TatTwamAI / agents /tools /validation_tools.py
Jayashree Sridhar
included privateaatr in all tools
c5514db
raw
history blame
35.6 kB
"""
Response validation tools for ensuring safe and appropriate responses
"""
import re
from typing import Dict, List, Tuple, Optional, Any
from dataclasses import dataclass
import json
from transformers import pipeline
import torch
from pydantic import PrivateAttr
from crewai.tools import BaseTool
#from .base_tool import BaseTool
# @dataclass
class ValidationResult:
"""Result of validation check"""
is_valid: bool
issues: List[str]
warnings: List[str]
suggestions: List[str]
confidence: float
refined_text: Optional[str] = None
# class ValidationTools:
# """Tools for validating responses and ensuring safety"""
# def __init__(self, config):
# self.config = config
# # Initialize sentiment analyzer for tone checking
# self.sentiment_analyzer = pipeline(
# "sentiment-analysis",
# model="nlptown/bert-base-multilingual-uncased-sentiment",
# device=0 if torch.cuda.is_available() else -1
# )
# # Prohibited patterns for different categories
# self.prohibited_patterns = {
# 'medical': [
# r'\b(?:diagnos|prescrib|medicat|cure|treat|therap)\w*\b',
# r'\b(?:disease|illness|disorder|syndrome)\s+(?:is|are|can be)\b',
# r'\b(?:take|consume|dose|dosage)\s+\d+\s*(?:mg|ml|pill|tablet)',
# r'\b(?:medical|clinical|physician|doctor)\s+(?:advice|consultation|opinion)',
# ],
# 'legal': [
# r'\b(?:legal advice|lawsuit|sue|court|litigation)\b',
# r'\b(?:illegal|unlawful|crime|criminal|prosecut)\w*\b',
# r'\b(?:you should|must|have to)\s+(?:sign|agree|consent|contract)',
# r'\b(?:rights|obligations|liability|damages)\s+(?:are|include)\b',
# ],
# 'financial': [
# r'\b(?:invest|buy|sell|trade)\s+(?:stock|crypto|bitcoin|forex)\b',
# r'\b(?:guaranteed|promise)\s+(?:return|profit|income|earnings)\b',
# r'\b(?:financial advisor|investment advice|trading strategy)\b',
# r'\b(?:tax|accounting|financial planning)\s+(?:advice|consultation)',
# ],
# 'harmful': [
# r'\b(?:suicide|suicidal|kill\s+(?:your|my)self|end\s+(?:it|life))\b',
# r'\b(?:self[\-\s]?harm|hurt\s+(?:your|my)self|cutting)\b',
# r'\b(?:violence|violent|weapon|attack|assault)\b',
# r'\b(?:hate|discriminat|racist|sexist|homophobic)\b',
# ],
# 'absolute': [
# r'\b(?:always|never|every|all|none|no one|everyone)\s+(?:will|must|should|is|are)\b',
# r'\b(?:definitely|certainly|guaranteed|assured|promise)\b',
# r'\b(?:only way|only solution|must do|have to)\b',
# ]
# }
# # Required elements for supportive responses
# self.supportive_elements = {
# 'empathy': [
# 'understand', 'hear', 'feel', 'acknowledge', 'recognize',
# 'appreciate', 'empathize', 'relate', 'comprehend'
# ],
# 'validation': [
# 'valid', 'normal', 'understandable', 'natural', 'okay',
# 'reasonable', 'makes sense', 'legitimate'
# ],
# 'support': [
# 'support', 'help', 'here for you', 'together', 'alongside',
# 'assist', 'guide', 'accompany', 'with you'
# ],
# 'hope': [
# 'can', 'possible', 'able', 'capable', 'potential',
# 'opportunity', 'growth', 'improve', 'better', 'progress'
# ],
# 'empowerment': [
# 'choice', 'decide', 'control', 'power', 'strength',
# 'agency', 'capable', 'resource', 'ability'
# ]
# }
# # Crisis indicators
# self.crisis_indicators = [
# r'\b(?:want|going|plan)\s+to\s+(?:die|kill|end)\b',
# r'\b(?:no reason|point|hope)\s+(?:to|in)\s+(?:live|living|life)\b',
# r'\b(?:better off|world)\s+without\s+me\b',
# r'\bsuicide\s+(?:plan|method|attempt)\b',
# r'\b(?:final|last)\s+(?:goodbye|letter|message)\b'
# ]
# # Tone indicators
# self.negative_tone_words = [
# 'stupid', 'idiot', 'dumb', 'pathetic', 'worthless',
# 'loser', 'failure', 'weak', 'incompetent', 