Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -8,6 +8,7 @@ import PyPDF2
|
|
| 8 |
import docx
|
| 9 |
import io
|
| 10 |
import os
|
|
|
|
| 11 |
from typing import List, Optional
|
| 12 |
|
| 13 |
class DocumentRAG:
|
|
@@ -25,6 +26,7 @@ class DocumentRAG:
|
|
| 25 |
self.documents = []
|
| 26 |
self.index = None
|
| 27 |
self.is_indexed = False
|
|
|
|
| 28 |
|
| 29 |
def setup_llm(self):
|
| 30 |
"""Setup quantized Mistral model"""
|
|
@@ -91,79 +93,142 @@ class DocumentRAG:
|
|
| 91 |
self.tokenizer = None
|
| 92 |
print("β οΈ Using context-only mode (no text generation)")
|
| 93 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 94 |
def simple_context_answer(self, query: str, context: str) -> str:
|
| 95 |
-
"""Improved
|
| 96 |
if not context:
|
| 97 |
return "No relevant information found in the documents."
|
| 98 |
|
| 99 |
query_lower = query.lower()
|
| 100 |
|
|
|
|
|
|
|
|
|
|
| 101 |
# Handle "who is" questions specifically
|
| 102 |
if "who is" in query_lower:
|
| 103 |
-
|
| 104 |
-
|
| 105 |
|
| 106 |
-
#
|
| 107 |
-
|
| 108 |
-
name_info = []
|
| 109 |
-
professional_info = []
|
| 110 |
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
if
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 140 |
|
| 141 |
-
#
|
| 142 |
query_words = set(query_lower.split())
|
| 143 |
-
context_sentences = context.split('.')
|
|
|
|
|
|
|
|
|
|
|
|
|
| 144 |
|
| 145 |
-
# Find sentences that contain query keywords
|
| 146 |
-
relevant_sentences = []
|
| 147 |
for sentence in context_sentences:
|
| 148 |
-
|
| 149 |
-
if len(sentence) < 10: # Skip very short sentences
|
| 150 |
continue
|
| 151 |
-
|
| 152 |
sentence_words = set(sentence.lower().split())
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
if
|
| 156 |
-
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
|
| 160 |
-
return
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
|
| 164 |
-
if first_sentences:
|
| 165 |
-
return '. '.join([s.strip() for s in first_sentences if s.strip()]) + '.'
|
| 166 |
-
return "Based on the document content, I found some information but cannot provide a specific answer to your question."
|
| 167 |
|
| 168 |
def extract_text_from_file(self, file_path: str) -> str:
|
| 169 |
"""Extract text from various file formats"""
|
|
@@ -217,36 +282,61 @@ class DocumentRAG:
|
|
| 217 |
except Exception as e2:
|
| 218 |
return f"Error reading TXT: {str(e2)}"
|
| 219 |
|
| 220 |
-
def
|
| 221 |
-
"""
|
| 222 |
if not text.strip():
|
| 223 |
return []
|
| 224 |
|
| 225 |
-
# Split by sentences first, then group into chunks
|
| 226 |
-
sentences = text.replace('\n', ' ').split('. ')
|
| 227 |
chunks = []
|
|
|
|
|
|
|
|
|
|
| 228 |
current_chunk = ""
|
|
|
|
| 229 |
|
| 230 |
-
for
|
| 231 |
-
|
| 232 |
-
if not
|
| 233 |
continue
|
| 234 |
-
|
| 235 |
-
# Add sentence to current chunk
|
| 236 |
-
test_chunk = current_chunk + ". " + sentence if current_chunk else sentence
|
| 237 |
|
| 238 |
-
|
| 239 |
-
|
| 240 |
-
|
| 241 |
-
|
| 242 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 243 |
else:
|
| 244 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 245 |
|
| 246 |
# Add the last chunk
|
| 247 |
if current_chunk:
|
| 248 |
chunks.append(current_chunk.strip())
|
| 249 |
-
|
| 250 |
return chunks
|
| 251 |
|
| 252 |
def process_documents(self, files) -> str:
|
|
@@ -273,8 +363,11 @@ class DocumentRAG:
|
|
| 273 |
if not all_text.strip():
|
| 274 |
return "β No text extracted from files!"
|
| 275 |
|
| 276 |
-
#
|
| 277 |
-
self.
|
|
|
|
|
|
|
|
|
|
| 278 |
|
| 279 |
if not self.documents:
|
| 280 |
return "β No valid text chunks created!"
