neural-thinker's picture
feat: clean HuggingFace deployment with essential files only
824bf31
"""
Module: agents.machado_agent
Description: Machado de Assis - Textual Analysis Agent specialized in processing government documents
Author: Anderson H. Silva
Date: 2025-01-24
License: Proprietary - All rights reserved
"""
import asyncio
import hashlib
import re
from datetime import datetime, timedelta
from typing import Any, Dict, List, Optional, Tuple
from dataclasses import dataclass
from enum import Enum
import numpy as np
import pandas as pd
from pydantic import BaseModel, Field as PydanticField
from src.agents.deodoro import BaseAgent, AgentContext, AgentMessage, AgentResponse
from src.core import get_logger
from src.core.exceptions import AgentExecutionError, DataAnalysisError
class DocumentType(Enum):
"""Types of government documents."""
CONTRACT = "contract"
PUBLIC_TENDER = "edital"
LAW = "lei"
DECREE = "decreto"
ORDINANCE = "portaria"
RESOLUTION = "resolucao"
NORMATIVE_INSTRUCTION = "instrucao_normativa"
class AlertSeverity(Enum):
"""Severity levels for document alerts."""
LOW = 1
MEDIUM = 2
HIGH = 3
CRITICAL = 4
URGENT = 5
@dataclass
class EntityExtraction:
"""Extracted entities from document."""
organizations: List[str]
values: List[Dict[str, Any]] # {amount: float, context: str}
dates: List[Dict[str, Any]] # {date: str, event: str}
people: List[str]
locations: List[str]
legal_references: List[str]
@dataclass
class DocumentAlert:
"""Alert for suspicious or problematic content."""
alert_type: str
excerpt: str
legal_violation: Optional[str]
severity: AlertSeverity
confidence: float
explanation: str
recommendation: str
@dataclass
class TextualAnalysisResult:
"""Result of comprehensive textual analysis."""
document_id: str
document_type: DocumentType
entities: EntityExtraction
alerts: List[DocumentAlert]
complexity_score: float # Flesch adapted for PT-BR
transparency_score: float # 0.0 to 1.0
legal_compliance: float # 0.0 to 1.0
readability_grade: int
suspicious_patterns: List[str]
checksum: str
analysis_timestamp: datetime
class TextualAnalysisRequest(BaseModel):
"""Request for textual analysis of government documents."""
document_content: str = PydanticField(description="Full text of the document")
document_type: Optional[str] = PydanticField(default=None, description="Type of document")
document_metadata: Optional[Dict[str, Any]] = PydanticField(default=None, description="Document metadata")
focus_areas: Optional[List[str]] = PydanticField(default=None, description="Specific analysis focus areas")
legal_framework: Optional[List[str]] = PydanticField(default=None, description="Legal frameworks to check against")
complexity_threshold: float = PydanticField(default=0.7, description="Complexity alert threshold")
class MachadoAgent(BaseAgent):
"""
Machado de Assis - Textual Analysis Agent
Specialized in processing government documents, extracting structured information,
detecting inconsistencies, and identifying problematic clauses.
Inspired by Machado de Assis, master of Brazilian literature and language.
"""
def __init__(self):
super().__init__(
name="machado",
description="Textual Analysis Agent specialized in processing government documents",
capabilities=[
"document_parsing",
"named_entity_recognition",
"semantic_analysis",
"legal_compliance_checking",
"ambiguity_detection",
"readability_assessment",
"contract_analysis",
"tender_document_review",
"regulatory_text_processing",
"suspicious_clause_identification",
"linguistic_complexity_analysis",
"transparency_scoring"
]
)
self.logger = get_logger("agent.machado")
# Legal framework references
self._legal_frameworks = {
"CF88": "Constituição Federal de 1988",
"LEI8666": "Lei 8.666/93 - Licitações e Contratos",
"LEI14133": "Lei 14.133/21 - Nova Lei de Licitações",
"LAI": "Lei 12.527/11 - Lei de Acesso à Informação",
"LGPD": "Lei 13.709/18 - Lei Geral de Proteção de Dados"
}
# Suspicious patterns regex
self._suspicious_patterns = {
"urgency_abuse": r"(urgente|emergencial|inadiável)(?!.*justificativa)",
"vague_specifications": r"(conforme|adequado|satisfatório|apropriado)\s+(?!critério|norma)",
"exclusive_criteria": r"(exclusivamente|unicamente|somente)(?=.*fornecedor|empresa)",
"price_manipulation": r"(valor\s+aproximado|preço\s+estimado)(?=.*sigiloso|confidencial)",
"favoritism_indicators": r"(experiência\s+mínima\s+\d+\s+anos?)(?=.*específic)",
}
# NER patterns for Brazilian documents
self._ner_patterns = {
"cnpj": r"\d{2}\.\d{3}\.\d{3}/\d{4}-\d{2}",
"cpf": r"\d{3}\.\d{3}\.\d{3}-\d{2}",
"money": r"R\$\s*[\d,.]+",
"percentage": r"\d+(?:,\d+)?%",
"law_reference": r"Lei\s+n?º?\s*[\d./-]+",
"article": r"Art\.?\s*\d+[º°]?",
}
async def process(
self,
message: AgentMessage,
context: AgentContext,
) -> AgentResponse:
"""
Process textual analysis request.
