Spaces:
Build error
Build error
File size: 18,479 Bytes
40f6dcf | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 | """
Self-Reflection RAG - Advanced RAG Pattern
RAG system with self-reflection and correction capabilities.
"""
import asyncio
import logging
from dataclasses import dataclass, field
from typing import Any, Dict, List, Optional, Tuple
import time
from ..config.pipeline_configs.rag_pipeline import RAGPipeline, RAGResponse
logger = logging.getLogger(__name__)
@dataclass
class ReflectionResult:
"""Result from self-reflection process."""
needs_correction: bool
confidence_improvement: float
corrected_answer: Optional[str] = None
reasoning: Optional[str] = None
issues_found: List[str] = field(default_factory=list)
@dataclass
class ReflectionRound:
"""Single round of reflection."""
round_number: int
original_query: str
original_answer: str
original_sources: List[Dict[str, Any]]
reflection_result: ReflectionResult
timestamp: float = field(default_factory=time.time)
class SelfReflectionRAG:
"""RAG system with self-reflection and correction capabilities."""
def __init__(self, base_pipeline: RAGPipeline, config: Optional[Dict[str, Any]] = None):
self.pipeline = base_pipeline
self.config = config or {}
# Reflection settings
self.max_reflection_rounds = self.config.get("max_reflection_rounds", 2)
self.confidence_threshold = self.config.get("confidence_threshold", 0.7)
self.enable_fact_checking = self.config.get("enable_fact_checking", True)
self.enable_coherence_checking = self.config.get("enable_coherence_checking", True)
self.enable_source_verification = self.config.get("enable_source_verification", True)
# LLM settings for reflection
self.reflection_model = self.config.get("reflection_model", "gpt-4")
self.correction_model = self.config.get("correction_model", "gpt-4")
async def query_with_reflection(
self, query: str, max_rounds: Optional[int] = None
) -> Dict[str, Any]:
"""Execute query with self-reflection and correction."""
start_time = time.time()
# Initial query
reflection_rounds = []
current_query = query
current_answer = None
current_sources = None
total_confidence_improvement = 0.0
max_rounds = min(max_rounds or self.max_reflection_rounds, self.max_reflection_rounds)
for round_num in range(max_rounds):
logger.info(f"Reflection round {round_num + 1}/{max_rounds}")
# Execute query
response = await self.pipeline.query(
query=current_query, top_k=5, include_sources=True, include_confidence=True
)
current_answer = response.answer
current_sources = response.sources
current_confidence = response.confidence
# Perform self-reflection
if round_num < max_rounds - 1: # Don't reflect on final round
reflection_result = await self._reflect_on_answer(
query, current_answer, current_sources, reflection_rounds
)
# Decide if correction is needed
if reflection_result.needs_correction and reflection_result.corrected_answer:
current_query = reflection_result.corrected_answer
total_confidence_improvement += reflection_result.confidence_improvement
# Create reflection round record
reflection_round = ReflectionRound(
round_number=round_num + 1,
original_query=query,
original_answer=current_answer,
original_sources=current_sources,
reflection_result=reflection_result,
)
reflection_rounds.append(reflection_round)
else:
# No correction needed, this is our final answer
reflection_round = ReflectionRound(
round_number=round_num + 1,
original_query=query,
original_answer=current_answer,
original_sources=current_sources,
reflection_result=reflection_result,
)
reflection_rounds.append(reflection_round)
break
else:
# Final round
reflection_round = ReflectionRound(
round_number=round_num + 1,
original_query=query,
original_answer=current_answer,
original_sources=current_sources,
reflection_result=ReflectionResult(needs_correction=False),
)
reflection_rounds.append(reflection_round)
total_time = (time.time() - start_time) * 1000
return {
"original_query": query,
"final_answer": current_answer,
"final_sources": current_sources,
"final_confidence": current_confidence,
"reflection_rounds": reflection_rounds,
"total_rounds": len(reflection_rounds),
"total_confidence_improvement": total_confidence_improvement,
"total_time_ms": total_time,
"self_corrected": total_confidence_improvement > 0,
"metadata": {
"max_reflection_rounds": max_rounds,
"reflection_threshold": self.confidence_threshold,
},
}
async def _reflect_on_answer(
self,
query: str,
answer: str,
sources: List[Dict[str, Any]],
previous_rounds: List[ReflectionRound],
) -> ReflectionResult:
"""Perform self-reflection on the answer."""
