Fiddler Agent Integration

Overview

Fiddler AI provides enterprise-grade guardrails for AI applications, enabling you to build safe, compliant, and trustworthy AI systems. The AnswerAgentAI integration with Fiddler delivers comprehensive protection across three critical dimensions:

With this integration, you can:

Validate content safety across 11 dimensions (harmful, violent, unethical, illegal, sexual, racist, jailbreaking, harassing, hateful, sexist, roleplaying)
Detect and redact PII across 15+ entity types (SSN, credit card, email, phone, address, passport, driver's license, etc.)
Prevent hallucinations in RAG systems with faithfulness validation
Enforce compliance with HIPAA, GDPR, financial regulations, and industry standards
Customize thresholds per dimension and PII type for fine-grained control
Fail-open reliability with circuit breakers and <150ms latency

Fiddler's guardrails run in real-time as part of your AI workflows, automatically blocking unsafe content, redacting sensitive information, and flagging potential hallucinations before they reach users. This makes it ideal for customer-facing chatbots, internal knowledge bases, content moderation systems, and any AI application where safety and compliance are critical.

Quick Start

Obtaining Credentials

Fiddler Guardrails requires API credentials from your Fiddler account.

Required Credentials:

Fiddler API Key: Your authentication token
Fiddler API URL: Your guardrails endpoint (default: https://guardrails.cloud.fiddler.ai)

How to obtain:

Sign Up for Fiddler AI:
- Visit https://www.fiddler.ai/
- Click Get Started or Request Demo
- Complete the signup process or contact sales
Access API Keys:
- Log in to your Fiddler dashboard
- Navigate to Settings → API Keys
- Click Generate API Key
Copy Credentials:
- API Key: Copy the generated key (format: fiddler_api_key_...)
- API URL: Use default https://guardrails.cloud.fiddler.ai or your custom endpoint

Security Notice

The Fiddler API Key has full access to your guardrails services. Best practices:

Store credentials securely (use AnswerAgentAI credential system)
Never commit API keys to version control
Rotate keys regularly for security
Use organization-specific credentials for multi-tenant deployments

Documentation Reference: Fiddler Guardrails Documentation

Available Components

Auto-Generated

This section reflects the Fiddler Guardrails integration as implemented in packages/server/src/services/guardrails/. Last updated: December 4, 2025

This integration provides three core guardrail endpoints:

Safety Validation (11 Dimensions)

Endpoint: /v3/guardrails/ftl-safety

Description: Evaluates content across 11 safety dimensions using Fiddler's Fast Language Toolkit (FTL).

Safety Dimensions:

fdl_harmful - General harmful content
fdl_violent - Violence, gore, or aggressive behavior
fdl_unethical - Unethical behavior or advice
fdl_illegal - Illegal activities or instructions
fdl_sexual - Sexual or adult content
fdl_racist - Racist or discriminatory content
fdl_jailbreaking - Attempts to bypass AI safety controls
fdl_harassing - Harassment or bullying
fdl_hateful - Hate speech or derogatory content
fdl_sexist - Sexist or gender-discriminatory content
fdl_roleplaying - Unsafe roleplaying scenarios

Configuration:

Global Threshold (default: 0.1) - Higher scores indicate MORE unsafe content
Per-Dimension Thresholds - Override global threshold for specific dimensions
Per-Dimension Actions - block, warn, or continue per dimension

Use Cases:

Customer-facing chatbot content moderation
User-generated content screening
Internal AI application safety
Compliance with content policies

PII Detection & Redaction (15+ Types)

Endpoint: /v3/guardrails/sensitive-information

Description: Detects and optionally redacts personally identifiable information (PII) across 15+ entity types.

