Spaces:
Sleeping
Sleeping
karthikeyan-r
commited on
Commit
•
be240ea
1
Parent(s):
4828c45
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,185 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import List
|
2 |
+
import pandas as pd
|
3 |
+
from transformers import AutoTokenizer, AutoModel
|
4 |
+
import torch
|
5 |
+
from langchain_community.document_loaders import PyPDFLoader
|
6 |
+
from IPython.display import display
|
7 |
+
import os
|
8 |
+
os.system('apt-get install poppler-utils')
|
9 |
+
from sklearn.metrics.pairwise import cosine_similarity
|
10 |
+
import numpy as np
|
11 |
+
import streamlit as st
|
12 |
+
|
13 |
+
class PDFProcessor:
|
14 |
+
"""
|
15 |
+
Class for processing PDF files to extract text content.
|
16 |
+
"""
|
17 |
+
def extract_text_from_pdfs(self, file_paths: List[str]) -> List[str]:
|
18 |
+
"""
|
19 |
+
Extract text content from a list of PDF files.
|
20 |
+
|
21 |
+
Args:
|
22 |
+
file_paths (List[str]): A list of file paths to the PDF documents.
|
23 |
+
|
24 |
+
Returns:
|
25 |
+
List[str]: A list of text content extracted from the PDF documents.
|
26 |
+
"""
|
27 |
+
texts = []
|
28 |
+
for file_path in file_paths:
|
29 |
+
try:
|
30 |
+
loader = PyPDFLoader(file_path)
|
31 |
+
pages = loader.load_and_split()
|
32 |
+
|
33 |
+
for page in pages:
|
34 |
+
if isinstance(page.page_content, bytes):
|
35 |
+
text = page.page_content.decode('utf-8', errors='ignore')
|
36 |
+
elif isinstance(page.page_content, str):
|
37 |
+
text = page.page_content
|
38 |
+
else:
|
39 |
+
print(f"Unexpected type: {type(page.page_content)}")
|
40 |
+
continue
|
41 |
+
texts.append(text)
|
42 |
+
except Exception as e:
|
43 |
+
print(f"Failed to process {file_path}: {e}")
|
44 |
+
|
45 |
+
return texts
|
46 |
+
|
47 |
+
class EmbeddingsProcessor:
|
48 |
+
"""
|
49 |
+
Class for processing text to obtain embeddings using a transformer model.
|
50 |
+
"""
|
51 |
+
def __init__(self, model_name: str):
|
52 |
+
"""
|
53 |
+
Initialize the EmbeddingsProcessor with a pre-trained model.
|
54 |
+
|
55 |
+
Args:
|
56 |
+
model_name (str): The name of the pre-trained model to use for generating embeddings.
|
57 |
+
"""
|
58 |
+
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
|
59 |
+
self.model = AutoModel.from_pretrained(model_name).to('cuda')
|
60 |
+
|
61 |
+
def get_embeddings(self, texts: List[str]) -> np.ndarray:
|
62 |
+
"""
|
63 |
+
Generate embeddings for a list of texts.
|
64 |
+
|
65 |
+
Args:
|
66 |
+
texts (List[str]): A list of text strings for which to generate embeddings.
|
67 |
+
|
68 |
+
Returns:
|
69 |
+
np.ndarray: A NumPy array of embeddings for the provided texts.
|
70 |
+
"""
|
71 |
+
encoded_input = self.tokenizer(texts, padding=True, truncation=True, return_tensors="pt")
|
72 |
+
encoded_input = {k: v.to('cuda') for k, v in encoded_input.items()}
|
73 |
+
model_output = self.model(**encoded_input)
|
74 |
+
return model_output.last_hidden_state.mean(dim=1).detach().cpu().numpy()
|
75 |
+
|
76 |
+
def compute_similarity(template_embeddings: np.ndarray, contract_embeddings: np.ndarray) -> np.ndarray:
|
77 |
+
"""
|
78 |
+
Compute cosine similarity between template and contract embeddings.
|
79 |
+
|
80 |
+
Args:
|
81 |
+
template_embeddings (np.ndarray): A NumPy array of template embeddings.
|
82 |
+
contract_embeddings (np.ndarray): A NumPy array of contract embeddings.
|
83 |
+
|
84 |
+
Returns:
|
85 |
+
np.ndarray: A NumPy array of similarity scores between contracts and templates.
