|
1
|
|
|
"""TODO""" |
|
2
|
|
|
|
|
3
|
|
|
from __future__ import annotations |
|
4
|
|
|
|
|
5
|
|
|
import os.path |
|
6
|
|
|
from typing import TYPE_CHECKING, Any |
|
7
|
|
|
|
|
8
|
|
|
import numpy as np |
|
9
|
|
|
import tiktoken |
|
10
|
|
|
from openai import AzureOpenAI # Try using huggingface client |
|
11
|
|
|
from qdrant_client import QdrantClient, models |
|
12
|
|
|
|
|
13
|
|
|
from annif.exception import NotInitializedException, NotSupportedException |
|
14
|
|
|
from annif.suggestion import vector_to_suggestions |
|
15
|
|
|
|
|
16
|
|
|
from . import backend |
|
17
|
|
|
|
|
18
|
|
|
if TYPE_CHECKING: |
|
19
|
|
|
from collections.abc import Iterator |
|
20
|
|
|
|
|
21
|
|
|
from annif.corpus.document import DocumentCorpus |
|
22
|
|
|
|
|
23
|
|
|
|
|
24
|
|
|
class Vectorizer: |
|
25
|
|
|
def __init__(self, endpoint, model): |
|
26
|
|
|
self.model = model |
|
27
|
|
|
self.client = AzureOpenAI( # TODO Try AsyncAzureOpenAI( |
|
28
|
|
|
azure_endpoint=endpoint, |
|
29
|
|
|
api_key=os.getenv("AZURE_OPENAI_KEY"), |
|
30
|
|
|
api_version="2024-02-15-preview", |
|
31
|
|
|
) |
|
32
|
|
|
|
|
33
|
|
|
def transform(self, text): |
|
34
|
|
|
response = self.client.embeddings.create( |
|
35
|
|
|
input=text, |
|
36
|
|
|
model=self.model, |
|
37
|
|
|
# dimensions=dimensions, # TODO Try with reduced dimensions |
|
38
|
|
|
) |
|
39
|
|
|
return response.data[0].embedding |
|
40
|
|
|
|
|
41
|
|
|
|
|
42
|
|
|
class EmbeddingsBackend(backend.AnnifBackend): |
|
43
|
|
|
"""TODO xxx cector space similarity based backend for Annif""" |
|
44
|
|
|
|
|
45
|
|
|
name = "embeddings" |
|
46
|
|
|
is_trained = True |
|
47
|
|
|
|
|
48
|
|
|
# defaults for uninitialized instances |
|
49
|
|
|
_index = None |
|
50
|
|
|
|
|
51
|
|
|
DB_FILE = "qdrant-db" |
|
52
|
|
|
COLLECTION_NAME = "index-collection" |
|
53
|
|
|
BASE_MODEL = "text-embedding-3-large" |
|
54
|
|
|
VECTOR_DIMENSIONS = 3072 # For text-embedding-3-large |
|
55
|
|
|
MAX_TOKENS = 8192 # For text-embedding-3-large |
|
56
|
|
|
|
|
57
|
|
|
encoding = tiktoken.encoding_for_model(BASE_MODEL) |
|
58
|
|
|
|
|
59
|
|
|
def _initialize_index(self) -> None: |
|
60
|
|
|
if self._index is None: |
|
61
|
|
|
path = os.path.join(self.datadir, self.DB_FILE) |
|
62
|
|
|
self.debug("loading similarity index from {}".format(path)) |
|
63
|
|
|
if os.path.exists(path): |
|
64
|
|
|
self.qdclient = QdrantClient(path=path) |
|
65
|
|
|
else: |
|
66
|
|
|
raise NotInitializedException( |
|
67
|
|
|
"similarity index {} not found".format(path), |
|
68
|
|
|
backend_id=self.backend_id, |
|
69
|
|
|
) |
|
70
|
|
|
|
|
71
|
|
|
def initialize( |
|
72
|
|
|
self, |
|
73
|
|
|
) -> None: |
|
74
|
|
|
self.vectorizer = Vectorizer(self.params["endpoint"], self.params["model"]) |
|
75
|
|
|
self._initialize_index() |
|
76
|
|
|
|
|
77
|
|
|
def _create_index(self, corpus) -> None: |
|
78
|
|
|
self.vectorizer = Vectorizer(self.params["endpoint"], self.params["model"]) |
|
79
|
|
|
self.info("creating similarity index") |
|
80
|
|
|
path = os.path.join(self.datadir, self.DB_FILE) |
|
81
|
|
|
|
|
82
|
|
|
self.qdclient = QdrantClient(path=path) |
|
83
|
|
|
self.qdclient.recreate_collection( |
|
84
|
|
|
collection_name=self.COLLECTION_NAME, |
|
85
|
|
|
vectors_config=models.VectorParams( |
|
86
|
|
|
size=self.VECTOR_DIMENSIONS, |
|
87
|
|
|
distance=models.Distance.COSINE, |
|
88
|
|
|
), |
|
89
|
|
|
) |
|
90
|
|
|
|
|
91
|
|
|
veccorpus = ( |
|
92
|
|
|
( |
|
93
|
|
|
doc.subject_set, |