'useless'
# ]
# self.dismissive_phrases = [
# 'just get over it', 'stop complaining', 'not a big deal',
# 'being dramatic', 'overreacting', 'too sensitive'
# ]
# def validate_response(self, response: str, context: Dict[str, Any] = None) -> ValidationResult:
# """Comprehensive validation of response"""
# issues = []
# warnings = []
# suggestions = []
# # Check for prohibited content
# prohibited_check = self._check_prohibited_content(response)
# if prohibited_check["found"]:
# issues.extend(prohibited_check["violations"])
# suggestions.extend(prohibited_check["suggestions"])
# # Check tone and sentiment
# tone_check = self._check_tone(response)
# if not tone_check["appropriate"]:
# warnings.extend(tone_check["issues"])
# suggestions.extend(tone_check["suggestions"])
# # Check for supportive elements
# support_check = self._check_supportive_elements(response)
# if support_check["missing"]:
# warnings.append(f"Missing supportive elements: {', '.join(support_check['missing'])}")
# suggestions.extend(support_check["suggestions"])
# # Check for crisis content in context
# if context and context.get("user_input"):
# crisis_check = self._check_crisis_indicators(context["user_input"])
# if crisis_check["is_crisis"] and "crisis" not in response.lower():
# warnings.append("User may be in crisis but response doesn't address this")
# suggestions.append("Include crisis resources and immediate support options")
# # Calculate overall confidence
# confidence = self._calculate_confidence(issues, warnings)
# # Generate refined response if needed
# refined_text = None
# if issues or (warnings and confidence < 0.7):
# refined_text = self._refine_response(response, issues, warnings, suggestions)
# return ValidationResult(
# is_valid=len(issues) == 0,
# issues=issues,
# warnings=warnings,
# suggestions=suggestions,
# confidence=confidence,
# refined_text=refined_text
# )
# def _check_prohibited_content(self, text: str) -> Dict[str, Any]:
# """Check for prohibited content patterns"""
# found_violations = []
# suggestions = []
# for category, patterns in self.prohibited_patterns.items():
# for pattern in patterns:
# if re.search(pattern, text, re.IGNORECASE):
# found_violations.append(f"Contains {category} advice/content")
# # Add specific suggestions
# if category == "medical":
# suggestions.append("Replace with: 'Consider speaking with a healthcare professional'")
# elif category == "legal":
# suggestions.append("Replace with: 'For legal matters, consult with a qualified attorney'")
# elif category == "financial":
# suggestions.append("Replace with: 'For financial decisions, consider consulting a financial advisor'")
# elif category == "harmful":
# suggestions.append("Include crisis resources and express immediate concern for safety")
# elif category == "absolute":
# suggestions.append("Use qualifying language like 'often', 'might', 'could' instead of absolutes")
# break
# return {
# "found": len(found_violations) > 0,
# "violations": found_violations,
# "suggestions": suggestions
# }
# def _check_tone(self, text: str) -> Dict[str, Any]:
# """Check the tone and sentiment of the response"""
# issues = []
# suggestions = []
# # Check sentiment
# try:
# sentiment_result = self.sentiment_analyzer(text[:512])[0] # Limit length for model
# sentiment_score = sentiment_result['score']
# sentiment_label = sentiment_result['label']
# # Check if too negative
# if '1' in sentiment_label or '2' in sentiment_label: # 1-2 stars = negative
# issues.append("Response tone is too negative")
# suggestions.append("Add more supportive and hopeful language")
# except:
# pass
# # Check for negative words
# text_lower = text.lower()