PII Types Detected:

PERSON - Names and identities
EMAIL - Email addresses
PHONE_NUMBER - Phone numbers (all formats)
ADDRESS - Physical addresses
US_SOCIAL_SECURITY_NUMBER - SSN
CREDIT_CARD - Credit card numbers
US_PASSPORT - Passport numbers
US_DRIVER_LICENSE - Driver's license numbers
US_BANK_NUMBER - Bank account numbers
CREDIT_CARD_EXPIRATION - Card expiration dates
PIN - PINs and security codes
IBAN_CODE - International bank accounts
SWIFT_CODE - SWIFT/BIC codes
USERNAME - Usernames and handles
AGE - Age information

Configuration:

Global Confidence Threshold (default: 0.8) - Higher scores indicate HIGHER confidence
Per-Type Confidence Thresholds - Override for specific PII types
Per-Type Actions - block, redact, replace, warn, or continue
Enabled Types - Optionally filter which PII types to check

Use Cases:

HIPAA compliance (healthcare data)
GDPR compliance (EU personal data)
Financial data protection
PCI-DSS compliance (credit card data)
General privacy protection

Faithfulness Validation (RAG Hallucination Detection)

Endpoint: /v3/guardrails/ftl-response-faithfulness

Description: Validates that AI-generated responses are faithful to source documents in RAG (Retrieval-Augmented Generation) systems.

How It Works:

Compares AI output against source context/documents
Uses Fiddler's Fast Faithfulness model
Returns faithfulness score (inverted scale)
Score < 0.005 = unfaithful (hallucination detected)
Score ≥ 0.005 = faithful (accurate response)

Configuration:

Threshold (default: 0.005) - Fiddler's recommended threshold
Action (default: warn) - Output validation never blocks, only warns

Use Cases:

RAG chatbots validation
Knowledge base accuracy
Document Q&A systems
Customer support automation
Legal/medical AI applications

Use Cases

Common Scenarios

1. Customer-Facing Chatbot Safety

Automatically validate all user inputs and AI responses for safety and PII before displaying to users.

Workflow:

User submits message to chatbot
Fiddler validates input for:
- Safety violations (block harmful/illegal content)
- PII detection (redact sensitive information)
AI generates response based on validated/redacted input
Fiddler validates output for:
- Safety (warn on unsafe output)
- PII (redact any leaked PII)
- Faithfulness (detect hallucinations if using RAG)
Display safe, redacted response to user

Benefits:

Protect users from harmful content
Maintain compliance with privacy regulations
Prevent data leaks
Build trust with safe, accurate AI

Time Saved: 20+ hours per week on manual content review

2. HIPAA-Compliant Healthcare Chatbot

Ensure medical chatbots never expose protected health information (PHI).

Workflow:

Configure Fiddler PII detection for medical data:
- SSN, address, phone (block actions)
- Medical record numbers (redact)
- Names and DOB (redact)
Enable safety validation:
- Block harmful medical advice
- Warn on unethical suggestions
Enable faithfulness checking:
- Validate responses against medical knowledge base
- Flag potential hallucinations for review

Benefits:

HIPAA compliance
Prevent PHI leaks
Ensure accurate medical information
Reduce legal liability

Compliance Impact: Reduce PII incidents by 95%

3. Financial Services Content Moderation

Protect financial data and ensure regulatory compliance.

Workflow:

Detect and redact financial PII:
- Credit card numbers (block)
- Bank account numbers (block)
- SSN (block)
- PINs and security codes (block)
Safety validation:
- Block illegal financial advice
- Warn on unethical suggestions
Faithfulness checking:
- Validate responses against financial policies
- Prevent misinformation

Benefits:

PCI-DSS compliance
SOC 2 compliance
Regulatory compliance (SEC, FINRA)
Data breach prevention

Time Saved: 15+ hours per week on compliance reviews

4. RAG Knowledge Base with Hallucination Detection

Ensure AI-generated answers are grounded in source documents.

Workflow:

User asks question
RAG system retrieves relevant documents
AI generates response based on documents
Fiddler validates faithfulness:
- Compare response to source context
- Score < 0.005 = flag as potential hallucination
- Score ≥ 0.005 = accept as faithful
If unfaithful, trigger review or regenerate

Benefits:

Prevent hallucinations
Improve accuracy
Build user trust
Reduce misinformation

Accuracy Improvement: 85% reduction in hallucinations

5. Internal Knowledge Base Safety

Protect employees from unsafe or inappropriate content in internal AI tools.

