Spaces:
Sleeping
Sleeping
File size: 16,346 Bytes
50fcf88 2a7fd26 50fcf88 2a7fd26 50fcf88 2a7fd26 50fcf88 2a7fd26 50fcf88 2a7fd26 50fcf88 2a7fd26 50fcf88 2a7fd26 50fcf88 2a7fd26 50fcf88 |
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 |
"""
Verification agent module for answer validation against source documents.
"""
from typing import Dict, List, Optional, Literal
from langchain_core.documents import Document
from langchain_google_genai import ChatGoogleGenerativeAI
from pydantic import BaseModel, Field
import logging
from configuration.parameters import parameters
logger = logging.getLogger(__name__)
class VerificationResult(BaseModel):
"""Structured output model for verification results."""
supported: Literal["YES", "NO", "PARTIAL"] = Field(
description="Whether the answer is supported by the context"
)
confidence: Literal["HIGH", "MEDIUM", "LOW"] = Field(
default="MEDIUM",
description="Confidence level in the verification result"
)
unsupported_claims: List[str] = Field(
default_factory=list,
description="Claims not supported by context"
)
contradictions: List[str] = Field(
default_factory=list,
description="Contradictions between answer and context"
)
relevant: Literal["YES", "NO"] = Field(
description="Whether the answer is relevant to the question"
)
completeness: Literal["COMPLETE", "PARTIAL", "INCOMPLETE"] = Field(
default="PARTIAL",
description="How completely the answer addresses the question"
)
additional_details: str = Field(
default="",
description="Additional explanations and reasoning"
)
class BestAnswerSelection(BaseModel):
"""Structured output model for selecting the best answer from candidates."""
selected_index: int = Field(
description="The index (0-based) of the best answer from the candidates list"
)
reasoning: str = Field(
description="Explanation of why this answer was selected as the best"
)
confidence: Literal["HIGH", "MEDIUM", "LOW"] = Field(
default="MEDIUM",
description="Confidence level in the selection"
)
comparison_summary: str = Field(
default="",
description="Brief comparison of the candidate answers"
)
class VerificationAgent:
"""Agent for verifying answers against source documents."""
def __init__(
self,
llm: Optional[ChatGoogleGenerativeAI] = None,
max_context_chars: int = None,
max_output_tokens: int = None,
) -> None:
"""Initialize the verification agent."""
logger.info("Initializing VerificationAgent...")
self.max_context_chars = max_context_chars or parameters.VERIFICATION_MAX_CONTEXT_CHARS
self.max_output_tokens = max_output_tokens or parameters.VERIFICATION_MAX_OUTPUT_TOKENS
base_llm = llm or ChatGoogleGenerativeAI(
model=parameters.VERIFICATION_AGENT_MODEL,
google_api_key=parameters.GOOGLE_API_KEY,
temperature=0,
max_output_tokens=self.max_output_tokens,
)
self.llm = base_llm
self.structured_llm = base_llm.with_structured_output(VerificationResult)
self.selection_llm = base_llm.with_structured_output(BestAnswerSelection)
logger.info(f"VerificationAgent initialized (model={parameters.VERIFICATION_AGENT_MODEL})")
def generate_prompt(self, answer: str, context: str, question: Optional[str] = None) -> str:
"""Generate verification prompt."""
question_section = f"\n**Original Question:** {question}\n" if question else ""
return f"""Verify the following answer against the provided context.
**Check for:**
1. Factual support (YES/NO/PARTIAL)
2. Confidence level (HIGH/MEDIUM/LOW)
3. Unsupported claims
4. Contradictions
5. Relevance to question
6. Completeness (COMPLETE/PARTIAL/INCOMPLETE)
**Scoring:**
- HIGH: All claims directly stated, no ambiguity
- MEDIUM: Most claims supported, some inferred
- LOW: Significant claims unsupported
{question_section}
**Answer to Verify:**
{answer}
**Context:**
{context}
Provide your verification analysis."""
def format_verification_report(self, verification: VerificationResult) -> str:
"""Format verification result into readable report."""
