Spaces:
Running
Running
revert changes
Browse files- __pycache__/app.cpython-311.pyc +0 -0
- __pycache__/neural_searcher.cpython-311.pyc +0 -0
- app.py +3 -3
- ner.py +97 -0
- neural_searcher.py +19 -17
__pycache__/app.cpython-311.pyc
ADDED
Binary file (2.05 kB). View file
|
|
__pycache__/neural_searcher.cpython-311.pyc
ADDED
Binary file (2.84 kB). View file
|
|
app.py
CHANGED
@@ -13,11 +13,11 @@ app = FastAPI()
|
|
13 |
|
14 |
neural_searcher = NeuralSearcher(collection_name=os.getenv('COLLECTION_NAME'))
|
15 |
|
16 |
-
REQUEST_TIMEOUT_ERROR =
|
17 |
|
18 |
@app.get("/api/search")
|
19 |
-
def search(q: str):
|
20 |
-
data = neural_searcher.search(text=q)
|
21 |
return data
|
22 |
|
23 |
|
|
|
13 |
|
14 |
neural_searcher = NeuralSearcher(collection_name=os.getenv('COLLECTION_NAME'))
|
15 |
|
16 |
+
REQUEST_TIMEOUT_ERROR = 30
|
17 |
|
18 |
@app.get("/api/search")
|
19 |
+
async def search(q: str):
|
20 |
+
data = await neural_searcher.search(text=q)
|
21 |
return data
|
22 |
|
23 |
|
ner.py
ADDED
@@ -0,0 +1,97 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import AutoModelForTokenClassification, AutoTokenizer
|
2 |
+
import torch
|
3 |
+
|
4 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
5 |
+
tokenizer = AutoTokenizer.from_pretrained("kalusev/NER4Legal_SRB", use_auth_token=True)
|
6 |
+
model = AutoModelForTokenClassification.from_pretrained("kalusev/NER4Legal_SRB", use_auth_token=True).to(device)
|
7 |
+
|
8 |
+
id_to_label = {
|
9 |
+
0: 'O',
|
10 |
+
1: 'B-COURT',
|
11 |
+
2: 'B-DATE',
|
12 |
+
3: 'B-DECISION',
|
13 |
+
4: 'B-LAW',
|
14 |
+
5: 'B-MONEY',
|
15 |
+
6: 'B-OFFICIAL GAZZETE',
|
16 |
+
7: 'B-PERSON',
|
17 |
+
8: 'B-REFERENCE',
|
18 |
+
9: 'I-COURT',
|
19 |
+
10: 'I-LAW',
|
20 |
+
11: 'I-MONEY',
|
21 |
+
12: 'I-OFFICIAL GAZZETE',
|
22 |
+
13: 'I-PERSON',
|
23 |
+
14: 'I-REFERENCE'
|
24 |
+
}
|
25 |
+
|
26 |
+
def perform_ner(text):
|
27 |
+
try:
|
28 |
+
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True).to(device)
|
29 |
+
with torch.no_grad():
|
30 |
+
outputs = model(**inputs)
|
31 |
+
logits = outputs.logits
|
32 |
+
predictions = torch.argmax(logits, dim=2).squeeze().tolist()
|
33 |
+
|
34 |
+
except RuntimeError as e:
|
35 |
+
if "CUDA out of memory" in str(e):
|
36 |
+
print("Switching to CPU due to memory constraints.")
|
37 |
+
inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True)
|
38 |
+
with torch.no_grad():
|
39 |
+
outputs = model.cpu()(**inputs) # Run model on CPU
|
40 |
+
logits = outputs.logits
|
41 |
+
predictions = torch.argmax(logits, dim=2).squeeze().tolist()
|
42 |
+
else:
|
43 |
+
raise e
|
44 |
+
tokens = tokenizer.convert_ids_to_tokens(inputs["input_ids"].squeeze())
|
45 |
+
labels = [id_to_label[pred] for pred in predictions]
|
46 |
+
|
47 |
+
results = [
|
48 |
+
(token, label)
|
49 |
+
for token, label in zip(tokens, labels)
|
50 |
+
if token not in tokenizer.all_special_tokens
|
51 |
+
]
|
52 |
+
return results
|
53 |
+
|
54 |
+
text = """1
|
55 |
+
osnovni sud u bijelom polju je vrsio veliku nuzdu
|
56 |
+
"""
|
57 |
+
|
58 |
+
def merge_entities(token_label_pairs):
|
59 |
+
merged_words, merged_labels = [], []
|
60 |
+
current_word, current_label = "", None
|
61 |
+
|
62 |
+
for token, label in token_label_pairs:
|
63 |
+
if token.startswith("##"):
|
64 |
+
current_word += token[2:]
|
65 |
+
else:
|
66 |
+
if current_word:
|
67 |
+
merged_words.append(current_word)
|
68 |
+
merged_labels.append(current_label)
|
69 |
+
|
70 |
+
current_word, current_label = token, label
|
71 |
+
|
72 |
+
if current_word:
|
73 |
+
merged_words.append(current_word)
|
74 |
+
merged_labels.append(current_label)
|
75 |
+
|
76 |
+
final_words, final_labels = [], []
|
77 |
+
|
78 |
+