|
|
@@ -301,8 +394,8 @@ class DocumentRAG:
|
|
| 301 |
except Exception as e:
|
| 302 |
return f"β Error processing documents: {str(e)}"
|
| 303 |
|
| 304 |
-
def retrieve_context(self, query: str, k: int =
|
| 305 |
-
"""Retrieve relevant context
|
| 306 |
if not self.is_indexed:
|
| 307 |
return ""
|
| 308 |
|
|
@@ -312,23 +405,33 @@ class DocumentRAG:
|
|
| 312 |
faiss.normalize_L2(query_embedding)
|
| 313 |
|
| 314 |
# Search for similar chunks
|
| 315 |
-
scores, indices = self.index.search(query_embedding.astype('float32'), k)
|
| 316 |
|
| 317 |
-
# Get relevant documents with
|
| 318 |
relevant_docs = []
|
| 319 |
-
|
| 320 |
-
if idx < len(self.documents) and scores[0][i] > 0.05: # Much lower threshold
|
| 321 |
-
relevant_docs.append(self.documents[idx])
|
| 322 |
-
|
| 323 |
-
# If no high-similarity matches, take the top results anyway
|
| 324 |
-
if not relevant_docs:
|
| 325 |
-
for i, idx in enumerate(indices[0]):
|
| 326 |
-
if idx < len(self.documents):
|
| 327 |
-
relevant_docs.append(self.documents[idx])
|
| 328 |
-
if len(relevant_docs) >= 3: # Take at least 3 chunks
|
| 329 |
-
break
|
| 330 |
|
| 331 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 332 |
|
| 333 |
except Exception as e:
|
| 334 |
print(f"Error in retrieval: {e}")
|
|
@@ -345,33 +448,27 @@ class DocumentRAG:
|
|
| 345 |
is_mistral = 'mistral' in model_name
|
| 346 |
|
| 347 |
if is_mistral:
|
| 348 |
-
#
|
| 349 |
-
prompt = f"""<s>[INST]
|
| 350 |
|
| 351 |
-
|
| 352 |
-
|
| 353 |
-
Context from document:
|
| 354 |
-
{context[:1500]}
|
| 355 |
|
| 356 |
Question: {query}
|
| 357 |
|
| 358 |
-
Provide a
|
| 359 |
else:
|
| 360 |
-
#
|
| 361 |
-
prompt = f"""
|
| 362 |
-
|
| 363 |
-
Resume Information:
|
| 364 |
-
{context[:1000]}
|
| 365 |
|
| 366 |
Question: {query}
|
|
|
|
| 367 |
|
| 368 |
-
|
| 369 |
-
|
| 370 |
-
# Tokenize with proper handling
|
| 371 |
inputs = self.tokenizer(
|
| 372 |
prompt,
|
| 373 |
return_tensors="pt",
|
| 374 |
-
max_length=
|
| 375 |
truncation=True,
|
| 376 |
padding=True
|
| 377 |
)
|
|
@@ -380,17 +477,16 @@ Answer (be direct and concise):"""
|
|
| 380 |
if torch.cuda.is_available() and next(self.model.parameters()).is_cuda:
|
| 381 |
inputs = {k: v.cuda() for k, v in inputs.items()}
|
| 382 |
|
| 383 |
-
# Generate with
|
| 384 |
with torch.no_grad():
|
| 385 |
outputs = self.model.generate(
|
| 386 |
**inputs,
|
| 387 |
-
max_new_tokens=
|
| 388 |
-
temperature=0.
|
| 389 |
do_sample=True,
|
| 390 |
-
top_p=0.
|
| 391 |
-
num_beams=3,
|
| 392 |
early_stopping=True,
|
| 393 |
-
repetition_penalty=1.