try:
# Analyze different aspects of the answer
issues_found = []
needs_correction = False
corrected_answer = None
# 1. Confidence analysis
confidence_issue = await self._analyze_confidence(answer, sources)
if confidence_issue:
issues_found.extend(confidence_issue)
# 2. Fact checking
if self.enable_fact_checking:
fact_issues = await self._check_factual_accuracy(answer, sources)
issues_found.extend(fact_issues)
# 3. Coherence analysis
if self.enable_coherence_checking:
coherence_issues = await self._check_coherence(query, answer)
issues_found.extend(coherence_issues)
# 4. Source verification
if self.enable_source_verification:
source_issues = await self._verify_sources(answer, sources)
issues_found.extend(source_issues)
# Determine if correction is needed
if issues_found and self.confidence_threshold > 0.0:
avg_confidence = await self._estimate_confidence(answer, sources)
if avg_confidence < self.confidence_threshold:
needs_correction = True
corrected_answer = await self._generate_correction(query, answer, issues_found)
reasoning = self._generate_reflection_reasoning(issues_found, needs_correction)
confidence_improvement = 0.0
if corrected_answer:
confidence_improvement = await self._estimate_confidence_improvement(
answer, corrected_answer
)
return ReflectionResult(
needs_correction=needs_correction,
confidence_improvement=confidence_improvement,
corrected_answer=corrected_answer,
reasoning=reasoning,
issues_found=issues_found,
)
except Exception as e:
logger.error(f"Error in self-reflection: {e}")
return ReflectionResult(
needs_correction=False,
confidence_improvement=0.0,
reasoning=f"Reflection failed: {str(e)}",
)
async def _analyze_confidence(self, answer: str, sources: List[Dict[str, Any]]) -> List[str]:
"""Analyze confidence of the answer."""
issues = []
# Check for hedge words
hedge_phrases = [
"might be",
"could be",
"possibly",
"probably",
"seems like",
"I think",
"it appears",
"roughly",
"approximately",
]
lower_answer = answer.lower()
for phrase in hedge_phrases:
if phrase in lower_answer:
issues.append(f"Contains hedge phrase: '{phrase}'")
# Check for uncertainty indicators
uncertainty_phrases = [
"I'm not sure",
"I cannot confirm",
"there is insufficient information",
"based on limited data",
"this is speculation",
]
for phrase in uncertainty_phrases:
if phrase in lower_answer:
issues.append(f"Contains uncertainty: '{phrase}'")
# Check source quality impact on confidence
if sources:
source_scores = [source.get("score", 0.0) for source in sources]
avg_source_score = sum(source_scores) / len(source_scores)
if avg_source_score < 0.6:
issues.append(f"Low source relevance: {avg_source_score:.2f}")
return issues
async def _check_factual_accuracy(
self, answer: str, sources: List[Dict[str, Any]]
) -> List[str]:
"""Check factual accuracy against sources."""
issues = []
if not sources:
return ["No sources provided for fact-checking"]
# Extract key claims from answer
claims = self._extract_key_claims(answer)
# Check each claim against sources
for claim in claims:
is_supported = await self._verify_claim_against_sources(claim, sources)
if not is_supported:
issues.append(f"Unsupported claim: {claim[:100]}...")
return issues
async def _check_coherence(self, query: str, answer: str) -> List[str]:
"""Check answer coherence."""
issues = []
# Check for contradictions within the answer
sentences = answer.split(".")
for i, sentence in enumerate(sentences):
sentence = sentence.strip()
if len(sentence) < 10:
continue
# Check for contradictions with previous sentences
for j, prev_sentence in enumerate(sentences[:i]):
prev_sentence = prev_sentence.strip()
if len(prev_sentence) < 10:
continue
contradiction = await self._detect_contradiction(prev_sentence, sentence)
if contradiction:
issues.append(
f"Contradiction: '{prev_sentence[:50]}...' vs '{sentence[:50]}...'"
)
# Check answer relevance to query
query_words = set(query.lower().split())
answer_words = set(answer.lower().split())
overlap = len(query_words & answer_words) / len(query_words) if query_words else 0
if overlap < 0.3: # Less than 30% word overlap
issues.append(f"Low query relevance: {overlap:.1%}")
return issues
async def _verify_sources(self, answer: str, sources: List[Dict[str, Any]]) -> List[str]:
"""Verify source quality and relevance."""
issues = []
# Check source diversity
source_ids = set(source.get("document_id", "") for source in sources)
if len(source_ids) < 2 and len(sources) > 1:
issues.append("Low source diversity")
# Check source scores
for source in sources:
score = source.get("score", 0.0)
if score < 0.3:
issues.append(f"Low relevance source: {source.get('title', 'Unknown')}")
# Check for recent sources
# (This would require timestamp information in sources)
return issues
async def _generate_correction(
self, query: str, original_answer: str, issues: List[str]
) -> str:
"""Generate corrected answer."""
try:
# Create correction prompt
issues_text = "\n".join(f"- {issue}" for issue in issues)
correction_prompt = f"""The following answer has identified issues:
Original Query: {query}
Original Answer: {original_answer}
Issues Found:
{issues_text}
Please provide a corrected, more accurate and confident answer that addresses these issues.