# found_negative = [word for word in self.negative_tone_words if word in text_lower]
# if found_negative:
# issues.append(f"Contains negative/judgmental language: {', '.join(found_negative)}")
# suggestions.append("Replace judgmental terms with supportive language")
# # Check for dismissive phrases
# found_dismissive = [phrase for phrase in self.dismissive_phrases if phrase in text_lower]
# if found_dismissive:
# issues.append("Contains dismissive language")
# suggestions.append("Acknowledge and validate the person's feelings instead")
# return {
# "appropriate": len(issues) == 0,
# "issues": issues,
# "suggestions": suggestions
# }
# def _check_supportive_elements(self, text: str) -> Dict[str, Any]:
# """Check for presence of supportive elements"""
# text_lower = text.lower()
# missing_elements = []
# suggestions = []
# element_scores = {}
# for element, keywords in self.supportive_elements.items():
# found = any(keyword in text_lower for keyword in keywords)
# element_scores[element] = found
# if not found:
# missing_elements.append(element)
# # Generate suggestions for missing elements
# if 'empathy' in missing_elements:
# suggestions.append("Add empathetic language like 'I understand how difficult this must be'")
# if 'validation' in missing_elements:
# suggestions.append("Validate their feelings with phrases like 'Your feelings are completely valid'")
# if 'support' in missing_elements:
# suggestions.append("Express support with 'I'm here to support you through this'")
# if 'hope' in missing_elements:
# suggestions.append("Include hopeful elements about growth and positive change")
# if 'empowerment' in missing_elements:
# suggestions.append("Emphasize their agency and ability to make choices")
# return {
# "missing": missing_elements,
# "present": [k for k, v in element_scores.items() if v],
# "suggestions": suggestions
# }
# def _check_crisis_indicators(self, text: str) -> Dict[str, Any]:
# """Check for crisis indicators in text"""
# for pattern in self.crisis_indicators:
# if re.search(pattern, text, re.IGNORECASE):
# return {
# "is_crisis": True,
# "pattern_matched": pattern,
# "action": "Immediate crisis response needed"
# }
# return {"is_crisis": False}
# def _calculate_confidence(self, issues: List[str], warnings: List[str]) -> float:
# """Calculate confidence score for validation"""
# if issues:
# return 0.3 - (0.1 * len(issues)) # Major issues severely impact confidence
# confidence = 1.0
# confidence -= 0.1 * len(warnings) # Each warning reduces confidence
# return max(0.0, confidence)
# def _refine_response(self, response: str, issues: List[str], warnings: List[str], suggestions: List[str]) -> str:
# """Attempt to refine the response based on issues found"""
# refined = response
# # Add disclaimer for professional advice
# if any('advice' in issue for issue in issues):
# disclaimer = "\n\n*Please note: I'm here to provide support and guidance, but for specific professional matters, it's important to consult with qualified professionals.*"
# if disclaimer not in refined:
# refined += disclaimer
# # Add crisis resources if needed
# if any('crisis' in warning for warning in warnings):
# crisis_text = "\n\n**If you're in crisis, please reach out for immediate help:**\n- Crisis Hotline: 988 (US)\n- Crisis Text Line: Text HOME to 741741\n- International: findahelpline.com"
# if crisis_text not in refined:
# refined += crisis_text
# # Add supportive closing if missing hope
# if any('hope' in warning for warning in warnings):
# hopeful_closing = "\n\nRemember, you have the strength to navigate this challenge, and positive change is possible. I'm here to support you on this journey."