|
| 394 |
pad_token_id=self.tokenizer.pad_token_id,
|
| 395 |
eos_token_id=self.tokenizer.eos_token_id
|
| 396 |
)
|
|
@@ -398,72 +494,50 @@ Answer (be direct and concise):"""
|
|
| 398 |
# Decode response
|
| 399 |
full_response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 400 |
|
| 401 |
-
# Extract answer
|
| 402 |
if is_mistral and "[/INST]" in full_response:
|
| 403 |
answer = full_response.split("[/INST]")[-1].strip()
|
| 404 |
else:
|
| 405 |
-
|
| 406 |
-
if "Answer (be direct and concise):" in full_response:
|
| 407 |
-
answer = full_response.split("Answer (be direct and concise):")[-1].strip()
|
| 408 |
-
elif "Answer:" in full_response:
|
| 409 |
-
answer = full_response.split("Answer:")[-1].strip()
|
| 410 |
-
else:
|
| 411 |
-
answer = full_response[len(prompt):].strip()
|
| 412 |
|
| 413 |
-
# Clean
|
| 414 |
answer = self.clean_answer(answer)
|
| 415 |
|
| 416 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 417 |
|
| 418 |
except Exception as e:
|
| 419 |
print(f"Error in generation: {e}")
|
| 420 |
return self.simple_context_answer(query, context)
|
| 421 |
|
| 422 |
def clean_answer(self, answer: str) -> str:
|
| 423 |
-
"""Clean up the generated answer
|
| 424 |
if not answer or len(answer) < 5:
|
| 425 |
return ""
|
| 426 |
|
| 427 |
-
# Remove
|
| 428 |
-
answer =
|
| 429 |
-
answer =
|
|
|
|
|
|
|
| 430 |
|
| 431 |
-
#
|
| 432 |
-
|
| 433 |
-
cleaned_sentences = []
|
| 434 |
|
| 435 |
-
|
| 436 |
-
|
| 437 |
-
|
| 438 |
-
|
| 439 |
-
|
| 440 |
-
|
| 441 |
-
if any(pattern in sentence.lower() for pattern in [
|
| 442 |
-
'what are you doing', 'what do you think', 'how are you',
|
| 443 |
-
'i am an ai', 'i cannot', 'i don\'t know', 'linkedin: www',
|
| 444 |
-
'github:', 'email:', 'mobile:', '+91-'
|
| 445 |
-
]):
|
| 446 |
-
continue
|
| 447 |
-
|
| 448 |
-
# Clean up common formatting issues
|
| 449 |
-
sentence = sentence.replace(' ', ' ')
|
| 450 |
-
if sentence and len(sentence) > 3:
|
| 451 |
-
cleaned_sentences.append(sentence)
|
| 452 |
-
|
| 453 |
-
if not cleaned_sentences:
|
| 454 |
-
return ""
|
| 455 |
-
|
| 456 |
-
# Reconstruct answer
|
| 457 |
-
cleaned_answer = '. '.join(cleaned_sentences[:2]) # Limit to 2 sentences
|
| 458 |
-
|
| 459 |
-
# Add period if missing
|
| 460 |
-
if cleaned_answer and not cleaned_answer.endswith('.'):
|
| 461 |
-
cleaned_answer += '.'
|
| 462 |
|
| 463 |
-
return
|
| 464 |
|
| 465 |
def answer_question(self, query: str) -> str:
|
| 466 |
-
"""Main function to answer questions
|
| 467 |
if not query.strip():
|
| 468 |
return "β Please ask a question!"
|
| 469 |
|
|
@@ -472,19 +546,18 @@ Answer (be direct and concise):"""
|
|
| 472 |
|
| 473 |
try:
|
| 474 |
# Retrieve relevant context
|
| 475 |
-
context = self.retrieve_context(query, k=
|
| 476 |
|
| 477 |
if not context:
|
| 478 |
-
return "π No relevant information found in the uploaded documents
|
| 479 |
|
| 480 |
# Generate answer
|
| 481 |
answer = self.generate_answer(query, context)
|
| 482 |
|
| 483 |
-
if answer and len(answer) >
|
| 484 |
-
return
|
| 485 |
else:
|
| 486 |
-
|
| 487 |
-
return f"π **Based on the document content:**\n{context[:500]}..."