report = f"**Supported:** {verification.supported}\n"
report += f"**Confidence:** {verification.confidence}\n"
report += f"**Unsupported Claims:** {', '.join(verification.unsupported_claims) or 'None'}\n"
report += f"**Contradictions:** {', '.join(verification.contradictions) or 'None'}\n"
report += f"**Relevant:** {verification.relevant}\n"
report += f"**Completeness:** {verification.completeness}\n"
report += f"**Additional Details:** {verification.additional_details or 'None'}\n"
return report
def generate_feedback_for_research(self, verification: VerificationResult) -> Optional[str]:
"""Generate feedback for research agent if improvements needed."""
feedback_parts = []
if verification.supported == "NO":
feedback_parts.append("Answer lacks sufficient support from documents.")
elif verification.supported == "PARTIAL":
feedback_parts.append("Some parts are not well supported.")
if verification.unsupported_claims:
claims_str = "; ".join(verification.unsupported_claims[:3])
feedback_parts.append(f"Unsupported: {claims_str}")
if verification.contradictions:
contradictions_str = "; ".join(verification.contradictions[:3])
feedback_parts.append(f"Contradictions: {contradictions_str}")
if verification.completeness == "INCOMPLETE":
feedback_parts.append("Answer is incomplete.")
if verification.confidence == "LOW":
feedback_parts.append("Focus on directly verifiable claims.")
# Always add additional_details if present, even if other feedback exists
if verification.additional_details:
feedback_parts.append(f"Additional Details: {verification.additional_details}")
return " | ".join(feedback_parts) if feedback_parts else None
def should_retry_research(self, verification: VerificationResult, verification_report: str = None, feedback: Optional[str] = None) -> bool:
"""Determine if research should be retried."""
# Use structured fields first
if verification.supported == "NO" or verification.relevant == "NO":
return True
if verification.confidence == "LOW" and (
verification.unsupported_claims or verification.contradictions
):
return True
if verification.supported == "PARTIAL" and verification.contradictions:
return True
# Also check verification_report string for extra signals (legacy/fallback)
if verification_report:
if "Supported: NO" in verification_report:
logger.warning("[Re-Research] Answer not supported; triggering re-research.")
return True
elif "Relevant: NO" in verification_report:
logger.warning("[Re-Research] Answer not relevant; triggering re-research.")
return True
elif "Confidence: LOW" in verification_report and "Supported: PARTIAL" in verification_report:
logger.warning("[Re-Research] Low confidence with partial support; triggering re-research.")
return True
elif "Completeness: INCOMPLETE" in verification_report:
logger.warning("[Re-Research] Answer is incomplete; triggering re-research.")
return True
elif "Completeness: PARTIAL" in verification_report:
logger.warning("[Re-Research] Answer is partially complete; triggering re-research.")
return True
# Check feedback for contradiction/unsupported
if feedback and ("contradiction" in feedback.lower() or "unsupported" in feedback.lower()):
logger.warning("[Re-Research] Feedback indicates contradiction/unsupported; triggering re-research.")
return True
return False
def check(self, answer: str, documents: List[Document], question: Optional[str] = None) -> Dict:
"""
Verify answer against provided documents.
Args:
answer: The answer to verify
documents: Source documents for verification
question: Optional original question
Returns:
Dict with verification report, context, and metadata
"""
logger.info(f"Verifying answer ({len(answer)} chars) against {len(documents)} documents")
context = "\n\n".join([doc.page_content for doc in documents])
if len(context) > self.max_context_chars:
logger.debug(f"Context truncated: {len(context)} -> {self.max_context_chars}")
context = context[:self.max_context_chars]
prompt = self.generate_prompt(answer, context, question)
try:
logger.debug("Calling LLM for verification...")
verification_result: VerificationResult = self.structured_llm.invoke(prompt)
logger.info(f"Verification: {verification_result.supported} ({verification_result.confidence})")
except Exception as e:
logger.error(f"Structured output failed: {e}")
try:
response = self.llm.invoke(prompt)
report = response.content if hasattr(response, "content") else str(response)
verification_result = self._parse_unstructured_response(report.strip())
except Exception as fallback_error:
logger.error(f"Fallback failed: {fallback_error}")
verification_result = VerificationResult(
supported="NO",
confidence="LOW",
relevant="NO",
completeness="INCOMPLETE",
additional_details=f"Verification failed: {str(e)}"
)
verification_report = self.format_verification_report(verification_result)
feedback = self.generate_feedback_for_research(verification_result)
if feedback:
logger.debug(f"Generated feedback: {feedback[:80]}...")
return {
"verification_report": verification_report,
"context_used": context,
"structured_result": verification_result.model_dump(),
"should_retry": self.should_retry_research(verification_result, verification_report, feedback),
"feedback": feedback
}
def select_best_answer(
self,
candidate_answers: List[str],
documents: List[Document],
question: str
) -> Dict:
"""
Select the best answer from multiple candidates based on verification criteria.
