Update app.py
Browse files
app.py
CHANGED
@@ -15,8 +15,8 @@ logging.basicConfig(
|
|
15 |
logger = logging.getLogger()
|
16 |
|
17 |
# Load models
|
18 |
-
qa_model = pipeline("question-answering", model="distilbert-base-uncased-distilled-squad")
|
19 |
-
embedder = SentenceTransformer('all-MiniLM-L6-v2')
|
20 |
|
21 |
# Helper function to extract text from PDF
|
22 |
def extract_text_from_pdf(file_path):
|
@@ -24,19 +24,27 @@ def extract_text_from_pdf(file_path):
|
|
24 |
with open(file_path, "rb") as file:
|
25 |
pdf_reader = PyPDF2.PdfReader(file)
|
26 |
for page in pdf_reader.pages:
|
27 |
-
text += page.extract_text() + "\n"
|
28 |
return text
|
29 |
|
30 |
# Find the most relevant section in the document
|
|
|
31 |
def find_relevant_section(query, sections, section_embeddings):
|
|
|
|
|
|
|
|
|
|
|
32 |
stopwords = {"and", "the", "is", "for", "to", "a", "an", "of", "in", "on", "at", "with", "by", "it", "as", "so", "what"}
|
33 |
|
34 |
# Semantic search
|
35 |
query_embedding = embedder.encode(query, convert_to_tensor=True)
|
36 |
-
similarities = util.cos_sim(query_embedding, section_embeddings)[0]
|
37 |
best_idx = similarities.argmax().item()
|
38 |
best_section = sections[best_idx]
|
39 |
similarity_score = similarities[best_idx].item()
|
|
|
|
|
40 |
|
41 |
SIMILARITY_THRESHOLD = 0.4
|
42 |
if similarity_score >= SIMILARITY_THRESHOLD:
|
@@ -46,10 +54,14 @@ def find_relevant_section(query, sections, section_embeddings):
|
|
46 |
logger.info(f"Low similarity ({similarity_score}). Falling back to keyword search.")
|
47 |
|
48 |
# Keyword-based fallback search with stopword filtering
|
|
|
49 |
query_words = {word for word in query.lower().split() if word not in stopwords}
|
50 |
for section in sections:
|
51 |
section_words = {word for word in section.lower().split() if word not in stopwords}
|
52 |
common_words = query_words.intersection(section_words)
|
|
|
|
|
|
|
53 |
if len(common_words) >= 2:
|
54 |
logger.info(f"Keyword match found for query: {query} with common words: {common_words}")
|
55 |
return section
|
@@ -57,11 +69,14 @@ def find_relevant_section(query, sections, section_embeddings):
|
|
57 |
logger.info(f"No good keyword match found. Returning default fallback response.")
|
58 |
return "I don’t have enough information to answer that."
|
59 |
|
60 |
-
# Process the uploaded file with detailed logging
|
61 |
def process_file(file, state):
|
|
|
|
|
62 |
if file is None:
|
63 |
logger.info("No file uploaded.")
|
64 |
return [("Bot", "Please upload a file.")], state
|
|
|
|
|
65 |
|
66 |
file_path = file.name
|
67 |
if file_path.lower().endswith(".pdf"):
|
@@ -74,9 +89,12 @@ def process_file(file, state):
|
|
74 |
else:
|
75 |
logger.error(f"Unsupported file format: {file_path}")
|
76 |
return [("Bot", "Unsupported file format. Please upload a PDF or TXT file.")], state
|
77 |
-
|
|
|
78 |
sections = text.split('\n\n')
|
79 |
section_embeddings = embedder.encode(sections, convert_to_tensor=True)
|
|
|
|
|
80 |
state['document_text'] = text
|
81 |
state['sections'] = sections
|
82 |
state['section_embeddings'] = section_embeddings
|
@@ -87,56 +105,77 @@ def process_file(file, state):
|
|
87 |
logger.info(f"Processed file: {file_path}")
|
88 |
return state['chat_history'], state
|
89 |
|
90 |
-
|
91 |
def handle_input(user_input, state):
|
|
|
|
|
92 |
if state['mode'] == 'waiting_for_upload':
|
93 |
state['chat_history'].append(("Bot", "Please upload a file first."))