|
| 488 |
|
| 489 |
except Exception as e:
|
| 490 |
return f"β Error answering question: {str(e)}"
|
|
@@ -532,7 +605,7 @@ def create_interface():
|
|
| 532 |
with gr.Column():
|
| 533 |
question_input = gr.Textbox(
|
| 534 |
label="Your Question",
|
| 535 |
-
placeholder="
|
| 536 |
lines=3
|
| 537 |
)
|
| 538 |
ask_btn = gr.Button("π Get Answer", variant="primary")
|
|
@@ -540,7 +613,7 @@ def create_interface():
|
|
| 540 |
with gr.Column():
|
| 541 |
answer_output = gr.Textbox(
|
| 542 |
label="Answer",
|
| 543 |
-
lines=
|
| 544 |
interactive=False
|
| 545 |
)
|
| 546 |
|
|
@@ -553,11 +626,11 @@ def create_interface():
|
|
| 553 |
# Example questions
|
| 554 |
gr.Markdown("""
|
| 555 |
### π‘ Example Questions:
|
| 556 |
-
-
|
| 557 |
-
-
|
| 558 |
-
- What
|
| 559 |
-
-
|
| 560 |
-
-
|
| 561 |
""")
|
| 562 |
|
| 563 |
return demo
|
|
|
|
| 8 |
import docx
|
| 9 |
import io
|
| 10 |
import os
|
| 11 |
+
import re
|
| 12 |
from typing import List, Optional
|
| 13 |
|
| 14 |
class DocumentRAG:
|
|
|
|
| 26 |
self.documents = []
|
| 27 |
self.index = None
|
| 28 |
self.is_indexed = False
|
| 29 |
+
self.raw_text = "" # Store raw text for fallback
|
| 30 |
|
| 31 |
def setup_llm(self):
|
| 32 |
"""Setup quantized Mistral model"""
|
|
|
|
| 93 |
self.tokenizer = None
|
| 94 |
print("β οΈ Using context-only mode (no text generation)")
|
| 95 |
|
| 96 |
+
def extract_profile_info(self, text: str) -> dict:
|
| 97 |
+
"""Extract key profile information from resume text"""
|
| 98 |
+
profile = {
|
| 99 |
+
'name': '',
|
| 100 |
+
'role': '',
|
| 101 |
+
'skills': [],
|
| 102 |
+
'experience': [],
|
| 103 |
+
'education': [],
|
| 104 |
+
'projects': []
|
| 105 |
+
}
|
| 106 |
+
|
| 107 |
+
lines = text.split('\n')
|
| 108 |
+
current_section = None
|
| 109 |
+
|
| 110 |
+
for line in lines:
|
| 111 |
+
line = line.strip()
|
| 112 |
+
if not line:
|
| 113 |
+
continue
|
| 114 |
+
|
| 115 |
+
line_lower = line.lower()
|
| 116 |
+
|
| 117 |
+
# Extract name (usually first meaningful line)
|
| 118 |
+
if not profile['name'] and len(line.split()) <= 4 and not any(char in line for char in ['@', '.com', '+91', 'linkedin']):
|
| 119 |
+
if not any(word in line_lower for word in ['resume', 'cv', 'experience', 'education', 'skills']):
|
| 120 |
+
profile['name'] = line
|
| 121 |
+
|
| 122 |
+
# Look for role/title indicators
|
| 123 |
+
if any(keyword in line_lower for keyword in ['data scientist', 'software engineer', 'developer', 'analyst', 'intern']):
|
| 124 |
+
if 'data scientist' in line_lower:
|
| 125 |
+
profile['role'] = 'Data Scientist'
|
| 126 |
+
elif 'software engineer' in line_lower:
|
| 127 |
+
profile['role'] = 'Software Engineer'
|
| 128 |
+
elif 'developer' in line_lower:
|
| 129 |
+
profile['role'] = 'Developer'
|
| 130 |
+
elif 'analyst' in line_lower:
|
| 131 |
+
profile['role'] = 'Analyst'
|
| 132 |
+
|
| 133 |
+
# Extract skills
|
| 134 |
+
if any(keyword in line_lower for keyword in ['python', 'machine learning', 'react', 'javascript', 'sql']):
|
| 135 |
+
if 'python' in line_lower:
|
| 136 |
+
profile['skills'].append('Python')
|
| 137 |
+
if 'machine learning' in line_lower:
|
| 138 |
+
profile['skills'].append('Machine Learning')
|
| 139 |
+
if 'react' in line_lower:
|
| 140 |
+
profile['skills'].append('React')
|
| 141 |
+
if 'javascript' in line_lower:
|
| 142 |
+
profile['skills'].append('JavaScript')
|
| 143 |
+
|
| 144 |
+
return profile
|
| 145 |
+
|
| 146 |
def simple_context_answer(self, query: str, context: str) -> str:
|
| 147 |
+
"""Improved smart answering based on context analysis"""
|
| 148 |
if not context:
|
| 149 |
return "No relevant information found in the documents."
