Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -74,8 +74,8 @@ class DocumentRAG:
|
|
| 74 |
def setup_fallback_model(self):
|
| 75 |
"""Fallback to smaller model if Mistral fails"""
|
| 76 |
try:
|
| 77 |
-
# Use a
|
| 78 |
-
model_name = "
|
| 79 |
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
|
| 80 |
self.model = AutoModelForCausalLM.from_pretrained(model_name)
|
| 81 |
|
|
@@ -86,8 +86,38 @@ class DocumentRAG:
|
|
| 86 |
print("β
Fallback model loaded")
|
| 87 |
except Exception as e:
|
| 88 |
print(f"β Fallback model failed: {e}")
|
|
|
|
| 89 |
self.model = None
|
| 90 |
self.tokenizer = None
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 91 |
|
| 92 |
def extract_text_from_file(self, file_path: str) -> str:
|
| 93 |
"""Extract text from various file formats"""
|
|
@@ -251,7 +281,7 @@ class DocumentRAG:
|
|
| 251 |
return ""
|
| 252 |
|
| 253 |
def generate_answer(self, query: str, context: str) -> str:
|
| 254 |
-
"""Generate answer using the LLM with
|
| 255 |
if self.model is None or self.tokenizer is None:
|
| 256 |
return "β Model not available. Please try again."
|
| 257 |
|
|
@@ -261,28 +291,37 @@ class DocumentRAG:
|
|
| 261 |
is_mistral = 'mistral' in model_name
|
| 262 |
|
| 263 |
if is_mistral:
|
| 264 |
-
#
|
| 265 |
-
prompt = f"""<s>[INST] You are a
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 266 |
|
| 267 |
-
|
| 268 |
-
{context[:
|
| 269 |
|
| 270 |
-
|
| 271 |
|
| 272 |
-
|
| 273 |
else:
|
| 274 |
-
#
|
| 275 |
-
prompt = f"""
|
| 276 |
|
| 277 |
-
|
|
|
|
| 278 |
|
| 279 |
-
|
|
|
|
|
|
|
| 280 |
|
| 281 |
# Tokenize with proper handling
|
| 282 |
inputs = self.tokenizer(
|
| 283 |
prompt,
|
| 284 |
return_tensors="pt",
|
| 285 |
-
max_length=
|
| 286 |
truncation=True,
|
| 287 |
padding=True
|
| 288 |
)
|
|
@@ -291,15 +330,17 @@ Answer based on the context:"""
|
|
| 291 |
if torch.cuda.is_available() and next(self.model.parameters()).is_cuda:
|
| 292 |
inputs = {k: v.cuda() for k, v in inputs.items()}
|
| 293 |
|
| 294 |
-
# Generate with
|
| 295 |
with torch.no_grad():
|
| 296 |
outputs = self.model.generate(
|
| 297 |
**inputs,
|
| 298 |
-
max_new_tokens=
|
| 299 |
-
temperature=0.
|
| 300 |
-
do_sample=
|
| 301 |
-
|
| 302 |
-
|
|
|
|
|
|
|
| 303 |
pad_token_id=self.tokenizer.pad_token_id,
|
| 304 |
eos_token_id=self.tokenizer.eos_token_id
|
| 305 |
)
|
|
@@ -312,18 +353,64 @@ Answer based on the context:"""
|
|
| 312 |
answer = full_response.split("[/INST]")[-1].strip()
|
| 313 |
else:
|
| 314 |
# For other models, remove the prompt
|
| 315 |
-
|
|
|
|
|
|
|
|
|
|
| 316 |
|
| 317 |
-
#
|
| 318 |
-
answer =
|
| 319 |
|
| 320 |
return answer if answer else "I couldn't generate a proper response based on the context."
|
| 321 |
|
| 322 |
except Exception as e:
|
| 323 |
return f"β Error generating answer: {str(e)}"
|
| 324 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 325 |
def answer_question(self, query: str) -> str:
|
| 326 |
-
"""Main function to answer questions"""
|
| 327 |
if not query.strip():
|
| 328 |
return "β Please ask a question!"
|
| 329 |
|
|
@@ -337,17 +424,72 @@ Answer based on the context:"""
|
|
| 337 |
if not context:
|
| 338 |
return "π No relevant information found in the uploaded documents for your question."
|
| 339 |
|
| 340 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 341 |
answer = self.generate_answer(query, context)
|
| 342 |
|
| 343 |
-
#
|
| 344 |
if answer and not answer.startswith("β"):
|
|
|
|
|
|
|
|
|
|
|
|
|
| 345 |
return f"π‘ **Answer:** {answer}\n\nπ **Relevant Context:**\n{context[:400]}..."