# if not any(phrase in refined.lower() for phrase in ['journey', 'strength', 'possible']):
# refined += hopeful_closing
# return refined
# def validate_user_input(self, text: str) -> ValidationResult:
# """Validate user input for safety and process-ability"""
# issues = []
# warnings = []
# suggestions = []
# # Check if empty
# if not text or not text.strip():
# issues.append("Empty input received")
# suggestions.append("Please share what's on your mind")
# return ValidationResult(False, issues, warnings, suggestions, 0.0)
# # Check length
# if len(text) > 5000:
# warnings.append("Input is very long")
# suggestions.append("Consider breaking this into smaller parts")
# # Check for crisis indicators
# crisis_check = self._check_crisis_indicators(text)
# if crisis_check["is_crisis"]:
# warnings.append("Crisis indicators detected")
# suggestions.append("Prioritize safety and provide crisis resources")
# # Check for spam/repetition
# if self._is_spam(text):
# issues.append("Input appears to be spam or repetitive")
# suggestions.append("Please share genuine thoughts or concerns")
# confidence = self._calculate_confidence(issues, warnings)
# return ValidationResult(
# is_valid=len(issues) == 0,
# issues=issues,
# warnings=warnings,
# suggestions=suggestions,
# confidence=confidence
# )
# def _is_spam(self, text: str) -> bool:
# """Simple spam detection"""
# # Check for excessive repetition
# words = text.lower().split()
# if len(words) > 10:
# unique_ratio = len(set(words)) / len(words)
# if unique_ratio < 0.3: # Less than 30% unique words
# return True
# # Check for common spam patterns
# spam_patterns = [
# r'(?:buy|sell|click|visit)\s+(?:now|here|this)',
# r'(?:congratulations|winner|prize|lottery)',
# r'(?:viagra|pills|drugs|pharmacy)',
# r'(?:$$|money\s+back|guarantee)'
# ]
# for pattern in spam_patterns:
# if re.search(pattern, text, re.IGNORECASE):
# return True
# return False
# def get_crisis_resources(self, location: str = "global") -> Dict[str, Any]:
# """Get crisis resources based on location"""
# resources = {
# "global": {
# "name": "International Association for Suicide Prevention",
# "url": "https://www.iasp.info/resources/Crisis_Centres/",
# "text": "Find crisis centers worldwide"
# },
# "us": {
# "name": "988 Suicide & Crisis Lifeline",
# "phone": "988",
# "text": "Text HOME to 741741",
# "url": "https://988lifeline.org/"
# },
# "uk": {
# "name": "Samaritans",
# "phone": "116 123",
# "email": "jo@samaritans.org",
# "url": "https://www.samaritans.org/"
# },
# "india": {
# "name": "National Suicide Prevention Helpline",
# "phone": "91-9820466726",
# "additional": "Vandrevala Foundation: 9999666555"
# },
# "australia": {
# "name": "Lifeline",
# "phone": "13 11 14",
# "text": "Text 0477 13 11 14",
# "url": "https://www.lifeline.org.au/"
# }
# }
# return resources.get(location.lower(), resources["global"])
#from .base_tool import BaseTool
#from crewai_tools import BaseTool
class ValidateResponseTool(BaseTool):
name: str = "validate_response"
description: str = "Validates safety and helpfulness."
model_config = {"arbitrary_types_allowed": True}
_config: object = PrivateAttr()
def __init__(self, config=None, **data):
super().__init__(**data)
self._config = config
# ... any required initialization ...
def _run(self, response: str, context: dict = None):
# Place your actual validation logic here, include dummy for illustration
# For full validation logic, use your own code!
# """Result of validation check"""
is_valid: bool
issues: List[str]
warnings: List[str]
suggestions: List[str]
confidence: float
refined_text: Optional[str] = None
return {"is_valid", "issues", "warnings", "suggestions","confidence","refined_text"}
class ValidationTools:
#_model: ValidateResponseTool = PrivateAttr()
def __init__(self, config=None):
self._validate_response = ValidateResponseTool(config)