|
86 |
+
"""
|
87 |
+
return cosine_similarity(contract_embeddings, template_embeddings)
|
88 |
+
|
89 |
+
def clear_folder(path):
|
90 |
+
if not os.path.exists(path):
|
91 |
+
os.makedirs(path) # Create the directory if it doesn't exist
|
92 |
+
for file in os.listdir(path):
|
93 |
+
file_path = os.path.join(path, file)
|
94 |
+
try:
|
95 |
+
if os.path.isfile(file_path):
|
96 |
+
os.unlink(file_path)
|
97 |
+
except Exception as e:
|
98 |
+
print(f"Failed to delete {file_path}: {e}")
|
99 |
+
|
100 |
+
def save_uploaded_file(uploaded_file, path):
|
101 |
+
try:
|
102 |
+
with open(os.path.join(path, uploaded_file.name), "wb") as f:
|
103 |
+
f.write(uploaded_file.getbuffer())
|
104 |
+
return True
|
105 |
+
except:
|
106 |
+
return False
|
107 |
+
|
108 |
+
# Streamlit UI
|
109 |
+
st.title('PDF Similarity Checker')
|
110 |
+
|
111 |
+
col1, col2 = st.columns(2)
|
112 |
+
|
113 |
+
# Clear the templates and contracts folders before uploading new files
|
114 |
+
templates_folder = './templates'
|
115 |
+
contracts_folder = './contracts'
|
116 |
+
|
117 |
+
clear_folder(templates_folder)
|
118 |
+
clear_folder(contracts_folder)
|
119 |
+
|
120 |
+
with col1:
|
121 |
+
st.header("Upload Templates")
|
122 |
+
uploaded_files_templates = st.file_uploader("Choose PDF files", accept_multiple_files=True, type=['pdf'])
|
123 |
+
os.makedirs(templates_folder, exist_ok=True)
|
124 |
+
for uploaded_file in uploaded_files_templates:
|
125 |
+
if save_uploaded_file(uploaded_file, templates_folder):
|
126 |
+
st.write(f"Saved: {uploaded_file.name}")
|
127 |
+
|
128 |
+
with col2:
|
129 |
+
st.header("Upload Contracts")
|
130 |
+
uploaded_files_contracts = st.file_uploader("Choose PDF files", key="contracts", accept_multiple_files=True, type=['pdf'])
|
131 |
+
os.makedirs(contracts_folder, exist_ok=True)
|
132 |
+
for uploaded_file in uploaded_files_contracts:
|
133 |
+
if save_uploaded_file(uploaded_file, contracts_folder):
|
134 |
+
st.write(f"Saved: {uploaded_file.name}")
|
135 |
+
|
136 |
+
model_name = st.selectbox("Select Model", ['sentence-transformers/multi-qa-mpnet-base-dot-v1'], index=0)
|
137 |
+
|
138 |
+
if st.button("Compute Similarities"):
|
139 |
+
pdf_processor = PDFProcessor()
|
140 |
+
embedding_processor = EmbeddingsProcessor(model_name)
|
141 |
+
|
142 |
+
# Process templates
|
143 |
+
template_files = [os.path.join(templates_folder, f) for f in os.listdir(templates_folder)]
|
144 |
+
template_texts = [pdf_processor.extract_text_from_pdfs([f])[0] for f in template_files if pdf_processor.extract_text_from_pdfs([f])]
|
145 |
+
template_embeddings = embedding_processor.get_embeddings(template_texts)
|
146 |
+
|
147 |
+
# Process contracts
|
148 |
+
contract_files = [os.path.join(contracts_folder, f) for f in os.listdir(contracts_folder)]
|
149 |
+
contract_texts = [pdf_processor.extract_text_from_pdfs([f])[0] for f in contract_files if pdf_processor.extract_text_from_pdfs([f])]
|
150 |
+
contract_embeddings = embedding_processor.get_embeddings(contract_texts)
|
151 |
+
|
152 |
+
# Compute similarities
|
153 |
+
similarities = compute_similarity(template_embeddings, contract_embeddings)
|
154 |
+
|
155 |
+
# Display results in a table format
|
156 |
+
similarity_data = []
|
157 |
+
for i, contract_file in enumerate(contract_files):
|
158 |
+
row = [i + 1, os.path.basename(contract_file)] # SI No and contract file name
|
159 |
+
for j in range(len(template_files)):
|
160 |
+
if j < similarities.shape[1] and i < similarities.shape[0]: # Check if indices are within bounds
|
161 |
+
row.append(f"{similarities[i, j] * 100:.2f}%") # Format as percentage
|
162 |
+
else:
|
163 |
+
row.append("N/A") # Handle out-of-bounds indices gracefully
|
164 |
+
similarity_data.append(row)
|
165 |
+
|
166 |
+
# Create a DataFrame for the table
|
167 |
+
columns = ["SI No", "Contract"] + [os.path.basename(template_files[j]) for j in range(len(template_files))]
|
168 |
+
similarity_df = pd.DataFrame(similarity_data, columns=columns)
|
169 |
+
|
170 |
+
# Display maximize option
|
171 |
+
if st.checkbox("Maximize Table View"):
|
172 |
+
st.write("Similarity Scores Table (Maximized):")
|
173 |
+
st.dataframe(similarity_df) # Maximized view
|
174 |
+
else:
|
175 |
+
st.write("Similarity Scores Table:")
|
176 |
+
st.table(similarity_df.style.hide(axis="index")) # Normal view
|
177 |
+
|
178 |
+
# Download option
|
179 |
+
csv = similarity_df.to_csv(index=False).encode('utf-8')
|
180 |
+
st.download_button(
|
181 |
+
label="Download Similarity Table as CSV",
|
182 |
+
data=csv,
|
183 |
+
file_name='similarity_scores.csv',
|
184 |
+
mime='text/csv',
|
185 |
+
)
|