Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
@@ -1,7 +1,9 @@
|
|
1 |
import streamlit as st
|
2 |
import torch
|
|
|
3 |
from transformers import BertTokenizer, BertModel
|
4 |
import pdfplumber
|
|
|
5 |
|
6 |
# Load the pre-trained BERT model and tokenizer once
|
7 |
model_name = "bert-base-uncased"
|
@@ -27,12 +29,11 @@ def get_embeddings(text):
|
|
27 |
with torch.no_grad(): # Disable gradient calculation for inference
|
28 |
outputs = model(**inputs)
|
29 |
|
30 |
-
#
|
31 |
if hasattr(outputs, 'last_hidden_state'):
|
32 |
-
# Extract the embeddings from the last hidden state
|
33 |
return outputs.last_hidden_state[:, 0, :].detach().cpu().numpy() # Move to CPU before converting to numpy
|
34 |
else:
|
35 |
-
raise ValueError("Model output does not contain 'last_hidden_state'.
|
36 |
|
37 |
# Extract text from PDF
|
38 |
def extract_text_from_pdf(pdf_file):
|
@@ -42,9 +43,9 @@ def extract_text_from_pdf(pdf_file):
|
|
42 |
text += page.extract_text() + "\n" # Add newline for better separation
|
43 |
return text
|
44 |
|
45 |
-
#
|
46 |
-
|
47 |
-
|
48 |
|
49 |
# Streamlit app
|
50 |
st.title("PDF Chatbot using BERT")
|
@@ -52,10 +53,15 @@ st.title("PDF Chatbot using BERT")
|
|
52 |
# PDF file upload
|
53 |
pdf_file = st.file_uploader("Upload a PDF file", type=["pdf"])
|
54 |
|
|
|
|
|
|
|
|
|
55 |
if pdf_file:
|
56 |
pdf_text = extract_text_from_pdf(pdf_file)
|
57 |
try:
|
58 |
-
|
|
|
59 |
st.success("PDF loaded successfully!")
|
60 |
except Exception as e:
|
61 |
st.error(f"Error while processing PDF: {e}")
|
@@ -64,16 +70,23 @@ if pdf_file:
|
|
64 |
user_input = st.text_input("Ask a question about the PDF:")
|
65 |
|
66 |
if st.button("Get Response"):
|
67 |
-
if
|
68 |
st.warning("Please upload a PDF file first.")
|
|
|
|
|
69 |
else:
|
70 |
-
# Get embeddings for user input
|
71 |
try:
|
72 |
user_embeddings = get_embeddings(user_input)
|
73 |
-
|
74 |
-
|
|
|
|
|
|
|
|
|
|
|
75 |
st.write("### Response:")
|
76 |
-
st.write(
|
|
|
77 |
except Exception as e:
|
78 |
st.error(f"Error while processing user input: {e}")
|
79 |
|
|
|
1 |
import streamlit as st
|
2 |
import torch
|
3 |
+
import numpy as np
|
4 |
from transformers import BertTokenizer, BertModel
|
5 |
import pdfplumber
|
6 |
+
from sklearn.metrics.pairwise import cosine_similarity
|
7 |
|
8 |
# Load the pre-trained BERT model and tokenizer once
|
9 |
model_name = "bert-base-uncased"
|
|
|
29 |
with torch.no_grad(): # Disable gradient calculation for inference
|
30 |
outputs = model(**inputs)
|
31 |
|
32 |
+
# Extract the embeddings from the last hidden state
|
33 |
if hasattr(outputs, 'last_hidden_state'):
|
|
|
34 |
return outputs.last_hidden_state[:, 0, :].detach().cpu().numpy() # Move to CPU before converting to numpy
|
35 |
else:
|
36 |
+
raise ValueError("Model output does not contain 'last_hidden_state'.")
|
37 |
|
38 |
# Extract text from PDF
|
39 |
def extract_text_from_pdf(pdf_file):
|
|
|
43 |
text += page.extract_text() + "\n" # Add newline for better separation
|
44 |
return text
|
45 |
|
46 |
+
# Split text into sentences for better matching
|
47 |
+
def split_text_into_sentences(text):
|
48 |
+
return text.split('\n') # Split by newlines; adjust as needed
|
49 |
|
50 |
# Streamlit app
|
51 |
st.title("PDF Chatbot using BERT")
|
|
|
53 |
# PDF file upload
|
54 |
pdf_file = st.file_uploader("Upload a PDF file", type=["pdf"])
|
55 |
|
56 |
+
# Store the PDF text and embeddings
|
57 |
+
pdf_text = ""
|
58 |
+
pdf_embeddings = None
|
59 |
+
|
60 |
if pdf_file:
|
61 |
pdf_text = extract_text_from_pdf(pdf_file)
|
62 |
try:
|
63 |
+
pdf_sentences = split_text_into_sentences(pdf_text) # Split PDF text into sentences
|
64 |
+
pdf_embeddings = np.array([get_embeddings(sentence) for sentence in pdf_sentences]) # Get embeddings for each sentence
|
65 |
st.success("PDF loaded successfully!")
|
66 |
except Exception as e:
|
67 |
st.error(f"Error while processing PDF: {e}")
|
|
|
70 |
user_input = st.text_input("Ask a question about the PDF:")
|
71 |
|
72 |
if st.button("Get Response"):
|
73 |
+
if not pdf_sentences:
|
74 |
st.warning("Please upload a PDF file first.")
|
75 |
+
elif not user_input.strip():
|
76 |
+
st.warning("Please enter a question.")
|
77 |
else:
|
|
|
78 |
try:
|
79 |
user_embeddings = get_embeddings(user_input)
|
80 |
+
user_embeddings = user_embeddings.reshape(1, -1) # Reshape for cosine similarity calculation
|
81 |
+
|
82 |
+
# Calculate cosine similarity between user input and PDF sentence embeddings
|
83 |
+
similarities = cosine_similarity(user_embeddings, pdf_embeddings)
|
84 |
+
best_match_index = np.argmax(similarities) # Get the index of the best match
|
85 |
+
|
86 |
+
# Display the most relevant sentence
|
87 |
st.write("### Response:")
|
88 |
+
st.write(pdf_sentences[best_match_index]) # Return the most relevant sentence
|
89 |
+
|
90 |
except Exception as e:
|
91 |
st.error(f"Error while processing user input: {e}")
|
92 |
|