# Add more tools as needed (check_safety, refine_response, etc.)
# # Initialize sentiment analyzer for tone checking
self.sentiment_analyzer = pipeline(
"sentiment-analysis",
model="nlptown/bert-base-multilingual-uncased-sentiment",
device=0 if torch.cuda.is_available() else -1
)
# Prohibited patterns for different categories
self.prohibited_patterns = {
'medical': [
r'\b(?:diagnos|prescrib|medicat|cure|treat|therap)\w*\b',
r'\b(?:disease|illness|disorder|syndrome)\s+(?:is|are|can be)\b',
r'\b(?:take|consume|dose|dosage)\s+\d+\s*(?:mg|ml|pill|tablet)',
r'\b(?:medical|clinical|physician|doctor)\s+(?:advice|consultation|opinion)',
],
'legal': [
r'\b(?:legal advice|lawsuit|sue|court|litigation)\b',
r'\b(?:illegal|unlawful|crime|criminal|prosecut)\w*\b',
r'\b(?:you should|must|have to)\s+(?:sign|agree|consent|contract)',
r'\b(?:rights|obligations|liability|damages)\s+(?:are|include)\b',
],
'financial': [
r'\b(?:invest|buy|sell|trade)\s+(?:stock|crypto|bitcoin|forex)\b',
r'\b(?:guaranteed|promise)\s+(?:return|profit|income|earnings)\b',
r'\b(?:financial advisor|investment advice|trading strategy)\b',
r'\b(?:tax|accounting|financial planning)\s+(?:advice|consultation)',
],
'harmful': [
r'\b(?:suicide|suicidal|kill\s+(?:your|my)self|end\s+(?:it|life))\b',
r'\b(?:self[\-\s]?harm|hurt\s+(?:your|my)self|cutting)\b',
r'\b(?:violence|violent|weapon|attack|assault)\b',
r'\b(?:hate|discriminat|racist|sexist|homophobic)\b',
],
'absolute': [
r'\b(?:always|never|every|all|none|no one|everyone)\s+(?:will|must|should|is|are)\b',
r'\b(?:definitely|certainly|guaranteed|assured|promise)\b',
r'\b(?:only way|only solution|must do|have to)\b',
]
}
# Required elements for supportive responses
self.supportive_elements = {
'empathy': [
'understand', 'hear', 'feel', 'acknowledge', 'recognize',
'appreciate', 'empathize', 'relate', 'comprehend'
],
'validation': [
'valid', 'normal', 'understandable', 'natural', 'okay',
'reasonable', 'makes sense', 'legitimate'
],
'support': [
'support', 'help', 'here for you', 'together', 'alongside',
'assist', 'guide', 'accompany', 'with you'
],
'hope': [
'can', 'possible', 'able', 'capable', 'potential',
'opportunity', 'growth', 'improve', 'better', 'progress'
],
'empowerment': [
'choice', 'decide', 'control', 'power', 'strength',
'agency', 'capable', 'resource', 'ability'
]
}
# Crisis indicators
self.crisis_indicators = [
r'\b(?:want|going|plan)\s+to\s+(?:die|kill|end)\b',
r'\b(?:no reason|point|hope)\s+(?:to|in)\s+(?:live|living|life)\b',
r'\b(?:better off|world)\s+without\s+me\b',
r'\bsuicide\s+(?:plan|method|attempt)\b',
r'\b(?:final|last)\s+(?:goodbye|letter|message)\b'
]
# Tone indicators
self.negative_tone_words = [
'stupid', 'idiot', 'dumb', 'pathetic', 'worthless',
'loser', 'failure', 'weak', 'incompetent', 'useless'
]
self.dismissive_phrases = [
'just get over it', 'stop complaining', 'not a big deal',
'being dramatic', 'overreacting', 'too sensitive'
]
def validate_response(self, response: str, context: Dict[str, Any] = None) -> ValidationResult:
"""Comprehensive validation of response"""
issues = []
warnings = []
suggestions = []
# Check for prohibited content
prohibited_check = self._check_prohibited_content(response)
if prohibited_check["found"]:
issues.extend(prohibited_check["violations"])
suggestions.extend(prohibited_check["suggestions"])
# Check tone and sentiment
tone_check = self._check_tone(response)
if not tone_check["appropriate"]:
warnings.extend(tone_check["issues"])
suggestions.extend(tone_check["suggestions"])
# Check for supportive elements
support_check = self._check_supportive_elements(response)
if support_check["missing"]:
warnings.append(f"Missing supportive elements: {', '.join(support_check['missing'])}")
suggestions.extend(support_check["suggestions"])
# Check for crisis content in context
if context and context.get("user_input"):
crisis_check = self._check_crisis_indicators(context["user_input"])
if crisis_check["is_crisis"] and "crisis" not in response.lower():
warnings.append("User may be in crisis but response doesn't address this")
suggestions.append("Include crisis resources and immediate support options")
