
|
123
|
|
|
continue # TODO: handle this error |
|
124
|
|
|
suggestions = [] |
|
125
|
|
|
for label, score in llm_result.items(): |
|
126
|
|
|
print(label, score) |
|
127
|
|
|
subj_id = self.project.subjects.by_label( |
|
128
|
|
|
label, "en" |
|
129
|
|
|
) # TODO: make language selectable |
|
130
|
|
|
# print(subj_id) |
|
131
|
|
|
suggestions.append(SubjectSuggestion(subject_id=subj_id, score=score)) |
|
132
|
|
|
|
|
133
|
|
|
llm_batch_suggestions.append(suggestions) |
|
134
|
|
|
|
|
135
|
|
|
return SuggestionBatch.from_sequence( |
|
136
|
|
|
llm_batch_suggestions, self.project.subjects |
|
137
|
|
|
) |
|
138
|
|
|
|
|
139
|
|
|
def _truncate_text(self, text, encoding): |
|
140
|
|
|
"""truncate text so it contains at most MAX_PROMPT_TOKENS according to the |
|
141
|
|
|
OpenAI tokenizer""" |
|
142
|
|
|
|
|
143
|
|
|
MAX_PROMPT_TOKENS = 14000 |
|
144
|
|
|
tokens = encoding.encode(text) |
|
145
|
|
|
return encoding.decode(tokens[:MAX_PROMPT_TOKENS]) |
|
146
|
|
|
|
|
147
|
|
|
def _call_llm(self, prompt: str, model: str): |
|
148
|
|
|
messages = [ |
|
149
|
|
|
{"role": "system", "content": self.system_prompt}, |
|
150
|
|
|
{"role": "user", "content": prompt}, |
|
151
|
|
|
] |
|
152
|
|
|
# print(prompt) #[-10000:]) |
|
153
|
|
|
try: |
|
154
|
|
|
completion = self.client.chat.completions.create( |
|
155
|
|
|
model=model, |
|
156
|
|
|
messages=messages, |
|
157
|
|
|
temperature=0.0, |
|
158
|
|
|
seed=0, |
|
159
|
|
|
max_tokens=1800, |
|
160
|
|
|
top_p=0.95, |
|
161
|
|
|
frequency_penalty=0, |
|
162
|
|
|
presence_penalty=0, |
|
163
|
|
|
stop=None, |
|
164
|
|
|
response_format={"type": "json_object"}, |
|
165
|
|
|
) |
|
166
|
|
|
|
|
167
|
|
|
completion = completion.choices[0].message.content |
|
168
|
|
|
return completion |
|
169
|
|
|
except BadRequestError as err: # openai.RateLimitError |
|
170
|
|
|
print(err) |
|
171
|
|
|
return "{}" |
|
172
|
|
|
|
NatLibFi /
Annif