for i, (word, label) in enumerate(zip(merged_words, merged_labels)):
|
79 |
+
if final_labels and (
|
80 |
+
label == final_labels[-1] or
|
81 |
+
(label.startswith("I-") and final_labels[-1].endswith(label[2:])) or
|
82 |
+
(label.startswith("B-") and final_labels[-1].endswith(label[2:]))
|
83 |
+
):
|
84 |
+
|
85 |
+
final_words[-1] += " " + word
|
86 |
+
else:
|
87 |
+
final_words.append(word)
|
88 |
+
final_labels.append(label)
|
89 |
+
|
90 |
+
return final_words, final_labels
|
91 |
+
|
92 |
+
results = perform_ner(text)
|
93 |
+
|
94 |
+
words,labels = merge_entities(results)
|
95 |
+
|
96 |
+
for i,b in zip(words,labels):
|
97 |
+
print(i + " ### " + b)
|
neural_searcher.py
CHANGED
@@ -5,45 +5,47 @@ from sentence_transformers import SentenceTransformer
|
|
5 |
import os
|
6 |
|
7 |
class NeuralSearcher:
|
8 |
-
|
9 |
def __init__(self, collection_name):
|
10 |
self.collection_name = collection_name
|
11 |
self.dense_model = SentenceTransformer(os.getenv('DENSE_MODEL'),device="cpu")
|
12 |
self.sparse_model = SparseTextEmbedding(os.getenv('SPARSE_MODEL'))
|
13 |
self.late_interaction_model = LateInteractionTextEmbedding(os.getenv('LATE_INTERACTION_MODEL'))
|
14 |
-
self.qdrant_client = QdrantClient(os.getenv('QDRANT_URL'),api_key=os.getenv('QDRANT_API_KEY')
|
15 |
|
16 |
-
def search(self, text: str):
|
17 |
-
|
18 |
dense_query = self.dense_model.encode(text).tolist()
|
19 |
sparse_query = next(self.sparse_model.query_embed(text))
|
20 |
-
late_query = next(self.late_interaction_model.query_embed(text))
|
21 |
|
22 |
prefetch = [
|
23 |
models.Prefetch(
|
24 |
query=dense_query,
|
25 |
using=os.getenv('DENSE_MODEL'),
|
26 |
-
limit=
|
27 |
),
|
28 |
models.Prefetch(
|
29 |
query=models.SparseVector(**sparse_query.as_object()),
|
30 |
using=os.getenv('SPARSE_MODEL'),
|
31 |
-
limit=
|
32 |
)
|
33 |
]
|
34 |
|
35 |
-
search_result = self.qdrant_client.
|
36 |
collection_name= self.collection_name,
|
37 |
-
group_by="dbid",
|
38 |
prefetch=prefetch,
|
39 |
-
|
40 |
-
|
41 |
-
|
|
|
42 |
with_payload=True,
|
43 |
-
score_threshold=0.8,
|
44 |
limit = 10
|
45 |
-
).
|
|
|
|
|
|
|
|
|
|
|
46 |
|
47 |
-
|
48 |
-
print(group)
|
49 |
-
return search_result
|
|
|
5 |
import os
|
6 |
|
7 |
class NeuralSearcher:
|
8 |
+
|
9 |
def __init__(self, collection_name):
|
10 |
self.collection_name = collection_name
|
11 |
self.dense_model = SentenceTransformer(os.getenv('DENSE_MODEL'),device="cpu")
|
12 |
self.sparse_model = SparseTextEmbedding(os.getenv('SPARSE_MODEL'))
|
13 |
self.late_interaction_model = LateInteractionTextEmbedding(os.getenv('LATE_INTERACTION_MODEL'))
|
14 |
+
self.qdrant_client = QdrantClient(os.getenv('QDRANT_URL'),api_key=os.getenv('QDRANT_API_KEY'))
|
15 |
|
16 |
+
async def search(self, text: str):
|
17 |
+
|
18 |
dense_query = self.dense_model.encode(text).tolist()
|
19 |
sparse_query = next(self.sparse_model.query_embed(text))
|
20 |
+
# late_query = next(self.late_interaction_model.query_embed(text))
|
21 |
|
22 |
prefetch = [
|
23 |
models.Prefetch(
|
24 |
query=dense_query,
|
25 |
using=os.getenv('DENSE_MODEL'),
|
26 |
+
limit=200
|
27 |
),
|
28 |
models.Prefetch(
|
29 |
query=models.SparseVector(**sparse_query.as_object()),
|
30 |
using=os.getenv('SPARSE_MODEL'),
|
31 |
+
limit=200
|
32 |
)
|
33 |
]
|
34 |
|
35 |
+
search_result = self.qdrant_client.query_points(
|
36 |
collection_name= self.collection_name,
|
|
|
37 |
prefetch=prefetch,
|
38 |
+
query=models.FusionQuery(
|
39 |
+
fusion=models.Fusion.RRF,
|
40 |
+
),
|
41 |
+
# using=os.getenv('LATE_INTERACTION_MODEL'),
|
42 |
with_payload=True,
|
|
|
43 |
limit = 10
|
44 |
+
).points
|
45 |
+
|
46 |
+
data = []
|
47 |
+
|
48 |
+
for hit in search_result:
|
49 |
+
data.append(hit.payload)
|
50 |
|
51 |
+
return data
|
|
|
|