|
94 |
logger.info("User attempted to interact without uploading a file.")
|
|
|
95 |
elif state['mode'] == 'waiting_for_query':
|
96 |
query = user_input
|
97 |
state['current_query'] = query
|
98 |
state['feedback_count'] = 0
|
|
|
|
|
99 |
context = find_relevant_section(query, state['sections'], state['section_embeddings'])
|
|
|
|
|
100 |
if context == "I don’t have enough information to answer that.":
|
101 |
answer = context
|
102 |
else:
|
103 |
result = qa_model(question=query, context=context)
|
104 |
answer = result["answer"]
|
|
|
105 |
state['last_answer'] = answer
|
106 |
state['mode'] = 'waiting_for_feedback'
|
107 |
state['chat_history'].append(("User", query))
|
108 |
state['chat_history'].append(("Bot", f"Answer: {answer}\nPlease provide feedback: good, too vague, not helpful."))
|
109 |
logger.info(f"Query: {query}, Answer: {answer}")
|
|
|
110 |
elif state['mode'] == 'waiting_for_feedback':
|
111 |
feedback = user_input.lower()
|
112 |
state['chat_history'].append(("User", feedback))
|
113 |
logger.info(f"Feedback: {feedback}")
|
|
|
|
|
|
|
114 |
if feedback == "good" or state['feedback_count'] >= 2:
|
115 |
state['mode'] = 'waiting_for_query'
|
|
|
116 |
if feedback == "good":
|
117 |
state['chat_history'].append(("Bot", "Thank you for your feedback. You can ask another question."))
|
118 |
logger.info("Feedback accepted as 'good'. Waiting for next query.")
|
|
|
119 |
else:
|
120 |
state['chat_history'].append(("Bot", "Maximum feedback iterations reached. You can ask another question."))
|
121 |
logger.info("Max feedback iterations reached. Waiting for next query.")
|
|
|
122 |
else:
|
123 |
query = state['current_query']
|
124 |
context = find_relevant_section(query, state['sections'], state['section_embeddings'])
|
|
|
125 |
if feedback == "too vague":
|
126 |
adjusted_answer = f"{state['last_answer']}\n\n(More details:\n{context[:500]}...)"
|
|
|
127 |
elif feedback == "not helpful":
|
128 |
adjusted_answer = qa_model(question=query + " Please provide more detailed information with examples.", context=context)['answer']
|
|
|
129 |
else:
|
130 |
state['chat_history'].append(("Bot", "Please provide valid feedback: good, too vague, not helpful."))
|
131 |
logger.info(f"Invalid feedback received: {feedback}")
|
132 |
return state['chat_history'], state
|
|
|
133 |
state['last_answer'] = adjusted_answer
|
134 |
state['feedback_count'] += 1
|
135 |
state['chat_history'].append(("Bot", f"Updated answer: {adjusted_answer}\nPlease provide feedback: good, too vague, not helpful."))
|
136 |
logger.info(f"Adjusted answer: {adjusted_answer}")
|
|
|
137 |
return state['chat_history'], state
|
138 |
|
139 |
# Function to return the up-to-date log file for download
|
|
|
140 |
def get_log_file():
|
141 |
# Flush all log handlers to ensure log file is current
|
142 |
for handler in logger.handlers:
|
@@ -148,7 +187,7 @@ def get_log_file():
|
|
148 |
logger.info("Log file downloaded by user.")
|
149 |
return "support_bot_log.txt"
|
150 |
|
151 |
-
# Initial state
|
152 |
initial_state = {
|
153 |
'document_text': None,
|
154 |
'sections': None,
|
@@ -182,4 +221,4 @@ with gr.Blocks() as demo:
|
|
182 |
# Set up download log button
|
183 |
download_btn.click(fn=get_log_file, inputs=[], outputs=download_file)
|
184 |
|
185 |
-
demo.launch(share=True)
|
|
|
15 |
logger = logging.getLogger()
|
16 |
|
17 |
# Load models
|
18 |
+
qa_model = pipeline("question-answering", model="distilbert-base-uncased-distilled-squad") # Load the Hugging Face QA model for extracting answers from retrieved context.