|
| 150 |
|
| 151 |
query_lower = query.lower()
|
| 152 |
|
| 153 |
+
# Extract profile information first
|
| 154 |
+
profile = self.extract_profile_info(self.raw_text if self.raw_text else context)
|
| 155 |
+
|
| 156 |
# Handle "who is" questions specifically
|
| 157 |
if "who is" in query_lower:
|
| 158 |
+
name_in_query = re.search(r'who is (\w+)', query_lower)
|
| 159 |
+
person_name = name_in_query.group(1) if name_in_query else "this person"
|
| 160 |
|
| 161 |
+
# Build answer from profile
|
| 162 |
+
answer_parts = []
|
|
|
|
|
|
|
| 163 |
|
| 164 |
+
if profile['name']:
|
| 165 |
+
if profile['role']:
|
| 166 |
+
answer_parts.append(f"{profile['name']} is a {profile['role']}")
|
| 167 |
+
else:
|
| 168 |
+
# Try to infer role from context
|
| 169 |
+
context_lower = context.lower()
|
| 170 |
+
if 'data scientist' in context_lower or ('python' in context_lower and 'machine learning' in context_lower):
|
| 171 |
+
answer_parts.append(f"{profile['name']} is a Data Scientist")
|
| 172 |
+
elif 'software' in context_lower and 'developer' in context_lower:
|
| 173 |
+
answer_parts.append(f"{profile['name']} is a Software Developer")
|
| 174 |
+
else:
|
| 175 |
+
answer_parts.append(f"{profile['name']} is a professional")
|
| 176 |
+
else:
|
| 177 |
+
# Use name from query
|
| 178 |
+
context_lower = context.lower()
|
| 179 |
+
if 'data scientist' in context_lower or ('python' in context_lower and 'machine learning' in context_lower):
|
| 180 |
+
answer_parts.append(f"{person_name.title()} is a Data Scientist")
|
| 181 |
+
elif 'software' in context_lower and 'developer' in context_lower:
|
| 182 |
+
answer_parts.append(f"{person_name.title()} is a Software Developer")
|
| 183 |
+
else:
|
| 184 |
+
answer_parts.append(f"{person_name.title()} is a professional")
|
| 185 |
+
|
| 186 |
+
# Add key skills if available
|
| 187 |
+
if profile['skills']:
|
| 188 |
+
top_skills = profile['skills'][:3] # Top 3 skills
|
| 189 |
+
answer_parts.append(f"with expertise in {', '.join(top_skills)}")
|
| 190 |
+
|
| 191 |
+
if answer_parts:
|
| 192 |
+
return '. '.join(answer_parts) + '.'
|
| 193 |
+
|
| 194 |
+
# Handle other question types
|
| 195 |
+
elif any(keyword in query_lower for keyword in ['what', 'skills', 'experience', 'work']):
|
| 196 |
+
if 'skills' in query_lower:
|
| 197 |
+
if profile['skills']:
|
| 198 |
+
return f"Key skills include: {', '.join(profile['skills'])}."
|
| 199 |
+
elif 'experience' in query_lower or 'work' in query_lower:
|
| 200 |
+
# Look for experience indicators in context
|
| 201 |
+
exp_lines = []
|
| 202 |
+
for line in context.split('\n'):
|
| 203 |
+
if any(word in line.lower() for word in ['experience', 'worked', 'internship', 'project']):
|
| 204 |
+
exp_lines.append(line.strip())
|
| 205 |
+
if exp_lines:
|
| 206 |
+
return exp_lines[0]
|
| 207 |
|
| 208 |
+
# Fallback to keyword matching
|
| 209 |
query_words = set(query_lower.split())
|
| 210 |
+
context_sentences = [s.strip() for s in context.split('.') if s.strip()]
|
| 211 |
+
|
| 212 |
+
# Find most relevant sentence
|
| 213 |
+
best_sentence = ""
|
| 214 |
+
max_matches = 0
|
| 215 |
|
|
|
|
|
|
|
| 216 |
for sentence in context_sentences:
|
| 217 |
+
if len(sentence) < 20: # Skip very short sentences
|
|
|
|
| 218 |
continue
|
| 219 |
+
|
| 220 |
sentence_words = set(sentence.lower().split())
|
| 221 |
+
matches = len(query_words.intersection(sentence_words))
|
| 222 |
+
|
| 223 |
+
if matches > max_matches:
|
| 224 |
+
max_matches = matches
|
| 225 |
+
best_sentence = sentence
|
| 226 |
+
|
| 227 |
+
if best_sentence:
|
| 228 |
+
return best_sentence + '.'
