Introduction

Guardrails Validators are critical tools used to ensure the quality and compliance of data processed through AI systems. They help maintain data integrity, protect user privacy, and ensure the generated content is appropriate and meaningful.

At SimplAI, we provide four essential validators:

1. Competitor Check
2. Detect PII
3. Gibberish Text
4. Toxic Language Check.

Validators

1. Competitor Check

The Competitor Check validator identifies and flags any references to competitors within the data. This is particularly useful in maintaining the integrity of brand-specific content and ensuring competitive information does not infiltrate proprietary datasets or outputs.

User Inputs

To configure the Competitor check validator, the user needs to provide the following inputs:

1. Select Field

   The user specifies whether the validator should be applied to the Input data or the Output data.

2. Competitor List

   The user provides a list of competitor names and related keywords to be flagged.

3. Specify Corrective Action

   The user specifies the corrective action to be taken when a competitor is detected

   • FIX: Replace the competitor name with an alternative or placeholder.
   • NOOP: No operation; only flag the content without making changes.
   • Exception: Throw an exception to halt processing and alert the user.

How it Works

• Database of Competitors: The system uses the provided list of competitors to check against the data.
• Text Analysis: During text processing, the validator scans the content for any mentions of these competitors based on the user-specified field (input or output).
• Flagging and Corrective Action: If a competitor name or related keyword is detected, the content is flagged and the specified corrective action is applied.

Use Cases

• Brand Content Creation: Ensures that generated marketing content does not inadvertently reference competitors.
• Internal Documentation: Keeps internal documents free from competitor mentions, maintaining focus and confidentiality.

2. Detect PII (Personally Identifiable Information)

The Detect PII validator identifies and flags any personally identifiable information within the data. This is crucial for compliance with privacy regulations such as GDPR, CCPA, and HIPAA.

User Inputs

To configure the Detect PII validator, the user needs to provide the following inputs:

1. Select Field:

   The user specifies whether the validator should be applied to the Input data or the Output data.

2. PII Entities:

   The user selects the types of PII to detect from the following list:

   • Email address
   • Phone number
   • IP Address
   • Location
   • Person
   • URL
   • PAN (Permanent Account Number)
   • Aadhaar
   • Vehicle registration

3. Specify Corrective Action:

   The user specifies the corrective action to be taken when PII is detected:

   • FIX: Mask or anonymize the detected PII.
   • NOOP: No operation; only flag the content without making changes.
   • Exception: Throw an exception to halt processing and alert the user.

How it Works

• Pattern Recognition: Uses predefined patterns to identify selected PII entities such as names, email addresses, phone numbers, social security numbers, etc.
• Machine Learning: Employs machine learning models trained on datasets containing PII to improve detection accuracy.
• Contextual Analysis: Analyzes the context in which data appears to distinguish between sensitive and non-sensitive information.
• Flagging and Corrective Action: If PII is detected, the content is flagged and the specified corrective action is applied.

Use Cases

• Data Processing: Ensures that datasets used for analysis do not contain PII, protecting user privacy.
• Customer Support: Filters PII from customer interactions to prevent unauthorized data exposure.

3. Gibberish Text

The Gibberish Text validator detects and filters out nonsensical or meaningless text. This helps maintain the quality and readability of content generated by AI systems.

User Inputs

To configure the Gibberish text validator, the user needs to provide the following inputs:

1. Select Field

   The user specifies whether the validator should be applied to the Input data or the Output data.

2. Threshold:

   The user sets a threshold value that determines the tolerance level for gibberish detection. A lower threshold will flag more content as gibberish, while a higher threshold will be more lenient.

3. Validation Method:

   • The user specifies the method of validation:
     • Sentence: Validates text on a sentence-by-sentence basis.
     • Full: Validates the entire text as a whole.

4. Specify Corrective Action:

   • The user specifies the corrective action to be taken when gibberish text is detected:
     • FIX: Replace or correct the gibberish text.
     • NOOP: No operation; only flag the content without making changes.
     • Exception: Throw an exception to halt processing and alert the user.

How it Works

• Lexical Analysis: Analyzes the text for known words and phrases, flagging sequences that do not conform to any recognized language patterns.
• Statistical Models: Uses statistical models to identify text that deviates significantly from normal language structures.
• Contextual Integrity: Ensures the content makes sense in the given context, flagging irrelevant or random sequences of characters.

Use Cases

• Content Generation: Ensures that AI-generated content is coherent and meaningful.
• Data Cleaning: Filters out irrelevant noise from datasets, improving the quality of data for analysis.

4. Toxic Language Check

The Toxic Language Check validator identifies and flags harmful or abusive language within the data. This is important for maintaining a safe and respectful environment, especially in user-generated content platforms.

User Inputs

1. Select Field

   The user specifies whether the validator should be applied to the Input data or the Output data.

2. Threshold

   The user sets a threshold value that determines the tolerance level for toxic language detection. A lower threshold will flag more content as toxic, while a higher threshold will be more lenient.

3. Validation Method:

   The user specifies the method of validation:

   • Sentence: Validates text on a sentence-by-sentence basis.
   • Full: Validates the entire text as a whole.

4. Specify Corrective Action:

   The user specifies the corrective action to be taken when toxic language is detected:

   • FIX: Replace or correct the toxic language.
   • NOOP: No operation; only flag the content without making changes.
   • Exception: Throw an exception to halt processing and alert the user.

How it Works

• Predefined Keywords: Uses a list of harmful and abusive words and phrases to scan the content.
• Sentiment Analysis: Employs sentiment analysis techniques to detect negative tones and hostile language.
• Machine Learning Models: Utilizes trained models to understand the context and severity of the language used.

Use Cases

• Community Management: Monitors and moderates user-generated content to prevent the spread of toxic language.
• Customer Interactions: Ensures that customer service communications remain respectful and professional.

Implementation

To implement these validators in your system, follow these general steps:

1. Configuration

Configure the validators based on your specific needs. Update competitor lists, customize PII patterns, and adjust sensitivity levels for gibberish and toxic language detection.

2. Testing

Conduct thorough testing to ensure the validators are functioning correctly. Use diverse datasets to evaluate performance and accuracy.

3. Monitoring

Continuously monitor the performance of the validators. Update patterns and models regularly to adapt to new threats and changes in language usage.

4. Review

Periodically review flagged content to assess the effectiveness of the validators and make necessary adjustments to improve accuracy.

By implementing these Guardrails Validators, you can significantly enhance the quality, compliance, and safety of your data and AI-generated content.