|
19 |
+
embedder = SentenceTransformer('all-MiniLM-L6-v2') # Loading SentenceTransformer to convert text into vector embeddings for cosine similarity search.
|
20 |
|
21 |
# Helper function to extract text from PDF
|
22 |
def extract_text_from_pdf(file_path):
|
|
|
24 |
with open(file_path, "rb") as file:
|
25 |
pdf_reader = PyPDF2.PdfReader(file)
|
26 |
for page in pdf_reader.pages:
|
27 |
+
text += page.extract_text() + "\n" # Extract text from each page and concatenate.
|
28 |
return text
|
29 |
|
30 |
# Find the most relevant section in the document
|
31 |
+
|
32 |
def find_relevant_section(query, sections, section_embeddings):
|
33 |
+
"""
|
34 |
+
1. First, it performs a semantic search using cosine similarity.
|
35 |
+
2. If the similarity score is below a threshold, it falls back to a keyword-based search.
|
36 |
+
"""
|
37 |
+
|
38 |
stopwords = {"and", "the", "is", "for", "to", "a", "an", "of", "in", "on", "at", "with", "by", "it", "as", "so", "what"}
|
39 |
|
40 |
# Semantic search
|
41 |
query_embedding = embedder.encode(query, convert_to_tensor=True)
|
42 |
+
similarities = util.cos_sim(query_embedding, section_embeddings)[0] # Compute cosine similarity between the query embedding and all section embeddings.
|
43 |
best_idx = similarities.argmax().item()
|
44 |
best_section = sections[best_idx]
|
45 |
similarity_score = similarities[best_idx].item()
|
46 |
+
|
47 |
+
# Defining a threshold to determine if semantic search is confident enough.
|
48 |
|
49 |
SIMILARITY_THRESHOLD = 0.4
|
50 |
if similarity_score >= SIMILARITY_THRESHOLD:
|
|
|
54 |
logger.info(f"Low similarity ({similarity_score}). Falling back to keyword search.")
|
55 |
|
56 |
# Keyword-based fallback search with stopword filtering
|
57 |
+
|
58 |
query_words = {word for word in query.lower().split() if word not in stopwords}
|
59 |
for section in sections:
|
60 |
section_words = {word for word in section.lower().split() if word not in stopwords}
|
61 |
common_words = query_words.intersection(section_words)
|
62 |
+
|
63 |
+
# If at least two words match, return this section.
|
64 |
+
|
65 |
if len(common_words) >= 2:
|
66 |
logger.info(f"Keyword match found for query: {query} with common words: {common_words}")
|
67 |
return section
|
|
|
69 |
logger.info(f"No good keyword match found. Returning default fallback response.")
|
70 |
return "I don’t have enough information to answer that."
|
71 |
|
|
|
72 |
def process_file(file, state):
|
73 |
+
"""Handles the uploaded file, processes its text, and prepares it for querying."""
|
74 |
+
|
75 |
if file is None:
|
76 |
logger.info("No file uploaded.")
|
77 |
return [("Bot", "Please upload a file.")], state
|
78 |
+
|
79 |
+
# Determine file type and extract text accordingly.
|
80 |
|
81 |
file_path = file.name
|
82 |
if file_path.lower().endswith(".pdf"):
|
|
|
89 |
else:
|
90 |
logger.error(f"Unsupported file format: {file_path}")
|
91 |
return [("Bot", "Unsupported file format. Please upload a PDF or TXT file.")], state
|
92 |
+
|
93 |
+
# Split document into sections and encode them into embeddings.
|
94 |
sections = text.split('\n\n')
|
95 |
section_embeddings = embedder.encode(sections, convert_to_tensor=True)
|
96 |
+
|
97 |
+
# Store extracted text and embeddings in the chatbot's state dictionary.
|
98 |
state['document_text'] = text
|
99 |
state['sections'] = sections
|
100 |
state['section_embeddings'] = section_embeddings
|
|
|
105 |
logger.info(f"Processed file: {file_path}")
|
106 |
return state['chat_history'], state
|
107 |
|
108 |
+
|
109 |
def handle_input(user_input, state):
|
110 |
+
"""Processes user queries, fetches answers, and handles feedback loops."""