Be more specific, better supported by sources, and more confident in your response."""
from openai import OpenAI
client = OpenAI()
response = client.chat.completions.create(
model=self.correction_model,
messages=[
{
"role": "system",
"content": "You are an expert at correcting and improving AI-generated answers to be more accurate and confident.",
},
{"role": "user", "content": correction_prompt},
],
temperature=0.1,
max_tokens=800,
)
corrected_answer = response.choices[0].message.content.strip()
logger.info(f"Generated correction for answer")
return corrected_answer
except Exception as e:
logger.error(f"Error generating correction: {e}")
return original_answer
def _extract_key_claims(self, text: str) -> List[str]:
"""Extract key claims from text."""
# Simple claim extraction - split by sentences and filter
sentences = [s.strip() for s in text.split(".") if len(s.strip()) > 15]
return sentences
async def _verify_claim_against_sources(
self, claim: str, sources: List[Dict[str, Any]]
) -> bool:
"""Verify if a claim is supported by sources."""
claim_words = set(claim.lower().split())
for source in sources:
source_text = source.get("content", "").lower()
source_words = set(source_text.split())
# Check for significant overlap
overlap = len(claim_words & source_words) / len(claim_words) if claim_words else 0
if overlap >= 0.5: # 50% overlap threshold
return True
return False
async def _detect_contradiction(self, sentence1: str, sentence2: str) -> bool:
"""Detect contradiction between two sentences."""
# Simple contradiction patterns
contradictions = [
("not", ""),
("never", "always"),
("no", "yes"),
("false", "true"),
("incorrect", "correct"),
("cannot", "can"),
("impossible", "possible"),
]
words1 = set(sentence1.lower().split())
words2 = set(sentence2.lower().split())
for neg, pos in contradictions:
if (neg in words1 and pos in words2) or (pos in words1 and neg in words2):
return True
return False
async def _estimate_confidence(self, answer: str, sources: List[Dict[str, Any]]) -> float:
"""Estimate confidence in the answer."""
# Base confidence on source quality
if sources:
source_scores = [source.get("score", 0.0) for source in sources]
source_confidence = sum(source_scores) / len(source_scores)
else:
source_confidence = 0.3 # Low confidence without sources
# Adjust based on answer characteristics
answer_length = len(answer.split())
# Long answers might be more comprehensive
length_factor = min(answer_length / 100, 1.2)
# Hedge words reduce confidence
hedge_words = ["might", "could", "possibly", "probably"]
hedge_count = sum(1 for word in hedge_words if word in answer.lower())
hedge_penalty = hedge_count * 0.1
estimated_confidence = source_confidence * length_factor - hedge_penalty
return max(0.0, min(1.0, estimated_confidence))
async def _estimate_confidence_improvement(
self, original_answer: str, corrected_answer: str
) -> float:
"""Estimate confidence improvement from correction."""
# Simple heuristic based on correction characteristics
if corrected_answer == original_answer:
return 0.0
# Corrections that add specificity and citations tend to improve confidence
original_length = len(original_answer.split())
corrected_length = len(corrected_answer.split())
if corrected_length > original_length * 1.2: # Significantly longer
return 0.3
elif corrected_length > original_length * 1.1:
return 0.2
elif corrected_length > original_length:
return 0.1
return 0.05
def _generate_reflection_reasoning(
self, issues_found: List[str], needs_correction: bool
) -> str:
"""Generate reasoning for reflection decision."""
if not issues_found:
return "No significant issues found in the answer."
reasoning_parts = ["Analysis identified the following issues:"]
reasoning_parts.extend(f"• {issue}" for issue in issues_found[:5])
if needs_correction:
reasoning_parts.append("Correction is recommended to improve accuracy and confidence.")
else:
reasoning_parts.append("No correction needed at this time.")
return " ".join(reasoning_parts)
async def get_reflection_stats(self, session_id: Optional[str] = None) -> Dict[str, Any]:
"""Get statistics about reflection performance."""
# This would connect to a metrics system in a full implementation
return {
"session_id": session_id,
"max_reflection_rounds": self.max_reflection_rounds,
"confidence_threshold": self.confidence_threshold,
"features_enabled": {
"fact_checking": self.enable_fact_checking,
"coherence_checking": self.enable_coherence_checking,
"source_verification": self.enable_source_verification,
},
}
|