Workflow:

Configure lenient safety thresholds for internal use:
- Higher thresholds (0.15) for internal content
- Warn actions instead of blocking
- Monitor but don't prevent roleplaying scenarios
PII detection:
- Warn on employee PII (not block)
- Track PII exposure for audit
Faithfulness checking:
- Validate responses against internal docs
- Flag low-confidence answers

Benefits:

Employee safety
Internal compliance
Audit trail for governance
Flexibility for internal use

Time Saved: 10+ hours per week on content review

Example Workflows

Example 1: Customer Support Chatbot with Full Guardrails

Goal: Create a safe, compliant customer support chatbot that never leaks PII or produces harmful content.

AnswerAgentAI Configuration:

// Organization Guardrails Config
{
  enabled: true,
  credentialId: "fiddler-credential-id",
  safety: {
    enabled: true,
    threshold: 0.1,  // Fiddler recommendation
    action: "block",
    dimensionActions: {
      fdl_illegal: "block",
      fdl_harmful: "block",
      fdl_sexual: "block",
      fdl_roleplaying: "warn"  // Less strict
    }
  },
  pii: {
    enabled: true,
    confidenceThreshold: 0.8,
    action: "redact",
    typeActions: {
      US_SOCIAL_SECURITY_NUMBER: "block",
      CREDIT_CARD: "block",
      EMAIL: "redact",
      PHONE_NUMBER: "redact",
      USERNAME: "warn"
    }
  },
  faithfulness: {
    enabled: true,
    threshold: 0.005,
    action: "warn"
  }
}

Example Interaction:

User: "My SSN is 123-45-6789 and I need help with my account"

Input Validation:
- Safety: ✓ No violations
- PII Detection: ✗ SSN detected (confidence: 0.95)
  Action: BLOCK (per typeActions config)

Response: "I cannot process messages containing sensitive information
like Social Security Numbers. Please remove this information and try again."

---

User: "I need help with my account login issues"

Input Validation:
- Safety: ✓ No violations
- PII: ✓ No PII detected

AI generates response based on RAG knowledge base...

Output Validation:
- Safety: ✓ No violations
- PII: ✓ No PII detected
- Faithfulness: ✓ Score 0.92 (faithful to source docs)

Response: "I can help you with login issues. First, let's try resetting your password..."

Result: Safe, compliant chatbot with <150ms guardrails latency.

Goal: Analyze user-generated content while ensuring GDPR compliance.

Configuration:

{
  pii: {
    enabled: true,
    confidenceThreshold: 0.75,  // More sensitive
    enabledTypes: [
      "PERSON",
      "EMAIL",
      "PHONE_NUMBER",
      "ADDRESS",
      "USERNAME"
    ],
    typeActions: {
      PERSON: "redact",
      EMAIL: "redact",
      PHONE_NUMBER: "redact",
      ADDRESS: "redact",
      USERNAME: "redact"
    }
  }
}

Example:

Input: "Contact John Doe at john.doe@example.com or call 555-1234"

PII Detection:
- PERSON: "John Doe" (confidence: 0.92) → Redact
- EMAIL: "john.doe@example.com" (confidence: 0.98) → Redact
- PHONE_NUMBER: "555-1234" (confidence: 0.88) → Redact

Redacted Output: "Contact [PERSON] at [EMAIL] or call [PHONE_NUMBER]"

Benefits:

GDPR compliance
Privacy protection
Audit trail
Automated redaction

Example 3: Medical Chatbot with Hallucination Prevention

Goal: Build a medical Q&A chatbot that never hallucinates medical information.

Configuration:

{
  safety: {
    enabled: true,
    threshold: 0.05,  // Strict for medical
    action: "block",
    dimensionActions: {
      fdl_harmful: "block",
      fdl_illegal: "block",
      fdl_unethical: "block"
    }
  },
  faithfulness: {
    enabled: true,
    threshold: 0.005,
    action: "warn"
  }
}

Workflow:

User: "What are the side effects of aspirin?"

RAG System:
1. Retrieves medical documents about aspirin
2. AI generates response based on documents

Faithfulness Validation:
- Input: Retrieved medical context
- Output: AI-generated response
- Faithfulness Score: 0.92 ✓ (faithful)

Response delivered to user.