# Calculate overall confidence
confidence = self._calculate_confidence(issues, warnings)
# Generate refined response if needed
refined_text = None
if issues or (warnings and confidence < 0.7):
refined_text = self._refine_response(response, issues, warnings, suggestions)
return ValidationResult(
is_valid=len(issues) == 0,
issues=issues,
warnings=warnings,
suggestions=suggestions,
confidence=confidence,
refined_text=refined_text
)
def _check_prohibited_content(self, text: str) -> Dict[str, Any]:
"""Check for prohibited content patterns"""
found_violations = []
suggestions = []
for category, patterns in self.prohibited_patterns.items():
for pattern in patterns:
if re.search(pattern, text, re.IGNORECASE):
found_violations.append(f"Contains {category} advice/content")
# Add specific suggestions
if category == "medical":
suggestions.append("Replace with: 'Consider speaking with a healthcare professional'")
elif category == "legal":
suggestions.append("Replace with: 'For legal matters, consult with a qualified attorney'")
elif category == "financial":
suggestions.append("Replace with: 'For financial decisions, consider consulting a financial advisor'")
elif category == "harmful":
suggestions.append("Include crisis resources and express immediate concern for safety")
elif category == "absolute":
suggestions.append("Use qualifying language like 'often', 'might', 'could' instead of absolutes")
break
return {
"found": len(found_violations) > 0,
"violations": found_violations,
"suggestions": suggestions
}
def _check_tone(self, text: str) -> Dict[str, Any]:
"""Check the tone and sentiment of the response"""
issues = []
suggestions = []
# Check sentiment
try:
sentiment_result = self.sentiment_analyzer(text[:512])[0] # Limit length for model
sentiment_score = sentiment_result['score']
sentiment_label = sentiment_result['label']
# Check if too negative
if '1' in sentiment_label or '2' in sentiment_label: # 1-2 stars = negative
issues.append("Response tone is too negative")
suggestions.append("Add more supportive and hopeful language")
except:
pass
# Check for negative words
text_lower = text.lower()
found_negative = [word for word in self.negative_tone_words if word in text_lower]
if found_negative:
issues.append(f"Contains negative/judgmental language: {', '.join(found_negative)}")
suggestions.append("Replace judgmental terms with supportive language")
# Check for dismissive phrases
found_dismissive = [phrase for phrase in self.dismissive_phrases if phrase in text_lower]
if found_dismissive:
issues.append("Contains dismissive language")
suggestions.append("Acknowledge and validate the person's feelings instead")
return {
"appropriate": len(issues) == 0,
"issues": issues,
"suggestions": suggestions
}
def _check_supportive_elements(self, text: str) -> Dict[str, Any]:
"""Check for presence of supportive elements"""
text_lower = text.lower()
missing_elements = []
suggestions = []
element_scores = {}
for element, keywords in self.supportive_elements.items():
found = any(keyword in text_lower for keyword in keywords)
element_scores[element] = found
if not found:
missing_elements.append(element)
# Generate suggestions for missing elements
if 'empathy' in missing_elements:
suggestions.append("Add empathetic language like 'I understand how difficult this must be'")
if 'validation' in missing_elements:
suggestions.append("Validate their feelings with phrases like 'Your feelings are completely valid'")
if 'support' in missing_elements:
suggestions.append("Express support with 'I'm here to support you through this'")
if 'hope' in missing_elements:
suggestions.append("Include hopeful elements about growth and positive change")
if 'empowerment' in missing_elements:
suggestions.append("Emphasize their agency and ability to make choices")
return {
"missing": missing_elements,
"present": [k for k, v in element_scores.items() if v],
"suggestions": suggestions
}
def _check_crisis_indicators(self, text: str) -> Dict[str, Any]:
"""Check for crisis indicators in text"""
for pattern in self.crisis_indicators:
if re.search(pattern, text, re.IGNORECASE):
return {
"is_crisis": True,
"pattern_matched": pattern,
"action": "Immediate crisis response needed"
