Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -1,23 +1,68 @@
|
|
1 |
import gradio as gr
|
2 |
from sentence_transformers import SentenceTransformer
|
|
|
|
|
|
|
3 |
|
4 |
-
#
|
5 |
-
|
6 |
|
7 |
-
|
8 |
-
def embed(text: str):
|
9 |
if not text.strip():
|
10 |
return {"error": "Input text is empty."}
|
11 |
-
|
12 |
-
return {"
|
13 |
-
|
14 |
-
#
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
)
|
22 |
-
|
23 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
import gradio as gr
|
2 |
from sentence_transformers import SentenceTransformer
|
3 |
+
from rank_bm25 import BM25Okapi
|
4 |
+
from transformers import AutoTokenizer, AutoModel
|
5 |
+
import torch
|
6 |
|
7 |
+
# 1. Dense embedding model (HF bi-encoder)
|
8 |
+
dense_model = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
|
9 |
|
10 |
+
def embed_dense(text: str):
|
|
|
11 |
if not text.strip():
|
12 |
return {"error": "Input text is empty."}
|
13 |
+
emb = dense_model.encode([text])[0]
|
14 |
+
return {"dense_embedding": emb.tolist()}
|
15 |
+
|
16 |
+
# 2. Sparse embedding model (BM25)
|
17 |
+
# Uses rank_bm25 to compute term weights
|
18 |
+
|
19 |
+
def embed_sparse(text: str):
|
20 |
+
if not text.strip():
|
21 |
+
return {"error": "Input text is empty."}
|
22 |
+
tokens = text.split()
|
23 |
+
bm25 = BM25Okapi([tokens])
|
24 |
+
scores = bm25.get_scores(tokens)
|
25 |
+
# Map each term to its BM25 weight
|
26 |
+
term_weights = {tok: float(score) for tok, score in zip(tokens, scores)}
|
27 |
+
return {"sparse_embedding": term_weights}
|
28 |
+
|
29 |
+
# 3. Late-interaction embedding model (ColBERT)
|
30 |
+
colbert_tokenizer = AutoTokenizer.from_pretrained('colbert-ir/colbertv2.0', use_fast=True)
|
31 |
+
colbert_model = AutoModel.from_pretrained('colbert-ir/colbertv2.0')
|
32 |
+
|
33 |
+
# Freeze model parameters for inference speed
|
34 |
+
for param in colbert_model.parameters():
|
35 |
+
param.requires_grad = False
|
36 |
+
|
37 |
+
|
38 |
+
def embed_colbert(text: str):
|
39 |
+
if not text.strip():
|
40 |
+
return {"error": "Input text is empty."}
|
41 |
+
inputs = colbert_tokenizer(text, return_tensors='pt', truncation=True, max_length=64)
|
42 |
+
with torch.no_grad():
|
43 |
+
outputs = colbert_model(**inputs)
|
44 |
+
# last_hidden_state: (1, seq_len, hidden_size)
|
45 |
+
embeddings = outputs.last_hidden_state.squeeze(0).tolist()
|
46 |
+
return {"colbert_embeddings": embeddings}
|
47 |
+
|
48 |
+
# Build Gradio interface with tabs for each model
|
49 |
+
with gr.Blocks(title="Text Embedding Playground") as demo:
|
50 |
+
gr.Markdown("# Text Embedding Playground\nChoose a model and input text to get embeddings.")
|
51 |
+
with gr.Tab("Dense (MiniLM-L6-v2)"):
|
52 |
+
txt1 = gr.Textbox(lines=3, label="Input Text")
|
53 |
+
out1 = gr.JSON(label="Embedding")
|
54 |
+
txt1.submit(embed_dense, txt1, out1)
|
55 |
+
gr.Button("Embed").click(embed_dense, txt1, out1)
|
56 |
+
with gr.Tab("Sparse (BM25)"):
|
57 |
+
txt2 = gr.Textbox(lines=3, label="Input Text")
|
58 |
+
out2 = gr.JSON(label="Term Weights")
|
59 |
+
txt2.submit(embed_sparse, txt2, out2)
|
60 |
+
gr.Button("Embed").click(embed_sparse, txt2, out2)
|
61 |
+
with gr.Tab("Late-Interaction (ColBERT)"):
|
62 |
+
txt3 = gr.Textbox(lines=3, label="Input Text")
|
63 |
+
out3 = gr.JSON(label="Embeddings per Token")
|
64 |
+
txt3.submit(embed_colbert, txt3, out3)
|
65 |
+
gr.Button("Embed").click(embed_colbert, txt3, out3)
|
66 |
+
|
67 |
+
if __name__ == "__main__":
|
68 |
+
demo.launch()
|