Args:
message: Document analysis request
context: Agent execution context
Returns:
Comprehensive textual analysis results
"""
try:
self.logger.info(
"Processing textual analysis request",
investigation_id=context.investigation_id,
message_type=message.type,
)
# Parse request
if isinstance(message.data, dict):
request = TextualAnalysisRequest(**message.data)
else:
request = TextualAnalysisRequest(document_content=str(message.data))
# Perform comprehensive textual analysis
analysis_result = await self._analyze_document(request, context)
# Generate insights and recommendations
insights = await self._generate_document_insights(analysis_result, request)
response_data = {
"document_id": analysis_result.document_id,
"timestamp": datetime.utcnow().isoformat(),
"agent": "machado",
"analysis_type": "textual_analysis",
"document_type": analysis_result.document_type.value,
"entities": {
"organizations": analysis_result.entities.organizations,
"values": analysis_result.entities.values,
"dates": analysis_result.entities.dates,
"people": analysis_result.entities.people,
"legal_references": analysis_result.entities.legal_references
},
"alerts": [
{
"type": alert.alert_type,
"excerpt": alert.excerpt,
"legal_violation": alert.legal_violation,
"severity": alert.severity.value,
"confidence": alert.confidence,
"explanation": alert.explanation
}
for alert in analysis_result.alerts
],
"metrics": {
"complexity_score": analysis_result.complexity_score,
"transparency_score": analysis_result.transparency_score,
"legal_compliance": analysis_result.legal_compliance,
"readability_grade": analysis_result.readability_grade
},
"suspicious_patterns": analysis_result.suspicious_patterns,
"insights": insights,
"checksum": analysis_result.checksum
}
self.logger.info(
"Textual analysis completed",
investigation_id=context.investigation_id,
document_type=analysis_result.document_type.value,
alerts_count=len(analysis_result.alerts),
transparency_score=analysis_result.transparency_score,
)
return AgentResponse(
agent_name=self.name,
response_type="textual_analysis",
data=response_data,
success=True,
context=context,
)
except Exception as e:
self.logger.error(
"Textual analysis failed",
investigation_id=context.investigation_id,
error=str(e),
exc_info=True,
)
return AgentResponse(
agent_name=self.name,
response_type="error",
data={"error": str(e), "analysis_type": "textual_analysis"},
success=False,
context=context,
)
async def _analyze_document(
self,
request: TextualAnalysisRequest,
context: AgentContext
) -> TextualAnalysisResult:
"""Perform comprehensive document analysis."""
self.logger.info(
"Starting textual analysis",
document_length=len(request.document_content),
document_type=request.document_type,
)
# Generate document ID
doc_id = hashlib.md5(request.document_content.encode()).hexdigest()[:12]
# Determine document type
doc_type = await self._classify_document_type(request.document_content)
# Extract entities using NER
entities = await self._extract_entities(request.document_content)
# Detect alerts and issues
alerts = await self._detect_document_alerts(request.document_content, doc_type)
# Calculate metrics
complexity = await self._calculate_complexity_score(request.document_content)
transparency = await self._calculate_transparency_score(request.document_content, entities)
compliance = await self._assess_legal_compliance(request.document_content, doc_type)
readability = await self._calculate_readability_grade(request.document_content)
# Detect suspicious patterns
suspicious = await self._detect_suspicious_patterns(request.document_content)
# Generate checksum
checksum = hashlib.md5(
f"{doc_id}{complexity}{transparency}{len(alerts)}".encode()
).hexdigest()
return TextualAnalysisResult(
document_id=doc_id,
document_type=doc_type,
entities=entities,
alerts=alerts,
complexity_score=complexity,
transparency_score=transparency,
legal_compliance=compliance,
readability_grade=readability,
suspicious_patterns=suspicious,
checksum=checksum,
analysis_timestamp=datetime.utcnow()
)
async def _classify_document_type(self, text: str) -> DocumentType:
"""Classify document type based on content patterns."""
