|
1
|
|
|
"""TODO""" |
|
2
|
|
|
|
|
3
|
|
|
from __future__ import annotations |
|
4
|
|
|
|
|
5
|
|
|
import os.path |
|
6
|
|
|
import tempfile |
|
7
|
|
|
from typing import TYPE_CHECKING, Any |
|
8
|
|
|
|
|
9
|
|
|
from openai import AzureOpenAI |
|
10
|
|
|
|
|
11
|
|
|
import annif.util |
|
12
|
|
|
from annif.exception import NotInitializedException, NotSupportedException |
|
13
|
|
|
from annif.suggestion import vector_to_suggestions |
|
14
|
|
|
|
|
15
|
|
|
from . import backend |
|
16
|
|
|
|
|
17
|
|
|
if TYPE_CHECKING: |
|
18
|
|
|
from collections.abc import Iterator |
|
19
|
|
|
|
|
20
|
|
|
from scipy.sparse._csr import csr_matrix |
|
21
|
|
|
|
|
22
|
|
|
from annif.corpus.document import DocumentCorpus |
|
23
|
|
|
|
|
24
|
|
|
|
|
25
|
|
|
class Vectorizer: |
|
26
|
|
|
def __init__(self, endpoint, model): |
|
27
|
|
|
self.model = model |
|
28
|
|
|
self.client = AzureOpenAI( # TODO Try AsyncAzureOpenAI( |
|
29
|
|
|
azure_endpoint=endpoint, |
|
30
|
|
|
api_key=os.getenv("AZURE_OPENAI_KEY"), |
|
31
|
|
|
api_version="2024-02-15-preview", |
|
32
|
|
|
) |
|
33
|
|
|
|
|
34
|
|
|
def vectorize(self, text): |
|
35
|
|
|
response = self.client.embeddings.create( |
|
36
|
|
|
input=text, |
|
37
|
|
|
model=self.model, |
|
38
|
|
|
# dimensions=dimensions, # TODO Try with reduced dimensions |
|
39
|
|
|
) |
|
40
|
|
|
return response.data[0].embedding |
|
41
|
|
|
|
|
42
|
|
|
|
|
43
|
|
View Code Duplication |
class EmbeddingsBackend(backend.AnnifBackend): |
|
|
|
|
|
|
44
|
|
|
"""TODO xxx cector space similarity based backend for Annif""" |
|
45
|
|
|
|
|
46
|
|
|
name = "embeddings" |
|
47
|
|
|
|
|
48
|
|
|
# defaults for uninitialized instances |
|
49
|
|
|
_index = None |
|
50
|
|
|
|
|
51
|
|
|
# INDEX_FILE = "tfidf-index" |
|
52
|
|
|
|
|
53
|
|
|
def _generate_subjects_from_documents( |
|
54
|
|
|
self, corpus: DocumentCorpus |
|
55
|
|
|
) -> Iterator[str]: |
|
56
|
|
|
with tempfile.TemporaryDirectory() as tempdir: |
|
57
|
|
|
subject_buffer = {} |
|
58
|
|
|
for subject_id in range(len(self.project.subjects)): |
|
59
|
|
|
subject_buffer[subject_id] = SubjectBuffer(tempdir, subject_id) |
|
60
|
|
|
|
|
61
|
|
|
for doc in corpus.documents: |
|
62
|
|
|
tokens = self.project.analyzer.tokenize_words(doc.text) |
|
63
|
|
|
for subject_id in doc.subject_set: |
|
64
|
|
|
subject_buffer[subject_id].write(" ".join(tokens)) |
|
65
|
|
|
|
|
66
|
|
|
for sid in range(len(self.project.subjects)): |
|
67
|
|
|
yield subject_buffer[sid].read() |
|
68
|
|
|
|
|
69
|
|
|
def _initialize_index(self) -> None: |
|
70
|
|
|
if self._index is None: |
|
71
|
|
|
path = os.path.join(self.datadir, self.INDEX_FILE) |
|
72
|
|
|
self.debug("loading similarity index from {}".format(path)) |
|
73
|
|
|
if os.path.exists(path): |
|
74
|
|
|
self._index = gensim.similarities.SparseMatrixSimilarity.load(path) |
|
75
|
|
|
else: |
|
76
|
|
|
raise NotInitializedException( |
|
77
|
|
|
"similarity index {} not found".format(path), |
|
78
|
|
|
backend_id=self.backend_id, |
|
79
|
|
|
) |
|
80
|
|
|
|
|
81
|
|
|
def initialize(self, parallel: bool = False) -> None: |
|
82
|
|
|
self.vectorizer = Vectorizer() |
|
83
|
|
|
self._initialize_index() |
|
84
|
|
|
|
|
85
|
|
|
def _create_index(self, veccorpus: csr_matrix) -> None: |
|
86
|
|
|
self.info("creating similarity index") |
|
87
|
|
|
gscorpus = Sparse2Corpus(veccorpus, documents_columns=False) |
|
88
|
|
|
self._index = gensim.similarities.SparseMatrixSimilarity( |
|
89
|
|
|
gscorpus, num_features=len(self.vectorizer.vocabulary_) |
|
90
|
|
|
) |
|
91
|
|
|
annif.util.atomic_save(self._index, self.datadir, self.INDEX_FILE) |
|
92
|
|
|
|
|
93
|
|
|
def _train( |
|
94
|
|
|
self, |
|
95
|
|
|
corpus: DocumentCorpus, |
|
96
|
|
|
params: dict[str, Any], |
|
97
|
|
|
jobs: int = 0, |
|
98
|
|
|
) -> None: |
|
99
|
|
|
if corpus == "cached": |
|
100
|
|
|
raise NotSupportedException( |
|
101
|
|
|
"Training tfidf project from cached data not supported." |
|
102
|
|
|
) |
|
103
|
|
|
if corpus.is_empty(): |
|
104
|
|
|
raise NotSupportedException("Cannot train tfidf project with no documents") |
|
105
|
|
|
self.info("transforming subject corpus") |
|
106
|
|
|
subjects = self._generate_subjects_from_documents(corpus) |
|
107
|
|
|
veccorpus = self.create_vectorizer(subjects) |
|
108
|
|
|
self._create_index(veccorpus) |
|
109
|
|
|
|
|
110
|
|
|
def _suggest(self, text: str, params: dict[str, Any]) -> Iterator: |
|
111
|
|
|
self.debug( |
|
112
|
|
|
'Suggesting subjects for text "{}..." (len={})'.format(text[:20], len(text)) |
|
113
|
|
|
) |
|
114
|
|
|
tokens = self.project.analyzer.tokenize_words(text) |
|
115
|
|
|
vectors = self.vectorizer.transform([" ".join(tokens)]) |
|
116
|
|
|
return vector_to_suggestions(self._index[vectors[0]], int(params["limit"])) |
|
117
|
|
|
|
NatLibFi /
Annif
Loading history...