|
| 229 |
+
|
| 230 |
+
# Final fallback
|
| 231 |
+
return "Based on the document, I found relevant information but cannot provide a specific answer."
|
|
|
|
|
|
|
|
|
|
| 232 |
|
| 233 |
def extract_text_from_file(self, file_path: str) -> str:
|
| 234 |
"""Extract text from various file formats"""
|
|
|
|
| 282 |
except Exception as e2:
|
| 283 |
return f"Error reading TXT: {str(e2)}"
|
| 284 |
|
| 285 |
+
def smart_chunk_text(self, text: str) -> List[str]:
|
| 286 |
+
"""Smart chunking that preserves important information together"""
|
| 287 |
if not text.strip():
|
| 288 |
return []
|
| 289 |
|
|
|
|
|
|
|
| 290 |
chunks = []
|
| 291 |
+
lines = text.split('\n')
|
| 292 |
+
|
| 293 |
+
# Create chunks based on semantic meaning
|
| 294 |
current_chunk = ""
|
| 295 |
+
chunk_type = None
|
| 296 |
|
| 297 |
+
for line in lines:
|
| 298 |
+
line = line.strip()
|
| 299 |
+
if not line:
|
| 300 |
continue
|
|
|
|
|
|
|
|
|
|
| 301 |
|
| 302 |
+
line_lower = line.lower()
|
| 303 |
+
|
| 304 |
+
# Identify section types
|
| 305 |
+
new_chunk_type = None
|
| 306 |
+
if any(keyword in line_lower for keyword in ['name', 'email', 'phone', 'linkedin', 'github']):
|
| 307 |
+
new_chunk_type = 'contact'
|
| 308 |
+
elif any(keyword in line_lower for keyword in ['experience', 'work', 'internship']):
|
| 309 |
+
new_chunk_type = 'experience'
|
| 310 |
+
elif any(keyword in line_lower for keyword in ['education', 'degree', 'university', 'college']):
|
| 311 |
+
new_chunk_type = 'education'
|
| 312 |
+
elif any(keyword in line_lower for keyword in ['skills', 'technologies', 'programming']):
|
| 313 |
+
new_chunk_type = 'skills'
|
| 314 |
+
elif any(keyword in line_lower for keyword in ['project', 'developed', 'built']):
|
| 315 |
+
new_chunk_type = 'projects'
|
| 316 |
+
|
| 317 |
+
# If section type changes, save current chunk and start new one
|
| 318 |
+
if new_chunk_type != chunk_type and current_chunk:
|
| 319 |
+
chunks.append(current_chunk.strip())
|
| 320 |
+
current_chunk = line
|
| 321 |
+
chunk_type = new_chunk_type
|
| 322 |
else:
|
| 323 |
+
# Add to current chunk
|
| 324 |
+
if current_chunk:
|
| 325 |
+
current_chunk += "\n" + line
|
| 326 |
+
else:
|
| 327 |
+
current_chunk = line
|
| 328 |
+
chunk_type = new_chunk_type
|
| 329 |
+
|
| 330 |
+
# Limit chunk size
|
| 331 |
+
if len(current_chunk.split()) > 150:
|
| 332 |
+
chunks.append(current_chunk.strip())
|
| 333 |
+
current_chunk = ""
|
| 334 |
+
chunk_type = None
|
| 335 |
|
| 336 |
# Add the last chunk
|
| 337 |
if current_chunk:
|
| 338 |
chunks.append(current_chunk.strip())
|
| 339 |
+
|
| 340 |
return chunks
|
| 341 |
|
| 342 |
def process_documents(self, files) -> str:
|
|
|
|
| 363 |
if not all_text.strip():
|
| 364 |
return "β No text extracted from files!"
|
| 365 |
|
| 366 |
+
# Store raw text for smart answering
|
| 367 |
+
self.raw_text = all_text
|
| 368 |
+
|
| 369 |
+
# Smart chunk the text
|
| 370 |
+
self.documents = self.smart_chunk_text(all_text)
|
| 371 |
|
| 372 |
if not self.documents:
|
| 373 |
return "β No valid text chunks created!"