text_lower = text.lower()
# Contract indicators
if any(keyword in text_lower for keyword in ["contrato", "contratação", "contratado"]):
return DocumentType.CONTRACT
# Public tender indicators
if any(keyword in text_lower for keyword in ["edital", "licitação", "pregão"]):
return DocumentType.PUBLIC_TENDER
# Law indicators
if any(keyword in text_lower for keyword in ["lei nº", "lei n°", "projeto de lei"]):
return DocumentType.LAW
# Decree indicators
if any(keyword in text_lower for keyword in ["decreto", "decreto nº"]):
return DocumentType.DECREE
# Default to contract if unsure
return DocumentType.CONTRACT
async def _extract_entities(self, text: str) -> EntityExtraction:
"""Extract named entities from document text."""
# Extract organizations (simplified)
organizations = []
org_patterns = [
r"(?:Ministério|Secretaria|Prefeitura|Câmara)\s+[\w\s]+",
r"(?:Empresa|Companhia|Sociedade)\s+[\w\s]+",
]
for pattern in org_patterns:
matches = re.findall(pattern, text, re.IGNORECASE)
organizations.extend(matches[:5]) # Limit to avoid clutter
# Extract monetary values
values = []
money_matches = re.findall(r"R\$\s*([\d,.]+)", text, re.IGNORECASE)
for match in money_matches[:10]: # Limit matches
try:
amount = float(match.replace(".", "").replace(",", "."))
values.append({
"amount": amount,
"context": f"Valor encontrado: R$ {match}"
})
except ValueError:
continue
# Extract dates
dates = []
date_patterns = [
r"(\d{1,2})/(\d{1,2})/(\d{4})",
r"(\d{1,2})\s+de\s+(\w+)\s+de\s+(\d{4})"
]
for pattern in date_patterns:
matches = re.findall(pattern, text)
for match in matches[:5]:
dates.append({
"date": "/".join(match) if "/" in pattern else " de ".join(match),
"event": "Data identificada no documento"
})
# Extract people names (simplified)
people = []
# This would need a proper NER model for better results
# Extract locations
locations = []
location_patterns = [
r"(?:Estado|Município)\s+(?:de|do|da)\s+([\w\s]+)",
r"(Brasília|São Paulo|Rio de Janeiro|Belo Horizonte)"
]
for pattern in location_patterns:
matches = re.findall(pattern, text, re.IGNORECASE)
locations.extend(matches[:5])
# Extract legal references
legal_refs = []
legal_patterns = [
r"Lei\s+n?º?\s*[\d./-]+",
r"Art\.?\s*\d+[º°]?",
r"CF/\d{2}",
]
for pattern in legal_patterns:
matches = re.findall(pattern, text, re.IGNORECASE)
legal_refs.extend(matches[:10])
return EntityExtraction(
organizations=list(set(organizations))[:10],
values=values,
dates=dates,
people=people,
locations=list(set(locations))[:5],
legal_references=list(set(legal_refs))[:10]
)
async def _detect_document_alerts(
self,
text: str,
doc_type: DocumentType
) -> List[DocumentAlert]:
"""Detect alerts and suspicious patterns in document."""
alerts = []
# Check for suspicious patterns
for pattern_name, pattern in self._suspicious_patterns.items():
matches = re.finditer(pattern, text, re.IGNORECASE)
for match in matches:
context_start = max(0, match.start() - 50)
context_end = min(len(text), match.end() + 50)
excerpt = text[context_start:context_end].strip()
alerts.append(DocumentAlert(
alert_type=pattern_name,
excerpt=excerpt,
legal_violation="Lei 8.666/93" if pattern_name in ["urgency_abuse", "exclusive_criteria"] else None,
severity=AlertSeverity.HIGH if pattern_name == "urgency_abuse" else AlertSeverity.MEDIUM,
confidence=0.75,
explanation=f"Padrão suspeito detectado: {pattern_name}",
recommendation="Revisar critérios e justificativas"
))
# Check for ambiguous language
ambiguous_terms = ["conforme", "adequado", "satisfatório", "apropriado", "razoável"]
for term in ambiguous_terms:
if term in text.lower() and text.lower().count(term) > 3:
alerts.append(DocumentAlert(
alert_type="ambiguity",
excerpt=f"Termo '{term}' usado frequentemente",
legal_violation=None,
severity=AlertSeverity.LOW,
confidence=0.6,
explanation=f"Uso excessivo de linguagem ambígua: '{term}'",
recommendation="Especificar critérios objetivos"
))
return alerts[:20] # Limit alerts
async def _calculate_complexity_score(self, text: str) -> float:
"""Calculate text complexity using adapted Flesch formula."""