---

If faithfulness score < 0.005:
- Log hallucination warning
- Optionally regenerate response
- Flag for human review

Benefits:

Medical accuracy
Patient safety
Liability reduction
Trust building

Advanced Configuration

Configuration Hierarchy

AnswerAgentAI supports a three-tier configuration hierarchy for guardrails:

Priority Order: Environment Variables → Organization Config → Chatflow Config

Environment Variables (Lowest Priority)

Set global defaults via environment variables:

# In .env
FIDDLER_API_KEY=fiddler_api_key_...
FIDDLER_API_URL=https://guardrails.cloud.fiddler.ai

Organization Config (Medium Priority)

Configure guardrails at the organization level:

// In Organization.organizationConfig
{
  guardrails: {
    enabled: true,
    credentialId: "org-fiddler-credential",
    safety: {
      enabled: true,
      threshold: 0.1,
      action: "block"
    },
    pii: {
      enabled: true,
      confidenceThreshold: 0.8,
      action: "redact"
    }
  }
}

Chatflow Config (Highest Priority)

Override organization settings per chatflow:

// In ChatFlow.chatflowConfig
{
  guardrails: {
    safety: {
      threshold: 0.05,  // More strict for this chatflow
      dimensionActions: {
        fdl_illegal: "block",
        fdl_harmful: "block"
      }
    }
  }
}

Safety Configuration Examples

Strict Safety (External Chatbots)

{
    "safety": {
        "enabled": true,
        "threshold": 0.05,
        "action": "block",
        "dimensionThresholds": {
            "fdl_illegal": 0.02,
            "fdl_harmful": 0.05,
            "fdl_sexual": 0.05
        },
        "dimensionActions": {
            "fdl_illegal": "block",
            "fdl_harmful": "block",
            "fdl_sexual": "block",
            "fdl_roleplaying": "warn"
        }
    }
}

Interpretation:

Global threshold: 0.05 (strict)
Illegal content: 0.02 (very strict)
Block most violations
Warn on roleplaying (more lenient)

Balanced Safety (General Purpose)

{
    "safety": {
        "enabled": true,
        "threshold": 0.1,
        "action": "block",
        "dimensionActions": {
            "fdl_illegal": "block",
            "fdl_harmful": "warn",
            "fdl_roleplaying": "continue"
        }
    }
}

Lenient Safety (Internal Tools)

{
    "safety": {
        "enabled": true,
        "threshold": 0.15,
        "action": "warn"
    }
}

PII Configuration Examples

HIPAA Compliance

{
    "pii": {
        "enabled": true,
        "confidenceThreshold": 0.75,
        "enabledTypes": ["PERSON", "US_SOCIAL_SECURITY_NUMBER", "ADDRESS", "PHONE_NUMBER", "EMAIL", "AGE"],
        "typeActions": {
            "US_SOCIAL_SECURITY_NUMBER": "block",
            "PERSON": "redact",
            "ADDRESS": "redact",
            "PHONE_NUMBER": "redact",
            "EMAIL": "redact",
            "AGE": "redact"
        }
    }
}

PCI-DSS Compliance (Financial)

{
    "pii": {
        "enabled": true,
        "confidenceThreshold": 0.8,
        "enabledTypes": ["CREDIT_CARD", "CREDIT_CARD_EXPIRATION", "US_BANK_NUMBER", "PIN", "SWIFT_CODE", "IBAN_CODE"],
        "typeActions": {
            "CREDIT_CARD": "block",
            "PIN": "block",
            "US_BANK_NUMBER": "block",
            "CREDIT_CARD_EXPIRATION": "redact",
            "SWIFT_CODE": "redact",
            "IBAN_CODE": "redact"
        }
    }
}

{
    "pii": {
        "enabled": true,
        "confidenceThreshold": 0.8,
        "enabledTypes": ["PERSON", "EMAIL", "PHONE_NUMBER", "ADDRESS"],
        "typeActions": {
            "PERSON": "redact",
            "EMAIL": "redact",
            "PHONE_NUMBER": "redact",
            "ADDRESS": "redact"
        }
    }
}

Circuit Breaker Configuration

Fiddler integration includes circuit breaker protection for reliability:

{
    "circuitBreaker": {
        "failureThreshold": 5,
        "resetTimeout": 30000,
        "successThreshold": 3
    }
}

Configuration:

failureThreshold - Number of failures before opening circuit (default: 5)
resetTimeout - Time in ms before attempting to close circuit (default: 30s)
successThreshold - Successes needed to close circuit (default: 3)

Behavior:

Closed: Normal operation, requests pass through
Open: After N failures, requests fail immediately (fail-open: return safe defaults)
Half-Open: After timeout, allow test requests to check if service recovered

Caching Configuration

Optional Redis caching for improved performance:

{
    "cache": {
        "enabled": true,
        "ttl": 3600
    }
}

Benefits:

Reduce API calls for identical inputs
Sub-millisecond response for cached results
Reduce costs
Improve latency

TTL: Time-to-live in seconds (default: 1 hour)

Frequently Asked Questions

Setup & Configuration

Q: Do I need a Fiddler account to use guardrails?