}
return {"is_crisis": False}
def _calculate_confidence(self, issues: List[str], warnings: List[str]) -> float:
"""Calculate confidence score for validation"""
if issues:
return 0.3 - (0.1 * len(issues)) # Major issues severely impact confidence
confidence = 1.0
confidence -= 0.1 * len(warnings) # Each warning reduces confidence
return max(0.0, confidence)
def _refine_response(self, response: str, issues: List[str], warnings: List[str], suggestions: List[str]) -> str:
"""Attempt to refine the response based on issues found"""
refined = response
# Add disclaimer for professional advice
if any('advice' in issue for issue in issues):
disclaimer = "\n\n*Please note: I'm here to provide support and guidance, but for specific professional matters, it's important to consult with qualified professionals.*"
if disclaimer not in refined:
refined += disclaimer
# Add crisis resources if needed
if any('crisis' in warning for warning in warnings):
crisis_text = "\n\n**If you're in crisis, please reach out for immediate help:**\n- Crisis Hotline: 988 (US)\n- Crisis Text Line: Text HOME to 741741\n- International: findahelpline.com"
if crisis_text not in refined:
refined += crisis_text
# Add supportive closing if missing hope
if any('hope' in warning for warning in warnings):
hopeful_closing = "\n\nRemember, you have the strength to navigate this challenge, and positive change is possible. I'm here to support you on this journey."
if not any(phrase in refined.lower() for phrase in ['journey', 'strength', 'possible']):
refined += hopeful_closing
return refined
def validate_user_input(self, text: str) -> ValidationResult:
"""Validate user input for safety and process-ability"""
issues = []
warnings = []
suggestions = []
# Check if empty
if not text or not text.strip():
issues.append("Empty input received")
suggestions.append("Please share what's on your mind")
return ValidationResult(False, issues, warnings, suggestions, 0.0)
# Check length
if len(text) > 5000:
warnings.append("Input is very long")
suggestions.append("Consider breaking this into smaller parts")
# Check for crisis indicators
crisis_check = self._check_crisis_indicators(text)
if crisis_check["is_crisis"]:
warnings.append("Crisis indicators detected")
suggestions.append("Prioritize safety and provide crisis resources")
# Check for spam/repetition
if self._is_spam(text):
issues.append("Input appears to be spam or repetitive")
suggestions.append("Please share genuine thoughts or concerns")
confidence = self._calculate_confidence(issues, warnings)
return ValidationResult(
is_valid=len(issues) == 0,
issues=issues,
warnings=warnings,
suggestions=suggestions,
confidence=confidence
)
def _is_spam(self, text: str) -> bool:
"""Simple spam detection"""
# Check for excessive repetition
words = text.lower().split()
if len(words) > 10:
unique_ratio = len(set(words)) / len(words)
if unique_ratio < 0.3: # Less than 30% unique words
return True
# Check for common spam patterns
spam_patterns = [
r'(?:buy|sell|click|visit)\s+(?:now|here|this)',
r'(?:congratulations|winner|prize|lottery)',
r'(?:viagra|pills|drugs|pharmacy)',
r'(?:$$|money\s+back|guarantee)'
]
for pattern in spam_patterns:
if re.search(pattern, text, re.IGNORECASE):
return True
return False
def get_crisis_resources(self, location: str = "global") -> Dict[str, Any]:
"""Get crisis resources based on location"""
resources = {
"global": {
"name": "International Association for Suicide Prevention",
"url": "https://www.iasp.info/resources/Crisis_Centres/",
"text": "Find crisis centers worldwide"
},
"us": {
"name": "988 Suicide & Crisis Lifeline",
"phone": "988",
"text": "Text HOME to 741741",
"url": "https://988lifeline.org/"
},
"uk": {
"name": "Samaritans",
"phone": "116 123",
"email": "jo@samaritans.org",
"url": "https://www.samaritans.org/"
},
"india": {
"name": "National Suicide Prevention Helpline",
"phone": "91-9820466726",
"additional": "Vandrevala Foundation: 9999666555"
},
"australia": {
"name": "Lifeline",
"phone": "13 11 14",
"text": "Text 0477 13 11 14",
"url": "https://www.lifeline.org.au/"
}
}
return resources.get(location.lower(), resources["global"])