|
111 |
+
|
112 |
if state['mode'] == 'waiting_for_upload':
|
113 |
state['chat_history'].append(("Bot", "Please upload a file first."))
|
114 |
logger.info("User attempted to interact without uploading a file.")
|
115 |
+
|
116 |
elif state['mode'] == 'waiting_for_query':
|
117 |
query = user_input
|
118 |
state['current_query'] = query
|
119 |
state['feedback_count'] = 0
|
120 |
+
|
121 |
+
# Finding the best matching section.
|
122 |
context = find_relevant_section(query, state['sections'], state['section_embeddings'])
|
123 |
+
|
124 |
+
# Generating an answer using the QA model.
|
125 |
if context == "I don’t have enough information to answer that.":
|
126 |
answer = context
|
127 |
else:
|
128 |
result = qa_model(question=query, context=context)
|
129 |
answer = result["answer"]
|
130 |
+
|
131 |
state['last_answer'] = answer
|
132 |
state['mode'] = 'waiting_for_feedback'
|
133 |
state['chat_history'].append(("User", query))
|
134 |
state['chat_history'].append(("Bot", f"Answer: {answer}\nPlease provide feedback: good, too vague, not helpful."))
|
135 |
logger.info(f"Query: {query}, Answer: {answer}")
|
136 |
+
|
137 |
elif state['mode'] == 'waiting_for_feedback':
|
138 |
feedback = user_input.lower()
|
139 |
state['chat_history'].append(("User", feedback))
|
140 |
logger.info(f"Feedback: {feedback}")
|
141 |
+
|
142 |
+
# Handling feedback responses.
|
143 |
+
|
144 |
if feedback == "good" or state['feedback_count'] >= 2:
|
145 |
state['mode'] = 'waiting_for_query'
|
146 |
+
|
147 |
if feedback == "good":
|
148 |
state['chat_history'].append(("Bot", "Thank you for your feedback. You can ask another question."))
|
149 |
logger.info("Feedback accepted as 'good'. Waiting for next query.")
|
150 |
+
|
151 |
else:
|
152 |
state['chat_history'].append(("Bot", "Maximum feedback iterations reached. You can ask another question."))
|
153 |
logger.info("Max feedback iterations reached. Waiting for next query.")
|
154 |
+
|
155 |
else:
|
156 |
query = state['current_query']
|
157 |
context = find_relevant_section(query, state['sections'], state['section_embeddings'])
|
158 |
+
|
159 |
if feedback == "too vague":
|
160 |
adjusted_answer = f"{state['last_answer']}\n\n(More details:\n{context[:500]}...)"
|
161 |
+
|
162 |
elif feedback == "not helpful":
|
163 |
adjusted_answer = qa_model(question=query + " Please provide more detailed information with examples.", context=context)['answer']
|
164 |
+
|
165 |
else:
|
166 |
state['chat_history'].append(("Bot", "Please provide valid feedback: good, too vague, not helpful."))
|
167 |
logger.info(f"Invalid feedback received: {feedback}")
|
168 |
return state['chat_history'], state
|
169 |
+
|
170 |
state['last_answer'] = adjusted_answer
|
171 |
state['feedback_count'] += 1
|
172 |
state['chat_history'].append(("Bot", f"Updated answer: {adjusted_answer}\nPlease provide feedback: good, too vague, not helpful."))
|
173 |
logger.info(f"Adjusted answer: {adjusted_answer}")
|
174 |
+
|
175 |
return state['chat_history'], state
|
176 |
|
177 |
# Function to return the up-to-date log file for download
|
178 |
+
|
179 |
def get_log_file():
|
180 |
# Flush all log handlers to ensure log file is current
|
181 |
for handler in logger.handlers:
|
|
|
187 |
logger.info("Log file downloaded by user.")
|
188 |
return "support_bot_log.txt"
|
189 |
|
190 |
+
# Initial state setup
|
191 |
initial_state = {
|
192 |
'document_text': None,
|
193 |
'sections': None,
|
|
|
221 |
# Set up download log button
|
222 |
download_btn.click(fn=get_log_file, inputs=[], outputs=download_file)
|
223 |
|
224 |
+
demo.launch(share=True)
|