A: Yes, you need a Fiddler AI account and API key. Visit https://www.fiddler.ai/ to sign up or request a demo.

Q: What's the difference between input and output validation?

Input Validation: Validates user-submitted text before AI processing
- Can block unsafe inputs (prevent processing)
- Can redact PII before sending to LLM
- Default action: block for safety, redact for PII
Output Validation: Validates AI-generated responses before showing to users
- Never blocks (always fail-open)
- Can redact PII in responses
- Can warn on faithfulness issues (hallucinations)
- Default action: warn for all violations

Q: Can I configure different guardrails for different chatflows?

A: Yes! AnswerAgentAI supports a three-tier hierarchy:

Environment: Global defaults
Organization: Org-wide settings
Chatflow: Per-chatflow overrides (highest priority)

Each tier can override specific settings from lower tiers.

Q: How do I set up guardrails for a single chatflow?

A: Configure chatflowConfig on the ChatFlow entity:

{
  guardrails: {
    safety: { threshold: 0.05 },
    pii: { action: "block" }
  }
}

Q: What happens if my Fiddler API key is invalid?

A: The integration uses fail-open design:

Invalid credentials → Log warning, continue without guardrails
API errors → Return safe defaults (no violations detected)
Circuit breaker open → Skip guardrails, allow requests through

This ensures your application never goes down due to guardrails service issues.

Usage & Best Practices

Q: How do safety thresholds work?

A: Fiddler safety scores indicate how unsafe content is:

Score > threshold = Unsafe (violation detected)
Score ≤ threshold = Safe (no violation)
Lower threshold = MORE strict (catches more violations)
Higher threshold = LESS strict (catches fewer violations)
Fiddler recommendation: 0.1 for production

Example:

Threshold: 0.1
Score 0.05: Safe ✓
Score 0.15: Unsafe ✗ (violation)

Q: How do PII confidence thresholds work?

A: PII scores indicate how confident the detection is:

Score > threshold = PII detected (process it)
Score ≤ threshold = Not PII (ignore)
Lower threshold = MORE sensitive (more detections)
Higher threshold = LESS sensitive (fewer detections)
Fiddler recommendation: 0.8 for production

Example:

Threshold: 0.8
Score 0.95: High confidence PII ✓ (redact)
Score 0.70: Low confidence ✗ (ignore)

Q: How do I set different actions for different safety dimensions?

A: Use dimensionActions in safety config:

{
    "safety": {
        "action": "block",
        "dimensionActions": {
            "fdl_illegal": "block",
            "fdl_harmful": "warn",
            "fdl_roleplaying": "continue"
        }
    }
}

Q: How do I set different actions for different PII types?

A: Use typeActions in PII config:

{
    "pii": {
        "action": "redact",
        "typeActions": {
            "US_SOCIAL_SECURITY_NUMBER": "block",
            "CREDIT_CARD": "block",
            "EMAIL": "redact",
            "USERNAME": "warn"
        }
    }
}

Q: What actions are available?

block - Reject input/output, return error message
redact - Replace PII with [PII_TYPE] placeholder
replace - Replace with alternative text (future)
warn - Log violation, allow content through
continue - No action (monitoring only)

Q: How does faithfulness checking work?

A: Fiddler compares AI output against source context (RAG documents):

Input: Source documents/context
Output: AI-generated response
Returns: Faithfulness score (inverted scale)
Score < 0.005 = Unfaithful (hallucination)
Score ≥ 0.005 = Faithful (accurate)

Note: Faithfulness only works for RAG systems with source context.

Q: Should I enable faithfulness checking for all chatflows?