|
| 346 |
else:
|
| 347 |
return answer
|
| 348 |
|
| 349 |
except Exception as e:
|
| 350 |
return f"β Error answering question: {str(e)}"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 351 |
|
| 352 |
# Initialize the RAG system
|
| 353 |
print("Initializing Document RAG System...")
|
|
|
|
| 74 |
def setup_fallback_model(self):
|
| 75 |
"""Fallback to smaller model if Mistral fails"""
|
| 76 |
try:
|
| 77 |
+
# Use a model that's better for factual Q&A and less prone to hallucination
|
| 78 |
+
model_name = "microsoft/DialoGPT-small"
|
| 79 |
self.tokenizer = AutoTokenizer.from_pretrained(model_name)
|
| 80 |
self.model = AutoModelForCausalLM.from_pretrained(model_name)
|
| 81 |
|
|
|
|
| 86 |
print("β
Fallback model loaded")
|
| 87 |
except Exception as e:
|
| 88 |
print(f"β Fallback model failed: {e}")
|
| 89 |
+
# Try an even simpler approach - return context-based answers without generation
|
| 90 |
self.model = None
|
| 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 |
+
"""Simple context-based answering when model is not available"""
|
| 96 |
+
if not context:
|
| 97 |
+
return "No relevant information found in the documents."
|
| 98 |
+
|
| 99 |
+
# Simple keyword matching approach
|
| 100 |
+
query_words = set(query.lower().split())
|
| 101 |
+
context_sentences = context.split('.')
|
| 102 |
+
|
| 103 |
+
# Find sentences that contain query keywords
|
| 104 |
+
relevant_sentences = []
|
| 105 |
+
for sentence in context_sentences:
|
| 106 |
+
sentence = sentence.strip()
|
| 107 |
+
if len(sentence) < 10: # Skip very short sentences
|
| 108 |
+
continue
|
| 109 |
+
|
| 110 |
+
sentence_words = set(sentence.lower().split())
|
| 111 |
+
# Check if sentence contains at least 2 query words or important keywords
|
| 112 |
+
common_words = query_words.intersection(sentence_words)
|
| 113 |
+
if len(common_words) >= 2 or any(word in sentence.lower() for word in ['name', 'education', 'experience', 'skill', 'project']):
|
| 114 |
+
relevant_sentences.append(sentence)
|
| 115 |
+
|
| 116 |
+
if relevant_sentences:
|
| 117 |
+
# Return the most relevant sentences
|
| 118 |
+
return '. '.join(relevant_sentences[:3]) + '.'
|
| 119 |
+
else:
|
| 120 |
+
return "The information needed to answer this question is not available in the provided documents."
|
| 121 |
|
| 122 |
def extract_text_from_file(self, file_path: str) -> str:
|
| 123 |
"""Extract text from various file formats"""
|
|
|
|
| 281 |
return ""
|
| 282 |
|
| 283 |
def generate_answer(self, query: str, context: str) -> str:
|
| 284 |
+
"""Generate answer using the LLM with anti-hallucination techniques"""
|
| 285 |
if self.model is None or self.tokenizer is None:
|
| 286 |
return "β Model not available. Please try again."
|
| 287 |
|
|
|
|
| 291 |
is_mistral = 'mistral' in model_name
|
| 292 |
|
| 293 |
if is_mistral:
|
| 294 |
+
# Anti-hallucination prompt for Mistral
|
| 295 |
+
prompt = f"""<s>[INST] You are a document analysis assistant. You must ONLY answer based on the provided context. Do NOT use any external knowledge.
|
| 296 |
+
|
| 297 |
+
STRICT RULES:
|
| 298 |
+
1. Answer ONLY using information from the context below
|
| 299 |
+
2. If the answer is not in the context, respond: "The information needed to answer this question is not available in the provided documents."
|
| 300 |
+
3. Do NOT make assumptions or add information not in the context
|
| 301 |
+
4. Quote relevant parts from the context when possible
|
| 302 |
|
| 303 |
+
CONTEXT:
|
| 304 |
+
{context[:1200]}
|
| 305 |
|
| 306 |
+
QUESTION: {query}
|
| 307 |
|
| 308 |
+
Remember: Use ONLY the context above. No external knowledge allowed. [/INST]"""
|
| 309 |
else:
|
| 310 |
+
# Anti-hallucination prompt for fallback models
|
| 311 |
+
prompt = f"""INSTRUCTIONS: Answer the question using ONLY the information provided in the context. Do not use external knowledge.