|
94
|
|
|
self.vectorizer.transform(self._truncate_text(" ".join(doc.text))), |
|
95
|
|
|
) |
|
96
|
|
|
for doc in corpus.documents |
|
97
|
|
|
) |
|
98
|
|
|
|
|
99
|
|
|
subject_sets, vectors = zip(*veccorpus) |
|
100
|
|
|
payloads = [{"subjects": [sid for sid in ss]} for ss in subject_sets] |
|
101
|
|
|
ids = list(range(len(vectors))) |
|
102
|
|
|
self.qdclient.upsert( |
|
103
|
|
|
collection_name=self.COLLECTION_NAME, |
|
104
|
|
|
points=models.Batch( |
|
105
|
|
|
ids=ids, |
|
106
|
|
|
vectors=vectors, |
|
107
|
|
|
payloads=payloads, |
|
108
|
|
|
), |
|
109
|
|
|
) |
|
110
|
|
|
print(self.qdclient.get_collection(collection_name=self.COLLECTION_NAME)) |
|
111
|
|
|
|
|
112
|
|
|
def _truncate_text(self, text): |
|
113
|
|
|
"""truncate text so it contains at most MAX_TOKENS according to the OpenAI |
|
114
|
|
|
tokenizer""" |
|
115
|
|
|
tokens = self.encoding.encode(text) |
|
116
|
|
|
return self.encoding.decode(tokens[: self.MAX_TOKENS]) |
|
117
|
|
|
|
|
118
|
|
|
def _train( |
|
119
|
|
|
self, |
|
120
|
|
|
corpus: DocumentCorpus, |
|
121
|
|
|
params: dict[str, Any], |
|
122
|
|
|
jobs: int = 0, |
|
123
|
|
|
) -> None: |
|
124
|
|
|
if corpus == "cached": |
|
125
|
|
|
raise NotSupportedException( |
|
126
|
|
|
"Training embeddings project from cached data not supported." |
|
127
|
|
|
) |
|
128
|
|
|
if corpus.is_empty(): |
|
129
|
|
|
raise NotSupportedException( |
|
130
|
|
|
"Cannot train embeddings project with no documents" |
|
131
|
|
|
) |
|
132
|
|
|
self.info("transforming subject corpus") |
|
133
|
|
|
self._create_index(corpus) |
|
134
|
|
|
|
|
135
|
|
|
def _suggest(self, text: str, params: dict[str, Any]) -> Iterator: |
|
136
|
|
|
self.debug( |
|
137
|
|
|
'Suggesting subjects for text "{}..." (len={})'.format(text[:20], len(text)) |
|
138
|
|
|
) |
|
139
|
|
|
truncated_text = self._truncate_text(" ".join(text)) |
|
140
|
|
|
vector = self.vectorizer.transform(truncated_text) |
|
141
|
|
|
# print(vector[:5]) |
|
142
|
|
|
info = self.qdclient.get_collection(collection_name=self.COLLECTION_NAME) |
|
143
|
|
|
self.debug(f"Collection info: {info}") |
|
144
|
|
|
results = self._search(vector, params) |
|
145
|
|
|
# print(results) |
|
146
|
|
|
return self._prediction_to_result(results, params) |
|
147
|
|
|
|
|
148
|
|
|
def _search(self, vector, params): |
|
149
|
|
|
result = self.qdclient.search( |
|
150
|
|
|
collection_name=self.COLLECTION_NAME, |
|
151
|
|
|
query_vector=vector, |
|
152
|
|
|
# score_threshold=1.0, # TODO parameterize this |
|
153
|
|
|
limit=int(params["limit"]), |
|
154
|
|
|
# search_params=models.SearchParams(hnsw_ef=128, exact=False), # TODO This |
|
155
|
|
|
) |
|
156
|
|
|
return [(sp.payload["subjects"], sp.score) for sp in result] |
|
157
|
|
|
|
|
158
|
|
|
def _combine_search_results(self, results): |
|
159
|
|
|
combined = [] |
|
160
|
|
|
for res in results: |
|
161
|
|
|
sids, weight = res[0], res[1] |
|
162
|
|
|
combined.extend([sid * weight for sid in sids]) |
|
163
|
|
|
return combined |
|
164
|
|
|
|
|
165
|
|
|
# From backend/mllm.py |
|
166
|
|
|
def _prediction_to_result( |
|
167
|
|
|
self, |
|
168
|
|
|
results, |
|
169
|
|
|
params, |
|
170
|
|
|
) -> Iterator: |
|
171
|
|
|
vector = np.zeros(len(self.project.subjects), dtype=np.float32) |
|
172
|
|
|
for subject_ids, score in results: |
|
173
|
|
|
for sid in subject_ids: |
|
174
|
|
|
vector[sid] += score |
|
175
|
|
|
return vector_to_suggestions(vector, int(params["limit"])) |
|
176
|
|
|
|