A: No, only for RAG systems:

Enable: RAG chatbots, Q&A systems, document search
Disable: Creative writing, general conversation, non-RAG use cases

Q: How do I monitor guardrails performance?

A: Guardrails metadata is included in chatflow responses:

{
    "text": "AI response",
    "guardrails": {
        "safety": {
            "violations": [...],
            "isUnsafe": false
        },
        "pii": {
            "detections": [...],
            "hasPII": false
        },
        "faithfulness": {
            "score": 0.92,
            "threshold": 0.005
        }
    }
}

Monitor this data in logs or analytics.

Troubleshooting

Q: Guardrails are not running

Causes:

Not enabled: Check enabled: true in config
No credentials: Verify Fiddler API key exists
Circuit breaker open: Too many failures opened circuit

Solutions:

Verify configuration hierarchy (env → org → chatflow)
Check credential exists and is valid
Check circuit breaker status in logs
Verify FIDDLER_API_KEY environment variable if using env config

Q: Getting "Circuit breaker is open" errors

Cause: Too many consecutive failures to Fiddler API

Solutions:

Check Fiddler API status (service outage?)
Verify API key is valid
Check network connectivity
Wait for reset timeout (default: 30s)
Manually reset circuit (admin action)

Q: PII detection not working

Possible causes:

Not enabled: pii.enabled: false
Threshold too high: confidenceThreshold: 0.95 misses lower-confidence detections
Type not enabled: Check enabledTypes array
Action is "continue": Detection happens but no action taken

Solutions:

Enable PII detection: pii.enabled: true
Lower threshold: Try 0.75 or 0.8
Check enabledTypes includes the PII type
Use redact or block action instead of continue

Q: Safety validation blocking too much

Cause: Threshold too low (too strict)

Solutions:

Increase global threshold: 0.1 → 0.15
Use per-dimension thresholds for fine-tuning
Change action from block to warn for less critical dimensions
Review blocked content to calibrate thresholds

Q: Safety validation not blocking enough

Cause: Threshold too high (too lenient)

Solutions:

Decrease global threshold: 0.15 → 0.1 or 0.05
Use per-dimension thresholds to be strict on specific dimensions
Ensure action is block not warn

Q: Faithfulness scores always 1.0 (no hallucinations detected)

Causes:

No context provided: Faithfulness requires source context
Faithfulness not enabled: faithfulness.enabled: false
Using non-RAG chatflow: Faithfulness only works for RAG systems

Solutions:

Ensure you're using RAG chatflow with source documents
Enable faithfulness: faithfulness.enabled: true
Verify context is passed to validation

Q: High latency (>200ms)

Causes:

Multiple sequential checks: Running safety, PII, faithfulness sequentially
No caching: Duplicate requests hitting API
Network issues: Slow connection to Fiddler API

Solutions:

AnswerAgentAI runs checks in parallel (already optimized)
Enable Redis caching: cache.enabled: true
Check network latency to Fiddler API
Consider increasing timeouts if network is slow

Q: Redacted text looks wrong

Example: "Contact [PERSON] at [EMAIL] or [PHONE_NUMBER]"

This is expected behavior. Redaction replaces PII with [PII_TYPE] placeholders.

To customize redaction:

Use replace action (future feature)
Post-process redacted text in your application
Use different actions per type (e.g., block SSN, redact email)

Q: Different results between input and output validation

This is expected. Input and output validation have different goals:

Input: Strict (blocks unsafe/PII inputs)
Output: Lenient (warns but doesn't block)

Differences in detection can occur due to:

Different thresholds configured
Different actions configured
Context differences (output includes AI-generated text)

Performance & Reliability

Q: What's the expected latency?

A: <150ms for most guardrails checks:

Safety: ~50-80ms
PII: ~40-70ms
Faithfulness: ~60-100ms
Parallel execution: All checks run simultaneously

Q: What happens if Fiddler API is down?

A: Fail-open design ensures your application stays online:

Circuit breaker opens after 5 failures
Requests skip guardrails (return safe defaults)
Circuit attempts to close after 30s
Your chatflows continue working without guardrails

Logged behavior:

Warning logged to console
Guardrails metadata indicates "fail-open" state
Monitor for service restoration

Q: Can I use Fiddler guardrails in production?