|
|
|
|
| 394 |
except Exception as e:
|
| 395 |
return f"β Error processing documents: {str(e)}"
|
| 396 |
|
| 397 |
+
def retrieve_context(self, query: str, k: int = 3) -> str:
|
| 398 |
+
"""Retrieve relevant context with improved filtering"""
|
| 399 |
if not self.is_indexed:
|
| 400 |
return ""
|
| 401 |
|
|
|
|
| 405 |
faiss.normalize_L2(query_embedding)
|
| 406 |
|
| 407 |
# Search for similar chunks
|
| 408 |
+
scores, indices = self.index.search(query_embedding.astype('float32'), min(k, len(self.documents)))
|
| 409 |
|
| 410 |
+
# Get relevant documents with reasonable threshold
|
| 411 |
relevant_docs = []
|
| 412 |
+
query_lower = query.lower()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 413 |
|
| 414 |
+
for i, idx in enumerate(indices[0]):
|
| 415 |
+
if idx < len(self.documents):
|
| 416 |
+
doc = self.documents[idx]
|
| 417 |
+
score = scores[0][i]
|
| 418 |
+
|
| 419 |
+
# For "who is" questions, prioritize contact/basic info chunks
|
| 420 |
+
if "who is" in query_lower:
|
| 421 |
+
doc_lower = doc.lower()
|
| 422 |
+
if any(keyword in doc_lower for keyword in ['name', 'email', 'linkedin', 'data scientist', 'developer']):
|
| 423 |
+
relevant_docs.insert(0, doc) # Put at beginning
|
| 424 |
+
elif score > 0.15: # Lower threshold for other relevant content
|
| 425 |
+
relevant_docs.append(doc)
|
| 426 |
+
else:
|
| 427 |
+
if score > 0.2: # Standard threshold
|
| 428 |
+
relevant_docs.append(doc)
|
| 429 |
+
|
| 430 |
+
# If no good matches for "who is", get the first few chunks
|
| 431 |
+
if "who is" in query_lower and not relevant_docs:
|
| 432 |
+
relevant_docs = self.documents[:2]
|
| 433 |
+
|
| 434 |
+
return "\n\n".join(relevant_docs[:3]) # Limit to top 3 chunks
|
| 435 |
|
| 436 |
except Exception as e:
|
| 437 |
print(f"Error in retrieval: {e}")
|
|
|
|
| 448 |
is_mistral = 'mistral' in model_name
|
| 449 |
|
| 450 |
if is_mistral:
|
| 451 |
+
# Focused prompt for Mistral
|
| 452 |
+
prompt = f"""<s>[INST] Answer the question about the person based on their resume. Be concise and direct.
|
| 453 |
|
| 454 |
+
Resume Information:
|
| 455 |
+
{context[:800]}
|
|
|
|
|
|
|
| 456 |
|
| 457 |
Question: {query}
|
| 458 |
|
| 459 |
+
Provide a brief, specific answer in 1 sentence. [/INST]"""
|
| 460 |
else:
|
| 461 |
+
# Focused prompt for fallback models
|
| 462 |
+
prompt = f"""Resume: {context[:600]}
|
|
|
|
|
|
|
|
|
|
| 463 |
|
| 464 |
Question: {query}
|
| 465 |
+
Answer briefly:"""
|
| 466 |
|
| 467 |
+
# Tokenize
|
|
|
|
|
|
|
| 468 |
inputs = self.tokenizer(
|
| 469 |
prompt,
|
| 470 |
return_tensors="pt",
|
| 471 |
+
max_length=600,
|
| 472 |
truncation=True,
|
| 473 |
padding=True
|
| 474 |
)
|
|
|
|
| 477 |
if torch.cuda.is_available() and next(self.model.parameters()).is_cuda:
|
| 478 |
inputs = {k: v.cuda() for k, v in inputs.items()}
|
| 479 |
|
| 480 |
+
# Generate with focused parameters
|
| 481 |
with torch.no_grad():
|
| 482 |
outputs = self.model.generate(
|
| 483 |
**inputs,
|
| 484 |
+
max_new_tokens=50, # Much shorter for focused answers
|
| 485 |
+
temperature=0.1, # Very low for deterministic responses
|
| 486 |
do_sample=True,
|
| 487 |
+
top_p=0.9,
|
|
|
|
| 488 |
early_stopping=True,
|
| 489 |
+
repetition_penalty=1.1,
|
| 490 |
pad_token_id=self.tokenizer.pad_token_id,
|
| 491 |
eos_token_id=self.tokenizer.eos_token_id
|
| 492 |
)
|
|
|
|
| 494 |
# Decode response
|
| 495 |