sentences = len(re.findall(r'[.!?]+', text))
words = len(text.split())
syllables = sum(self._count_syllables(word) for word in text.split())
if sentences == 0 or words == 0:
return 1.0 # Maximum complexity
avg_sentence_length = words / sentences
avg_syllables_per_word = syllables / words
# Adapted Flesch formula for Portuguese
flesch_score = 248.835 - 1.015 * avg_sentence_length - 84.6 * avg_syllables_per_word
# Convert to 0-1 scale (higher = more complex)
complexity = max(0.0, min(1.0, (100 - flesch_score) / 100))
return round(complexity, 3)
def _count_syllables(self, word: str) -> int:
"""Count syllables in a Portuguese word (simplified)."""
vowels = "aeiouAEIOU"
count = 0
previous_was_vowel = False
for char in word:
if char in vowels:
if not previous_was_vowel:
count += 1
previous_was_vowel = True
else:
previous_was_vowel = False
return max(1, count) # At least one syllable
async def _calculate_transparency_score(
self,
text: str,
entities: EntityExtraction
) -> float:
"""Calculate document transparency score."""
score = 0.0
# Check for specific information
if entities.values: # Has monetary values
score += 0.3
if entities.dates: # Has specific dates
score += 0.2
if entities.organizations: # Identifies organizations
score += 0.2
if entities.legal_references: # References legal framework
score += 0.2
# Check for transparency indicators
transparency_indicators = [
"justificativa", "critério", "metodologia", "público",
"transparente", "acesso", "divulgação"
]
indicator_count = sum(1 for indicator in transparency_indicators
if indicator in text.lower())
score += min(0.1, indicator_count / len(transparency_indicators))
return round(min(1.0, score), 3)
async def _assess_legal_compliance(self, text: str, doc_type: DocumentType) -> float:
"""Assess legal compliance based on document type."""
compliance_score = 0.5 # Base score
# Check for required legal references based on document type
if doc_type in [DocumentType.CONTRACT, DocumentType.PUBLIC_TENDER]:
if "8.666" in text or "14.133" in text:
compliance_score += 0.3
if "art." in text.lower() or "artigo" in text.lower():
compliance_score += 0.2
# Check for common compliance issues
compliance_issues = [
("urgente", -0.1), # Unjustified urgency
("sigiloso", -0.1), # Inappropriate secrecy
("exclusivo", -0.1), # Exclusive criteria
]
for term, penalty in compliance_issues:
if term in text.lower():
compliance_score += penalty
return round(max(0.0, min(1.0, compliance_score)), 3)
async def _calculate_readability_grade(self, text: str) -> int:
"""Calculate readability grade level."""
sentences = len(re.findall(r'[.!?]+', text))
words = len(text.split())
if sentences == 0:
return 20 # Maximum difficulty
avg_sentence_length = words / sentences
# Simplified grade calculation
if avg_sentence_length <= 10:
return 6 # Elementary
elif avg_sentence_length <= 15:
return 8 # Middle school
elif avg_sentence_length <= 20:
return 12 # High school
else:
return 16 # College level
async def _detect_suspicious_patterns(self, text: str) -> List[str]:
"""Detect suspicious patterns in document."""
patterns_found = []
for pattern_name, pattern in self._suspicious_patterns.items():
if re.search(pattern, text, re.IGNORECASE):
patterns_found.append(pattern_name)
return patterns_found
async def _generate_document_insights(
self,
analysis: TextualAnalysisResult,
request: TextualAnalysisRequest
) -> List[Dict[str, Any]]:
"""Generate actionable insights from document analysis."""
insights = []
# Complexity insight
if analysis.complexity_score > 0.8:
insights.append({
"type": "complexity_warning",
"message": "Documento apresenta alta complexidade linguística",
"recommendation": "Simplificar linguagem para melhor compreensão pública",
"impact": "high"
})
# Transparency insight
if analysis.transparency_score < 0.5:
insights.append({
"type": "transparency_concern",
"message": "Documento apresenta baixo nível de transparência",
"recommendation": "Incluir mais detalhes específicos e referências",
"impact": "medium"
})
# Alert summary
if analysis.alerts:
high_severity_alerts = [a for a in analysis.alerts if a.severity.value >= 3]
if high_severity_alerts:
insights.append({
"type": "compliance_risk",
"message": f"Identificados {len(high_severity_alerts)} alertas de alta gravidade",
"recommendation": "Revisar e corrigir questões identificadas antes da publicação",
"impact": "critical"
})
return insights