|
| 312 |
|
| 313 |
+
CONTEXT:
|
| 314 |
+
{context[:800]}
|
| 315 |
|
| 316 |
+
QUESTION: {query}
|
| 317 |
+
|
| 318 |
+
ANSWER (using only the context above):"""
|
| 319 |
|
| 320 |
# Tokenize with proper handling
|
| 321 |
inputs = self.tokenizer(
|
| 322 |
prompt,
|
| 323 |
return_tensors="pt",
|
| 324 |
+
max_length=600, # Further reduced to prevent truncation issues
|
| 325 |
truncation=True,
|
| 326 |
padding=True
|
| 327 |
)
|
|
|
|
| 330 |
if torch.cuda.is_available() and next(self.model.parameters()).is_cuda:
|
| 331 |
inputs = {k: v.cuda() for k, v in inputs.items()}
|
| 332 |
|
| 333 |
+
# Generate with anti-hallucination parameters
|
| 334 |
with torch.no_grad():
|
| 335 |
outputs = self.model.generate(
|
| 336 |
**inputs,
|
| 337 |
+
max_new_tokens=100, # Shorter responses to reduce hallucination
|
| 338 |
+
temperature=0.1, # Very low temperature for factual responses
|
| 339 |
+
do_sample=False, # Use greedy decoding for consistency
|
| 340 |
+
num_beams=3, # Beam search for better quality
|
| 341 |
+
early_stopping=True,
|
| 342 |
+
repetition_penalty=1.2,
|
| 343 |
+
no_repeat_ngram_size=3,
|
| 344 |
pad_token_id=self.tokenizer.pad_token_id,
|
| 345 |
eos_token_id=self.tokenizer.eos_token_id
|
| 346 |
)
|
|
|
|
| 353 |
answer = full_response.split("[/INST]")[-1].strip()
|
| 354 |
else:
|
| 355 |
# For other models, remove the prompt
|
| 356 |
+
if "ANSWER (using only the context above):" in full_response:
|
| 357 |
+
answer = full_response.split("ANSWER (using only the context above):")[-1].strip()
|
| 358 |
+
else:
|
| 359 |
+
answer = full_response[len(prompt):].strip()
|
| 360 |
|
| 361 |
+
# Post-process to remove hallucinations
|
| 362 |
+
answer = self.post_process_answer(answer, context, query)
|
| 363 |
|
| 364 |
return answer if answer else "I couldn't generate a proper response based on the context."
|
| 365 |
|
| 366 |
except Exception as e:
|
| 367 |
return f"β Error generating answer: {str(e)}"
|
| 368 |
|
| 369 |
+
def post_process_answer(self, answer: str, context: str, query: str) -> str:
|
| 370 |
+
"""Post-process answer to reduce hallucinations"""
|
| 371 |
+
if not answer or len(answer) < 5:
|
| 372 |
+
return "The information needed to answer this question is not available in the provided documents."
|
| 373 |
+
|
| 374 |
+
# Remove common hallucination patterns
|
| 375 |
+
hallucination_patterns = [
|
| 376 |
+
"what are you doing",
|
| 377 |
+
"what do you think",
|
| 378 |
+
"in my opinion",
|
| 379 |
+
"i believe",
|
| 380 |
+
"personally",
|
| 381 |
+
"from my experience",
|
| 382 |
+
"generally speaking",
|
| 383 |
+
"it is known that",
|
| 384 |
+
"everyone knows"
|
| 385 |
+
]
|
| 386 |
+
|
| 387 |
+
answer_lower = answer.lower()
|
| 388 |
+
for pattern in hallucination_patterns:
|
| 389 |
+
if pattern in answer_lower:
|
| 390 |
+
return "The information needed to answer this question is not available in the provided documents."
|
| 391 |
+
|
| 392 |
+
# Check if answer contains information that's not in context
|
| 393 |
+
# Simple check: if answer is much longer than query and doesn't reference context
|
| 394 |
+
if len(answer) > len(query) * 3 and not any(word in answer.lower() for word in context.lower().split()[:20]):
|
| 395 |
+
return "The information needed to answer this question is not available in the provided documents."
|
| 396 |
+
|
| 397 |
+
# Clean up the answer
|
| 398 |
+
answer = answer.strip()
|
| 399 |
+
|
| 400 |
+
# Remove repetitive parts
|
| 401 |
+
sentences = answer.split('.')