A: Yes, Fiddler is enterprise-grade and production-ready:

<150ms latency
Circuit breaker protection
Fail-open reliability
Connection pooling
Optional caching
99.9%+ uptime SLA (check with Fiddler)

Q: How do I optimize performance?

Enable caching: cache.enabled: true (sub-ms for cached results)
Parallel execution: AnswerAgentAI runs checks in parallel (already optimized)
Selective checks: Disable unused checks (e.g., faithfulness for non-RAG)
Connection pooling: Already enabled (50 sockets per host)
Per-dimension config: Only check dimensions you care about

Resources

Official Documentation

Fiddler Documentation - Complete Fiddler platform documentation
Guardrails Overview - Introduction to Fiddler Guardrails
Safety Guardrails - Safety dimension documentation
PII Detection - PII detection and redaction guide
Faithfulness Validation - RAG hallucination detection

AnswerAgentAI Implementation

Guardrails Service - Internal implementation reference
Guardrails Types - TypeScript type definitions
Configuration Guide - Server configuration patterns

Compliance Resources

HIPAA Compliance - Healthcare data protection
GDPR Compliance - EU data privacy regulations
PCI-DSS - Payment card industry standards
SOC 2 - Security and compliance framework

Community & Support

Fiddler Website - Company website
Request Demo - Book a Fiddler demo
Support - Fiddler customer support
Blog - Latest updates and best practices

Overview​

Quick Start​

Obtaining Credentials​

Available Components​

Safety Validation (11 Dimensions)​

PII Detection & Redaction (15+ Types)​

Faithfulness Validation (RAG Hallucination Detection)​

Use Cases​

Common Scenarios​

1. Customer-Facing Chatbot Safety​

2. HIPAA-Compliant Healthcare Chatbot​

3. Financial Services Content Moderation​

4. RAG Knowledge Base with Hallucination Detection​

5. Internal Knowledge Base Safety​

Example Workflows​

Example 1: Customer Support Chatbot with Full Guardrails​

Example 2: GDPR-Compliant Content Analysis​

Example 3: Medical Chatbot with Hallucination Prevention​

Advanced Configuration​

Configuration Hierarchy​

Environment Variables (Lowest Priority)​

Organization Config (Medium Priority)​

Chatflow Config (Highest Priority)​

Safety Configuration Examples​

Strict Safety (External Chatbots)​

Balanced Safety (General Purpose)​

Lenient Safety (Internal Tools)​

PII Configuration Examples​

HIPAA Compliance​

PCI-DSS Compliance (Financial)​

GDPR Compliance​

Circuit Breaker Configuration​

Caching Configuration​

Frequently Asked Questions​

Setup & Configuration​

Usage & Best Practices​

Troubleshooting​

Performance & Reliability​

Resources​

Official Documentation​

AnswerAgentAI Implementation​

Compliance Resources​

Community & Support​

Ask Alpha

Overview

Quick Start

Obtaining Credentials

Available Components

Safety Validation (11 Dimensions)

PII Detection & Redaction (15+ Types)

Faithfulness Validation (RAG Hallucination Detection)

Use Cases

Common Scenarios

1. Customer-Facing Chatbot Safety

2. HIPAA-Compliant Healthcare Chatbot

3. Financial Services Content Moderation

4. RAG Knowledge Base with Hallucination Detection

5. Internal Knowledge Base Safety

Example Workflows

Example 1: Customer Support Chatbot with Full Guardrails

Example 2: GDPR-Compliant Content Analysis

Example 3: Medical Chatbot with Hallucination Prevention

Advanced Configuration

Configuration Hierarchy

Environment Variables (Lowest Priority)

Organization Config (Medium Priority)

Chatflow Config (Highest Priority)

Safety Configuration Examples

Strict Safety (External Chatbots)

Balanced Safety (General Purpose)

Lenient Safety (Internal Tools)

PII Configuration Examples

HIPAA Compliance

PCI-DSS Compliance (Financial)

GDPR Compliance

Circuit Breaker Configuration

Caching Configuration

Frequently Asked Questions

Setup & Configuration

Usage & Best Practices

Troubleshooting

Performance & Reliability

Resources

Official Documentation

AnswerAgentAI Implementation

Compliance Resources

Community & Support