full_response = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 496 |
|
| 497 |
+
# Extract answer
|
| 498 |
if is_mistral and "[/INST]" in full_response:
|
| 499 |
answer = full_response.split("[/INST]")[-1].strip()
|
| 500 |
else:
|
| 501 |
+
answer = full_response[len(prompt):].strip()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 502 |
|
| 503 |
+
# Clean and validate answer
|
| 504 |
answer = self.clean_answer(answer)
|
| 505 |
|
| 506 |
+
# If answer is too long or poor quality, use fallback
|
| 507 |
+
if not answer or len(answer) > 200:
|
| 508 |
+
return self.simple_context_answer(query, context)
|
| 509 |
+
|
| 510 |
+
return answer
|
| 511 |
|
| 512 |
except Exception as e:
|
| 513 |
print(f"Error in generation: {e}")
|
| 514 |
return self.simple_context_answer(query, context)
|
| 515 |
|
| 516 |
def clean_answer(self, answer: str) -> str:
|
| 517 |
+
"""Clean up the generated answer"""
|
| 518 |
if not answer or len(answer) < 5:
|
| 519 |
return ""
|
| 520 |
|
| 521 |
+
# Remove unwanted patterns
|
| 522 |
+
answer = re.sub(r'--- \w+.*? ---', '', answer)
|
| 523 |
+
answer = re.sub(r'\b\w+@\w+\.\w+\b', '', answer) # Remove emails
|
| 524 |
+
answer = re.sub(r'\+91-?\d+', '', answer) # Remove phone numbers
|
| 525 |
+
answer = answer.replace('LinkedIn:', '').replace('Github:', '')
|
| 526 |
|
| 527 |
+
# Clean up whitespace
|
| 528 |
+
answer = ' '.join(answer.split())
|
|
|
|
| 529 |
|
| 530 |
+
# Take only the first sentence if multiple
|
| 531 |
+
sentences = answer.split('.')
|
| 532 |
+
if sentences:
|
| 533 |
+
first_sentence = sentences[0].strip()
|
| 534 |
+
if len(first_sentence) > 10:
|
| 535 |
+
return first_sentence + '.'
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 536 |
|
| 537 |
+
return answer.strip()
|
| 538 |
|
| 539 |
def answer_question(self, query: str) -> str:
|
| 540 |
+
"""Main function to answer questions"""
|
| 541 |
if not query.strip():
|
| 542 |
return "β Please ask a question!"
|
| 543 |
|
|
|
|
| 546 |
|
| 547 |
try:
|
| 548 |
# Retrieve relevant context
|
| 549 |
+
context = self.retrieve_context(query, k=3)
|
| 550 |
|
| 551 |
if not context:
|
| 552 |
+
return "π No relevant information found in the uploaded documents."
|
| 553 |
|
| 554 |
# Generate answer
|
| 555 |
answer = self.generate_answer(query, context)
|
| 556 |
|
| 557 |
+
if answer and len(answer) > 5:
|
| 558 |
+
return answer
|
| 559 |
else:
|
| 560 |
+
return "I couldn't generate a specific answer from the document content."
|
|
|
|
| 561 |
|
| 562 |
except Exception as e:
|
| 563 |
return f"β Error answering question: {str(e)}"
|
|
|
|
| 605 |
with gr.Column():
|
| 606 |
question_input = gr.Textbox(
|
| 607 |
label="Your Question",
|
| 608 |
+
placeholder="Who is Pradeep?",
|
| 609 |
lines=3
|
| 610 |
)
|
| 611 |
ask_btn = gr.Button("π Get Answer", variant="primary")
|
|
|
|
| 613 |
with gr.Column():
|
| 614 |
answer_output = gr.Textbox(
|
| 615 |
label="Answer",
|
| 616 |
+
lines=6,
|
| 617 |
interactive=False
|
| 618 |
)
|
| 619 |
|
|
|
|
| 626 |
# Example questions
|
| 627 |
gr.Markdown("""
|
| 628 |
### π‘ Example Questions:
|
| 629 |
+
- Who is [Name]?
|
| 630 |
+
- What are [Name]'s skills?
|
| 631 |
+
- What experience does [Name] have?
|
| 632 |
+
- What projects has [Name] worked on?
|
| 633 |
+
- What is [Name]'s educational background?
|
| 634 |
""")
|
| 635 |
|
| 636 |
return demo
|