Args:
candidate_answers: List of candidate answers to evaluate
documents: Source documents for verification
question: The original question
Returns:
Dict with selected answer, index, reasoning, and verification details
"""
logger.info(f"Selecting best answer from {len(candidate_answers)} candidates")
if len(candidate_answers) == 0:
logger.warning("No candidate answers provided")
return {
"selected_answer": "No answers were generated.",
"selected_index": -1,
"reasoning": "No candidates available",
"confidence": "LOW"
}
if len(candidate_answers) == 1:
logger.info("Only one candidate, returning it directly")
return {
"selected_answer": candidate_answers[0],
"selected_index": 0,
"reasoning": "Only one candidate answer was provided",
"confidence": "MEDIUM"
}
context = "\n\n".join([doc.page_content for doc in documents])
if len(context) > self.max_context_chars:
logger.debug(f"Context truncated: {len(context)} -> {self.max_context_chars}")
context = context[:self.max_context_chars]
candidates_text = ""
for i, answer in enumerate(candidate_answers):
candidates_text += f"\n**Candidate {i + 1}:**\n{answer}\n"
prompt = f"""You are evaluating multiple candidate answers to select the best one.
**Original Question:** {question}
**Candidate Answers:**
{candidates_text}
**Source Context:**
{context}
**Evaluation Criteria:**
1. **Factual Accuracy**: Which answer is most accurately supported by the context?
2. **Completeness**: Which answer most thoroughly addresses the question?
3. **Relevance**: Which answer stays most focused on what was asked?
4. **No Contradictions**: Which answer has the fewest contradictions with the source?
5. **Clarity**: Which answer is clearest and most well-structured?
Select the best answer by providing its index (0-based) and explain your reasoning."""
try:
logger.debug("Calling LLM for best answer selection...")
selection_result: BestAnswerSelection = self.selection_llm.invoke(prompt)
selected_index = selection_result.selected_index
if selected_index < 0 or selected_index >= len(candidate_answers):
logger.warning(f"Invalid selection index {selected_index}, defaulting to 0")
selected_index = 0
logger.info(f"Selected candidate {selected_index + 1} with {selection_result.confidence} confidence")
return {
"selected_answer": candidate_answers[selected_index],
"selected_index": selected_index,
"reasoning": selection_result.reasoning,
"confidence": selection_result.confidence,
"comparison_summary": selection_result.comparison_summary
}
except Exception as e:
logger.error(f"Best answer selection failed: {e}")
# Fallback: return the first candidate
return {
"selected_answer": candidate_answers[0],
"selected_index": 0,
"reasoning": f"Selection failed, using first candidate: {str(e)}",
"confidence": "LOW"
}
def _parse_unstructured_response(self, response_text: str) -> VerificationResult:
"""Parse unstructured response into VerificationResult (fallback)."""
try:
data = {
"supported": "NO",
"confidence": "LOW",
"unsupported_claims": [],
"contradictions": [],
"relevant": "NO",
"completeness": "INCOMPLETE",
"additional_details": ""
}
for line in response_text.split('\n'):
if ':' not in line:
continue
key, value = line.split(':', 1)
key = key.strip().lower().replace(' ', '_')
value = value.strip().upper()
if key == "supported":
data["supported"] = "YES" if "YES" in value else ("PARTIAL" if "PARTIAL" in value else "NO")
elif key == "confidence":
data["confidence"] = "HIGH" if "HIGH" in value else ("MEDIUM" if "MEDIUM" in value else "LOW")
elif key == "relevant":
data["relevant"] = "YES" if "YES" in value else "NO"
elif key == "completeness":
if "COMPLETE" in value and "INCOMPLETE" not in value:
data["completeness"] = "COMPLETE"
elif "PARTIAL" in value:
data["completeness"] = "PARTIAL"
return VerificationResult(**data)
except Exception as e:
logger.error(f"Failed to parse response: {e}")
return VerificationResult(
supported="NO",
confidence="LOW",
relevant="NO",
completeness="INCOMPLETE",
additional_details="Failed to parse verification response"
)
|