|
| 402 |
+
unique_sentences = []
|
| 403 |
+
for sentence in sentences:
|
| 404 |
+
sentence = sentence.strip()
|
| 405 |
+
if sentence and sentence not in unique_sentences:
|
| 406 |
+
unique_sentences.append(sentence)
|
| 407 |
+
|
| 408 |
+
cleaned_answer = '. '.join(unique_sentences)
|
| 409 |
+
|
| 410 |
+
return cleaned_answer if cleaned_answer else "The information needed to answer this question is not available in the provided documents."
|
| 411 |
+
|
| 412 |
def answer_question(self, query: str) -> str:
|
| 413 |
+
"""Main function to answer questions with anti-hallucination measures"""
|
| 414 |
if not query.strip():
|
| 415 |
return "β Please ask a question!"
|
| 416 |
|
|
|
|
| 424 |
if not context:
|
| 425 |
return "π No relevant information found in the uploaded documents for your question."
|
| 426 |
|
| 427 |
+
# If no model available, use simple context-based answering
|
| 428 |
+
if self.model is None:
|
| 429 |
+
answer = self.simple_context_answer(query, context)
|
| 430 |
+
return f"π‘ **Answer:** {answer}\n\nπ **Source:** {context[:300]}..."
|
| 431 |
+
|
| 432 |
+
# Generate answer using the model
|
| 433 |
answer = self.generate_answer(query, context)
|
| 434 |
|
| 435 |
+
# Additional validation to prevent hallucinations
|
| 436 |
if answer and not answer.startswith("β"):
|
| 437 |
+
# Check if answer seems to be hallucinated
|
| 438 |
+
if self.is_likely_hallucination(answer, context):
|
| 439 |
+
answer = "The information needed to answer this question is not available in the provided documents."
|
| 440 |
+
|
| 441 |
return f"π‘ **Answer:** {answer}\n\nπ **Relevant Context:**\n{context[:400]}..."
|
| 442 |
else:
|
| 443 |
return answer
|
| 444 |
|
| 445 |
except Exception as e:
|
| 446 |
return f"β Error answering question: {str(e)}"
|
| 447 |
+
|
| 448 |
+
def is_likely_hallucination(self, answer: str, context: str) -> bool:
|
| 449 |
+
"""Check if the answer is likely a hallucination"""
|
| 450 |
+
# Convert to lowercase for comparison
|
| 451 |
+
answer_lower = answer.lower()
|
| 452 |
+
context_lower = context.lower()
|
| 453 |
+
|
| 454 |
+
# Check for obvious hallucination patterns
|
| 455 |
+
hallucination_indicators = [
|
| 456 |
+
"what are you doing",
|
| 457 |
+
"what do you think",
|
| 458 |
+
"how are you",
|
| 459 |
+
"i think",
|
| 460 |
+
"in my opinion",
|
| 461 |
+
"from my experience",
|
| 462 |
+
"generally speaking",
|
| 463 |
+
"it is well known",
|
| 464 |
+
"everyone knows",
|
| 465 |
+
"obviously",
|
| 466 |
+
"clearly",
|
| 467 |
+
"of course"
|
| 468 |
+
]
|
| 469 |
+
|
| 470 |
+
for indicator in hallucination_indicators:
|
| 471 |
+
if indicator in answer_lower:
|
| 472 |
+
return True
|
| 473 |
+
|
| 474 |
+
# Check if answer contains words that are not in context
|
| 475 |
+
answer_words = set(answer_lower.split())
|
| 476 |
+
context_words = set(context_lower.split())
|
| 477 |
+
|
| 478 |
+
# Remove common words
|
| 479 |
+
common_words = {'the', 'a', 'an', 'and', 'or', 'but', 'in', 'on', 'at', 'to', 'for', 'of', 'with', 'by', 'is', 'are', 'was', 'were', 'be', 'been', 'have', 'has', 'had', 'will', 'would', 'could', 'should', 'this', 'that', 'these', 'those'}
|
| 480 |
+
|
| 481 |
+
answer_content_words = answer_words - common_words
|
| 482 |
+
context_content_words = context_words - common_words
|
| 483 |
+
|
| 484 |
+
# If more than 70% of content words in answer are not in context, likely hallucination
|
| 485 |
+
if len(answer_content_words) > 0:
|
| 486 |
+
overlap = len(answer_content_words.intersection(context_content_words))
|
| 487 |
+
overlap_ratio = overlap / len(answer_content_words)
|
| 488 |
+
|
| 489 |
+
if overlap_ratio < 0.3: # Less than 30% overlap
|
| 490 |
+
return True
|
| 491 |
+
|
| 492 |
+
return False
|
| 493 |
|
| 494 |
# Initialize the RAG system
|
| 495 |
print("Initializing Document RAG System...")
|