Consumer

Protect customer data, deter abuse, screen content and verify consent in public-facing AI services.

Industry starter · Regulated SaaS

What it checks

Policies in the Consumer template, in the order the gateway runs them
PolicyOnChecksAction
prompt-injectionRequest4 attack patterns; similarity threshold 0.76Block
bot-detectorRequestMore than 8 similar requests in 60 seconds; similarity ≥ 0.94Block
pii-detectorRequestPayment-card data (PCI mode); 4 custom identifier patternsRedact
dlp-filterRequest5 blocked terms such as “full card number” and “password reset token”; 4 custom patterns; near-miss spellingsRedact
safety-filterRequestBuilt-in education patterns; 7 extra blocked phrases such as “self-harm” and “suicide”; near-miss spellingsBlock
gdpr-complianceRequestRequires the X-User-Consent-Token headerBlock
quality-scorerResponseAccuracy ≥ 0.76; relevancy ≥ 0.80; combined score ≥ 0.78; at least 120 characters and 2 sentencesFallback
audit-loggerRequestAdds an allow-only audit marker to the decisionRecord

Rate limit: 900 requests per minute across all gateway replicas.

Policy YAML
pack:
  id: "cfg_demo_consumer"
  name: "demo-consumer-trust"
  version: "0.2.0"
  enabled: true
  author: "local"
  description: "Consumer AI starter with abuse controls, privacy minimization, and consent-aware response quality"

# Public-facing assistants receive bursty traffic, so limits should stay consistent across replicas.
distributed_rate_limit:
  backend: valkey
  url_env: VERDICTAN_REDIS_URL
global_rate_limit:
  max_requests: 900
  window_seconds: 60

# The current schema-safe way to avoid raw retention is metadata-only history capture with blocked-event visibility.
history:
  capture:
    enabled: true
    mode: metadata_only
    include_blocked: true
    fail_open: true

policies:
  chain: [prompt-injection, bot-detector, pii-detector, dlp-filter, safety-filter, gdpr-compliance, quality-scorer, audit-logger]

policy:
  # Attackers often try to override refund or checkout rules, so request-boundary defense comes first.
  prompt-injection:
    embedding_threshold: 0.76
    attack_patterns: ["ignore.*store.*policy", "override.*refund.*rules", "reveal.*internal.*prompt", "bypass.*checkout.*restriction"]
    encoding: {decode_base64: true, normalize_unicode: true, detect_homoglyphs: true}
    boundaries: {enforce_delimiters: true, reject_fake_boundaries: true}

  # Fingerprinting helps distinguish scripted traffic from genuine customers before abuse drives cost or latency.
  bot-detector:
    fingerprint_fields: [user-agent, x-forwarded-for, body, model]
    profile_window_seconds: 60
    similarity_threshold: 0.94
    max_requests_per_window: 8
    action: block

  # Redact customer and payment identifiers rather than forcing users to restate their issue.
  pii-detector:
    action: redact
    pci_mode: true
    detect_patterns: ['(?i)\b(?:order|return|ticket)\s*#?\s*[0-9]{6,12}\b', '(?i)\b(?:customer|member|loyalty)\s*(?:id|number)?\s*[:#-]?\s*[A-Z0-9]{6,14}\b', '(?i)\b(?:gift\s*card|payment\s*token)\s*[:#-]?\s*[A-Z0-9_-]{8,32}\b', '(?i)\b(?:shipment|tracking)\s*[:#-]?\s*[A-Z0-9]{8,24}\b']
    redaction:
      marker_format: label
      include_metadata: true
      custom_markers: {generic_id: "[REDACTED-CUSTOMER-ID]"}

  # Customer-service prompts often include order and auth artifacts that should be minimized before provider transit.
  dlp-filter:
    detect_patterns: ['(?i)\bORD-[0-9]{8,12}\b', '(?i)\bCUST-[A-Z0-9]{6,12}\b', '(?i)\b(?:session|reset|auth)[_-]?token\b\s*[:=]?\s*[A-Za-z0-9._-]{12,}\b', '(?i)\b(?:gift\s*card\s*pin|otp)\b\s*[:=]?\s*[0-9]{4,8}\b']
    blocked_terms: ["full card number", "password reset token", "government id upload", "complete purchase history export", "customer support backoffice password"]
    action: redact
    fuzzy_matching: true
    max_distance: 1

  # Consumer experiences must not amplify harmful, abusive, or fraudulent instructions.
  safety-filter:
    mode: education
    block_if: ["self-harm", "suicide", "violent threat", "illegal drug instructions", "phishing kit", "refund fraud", "sexual content involving minors"]
    action: block
    fuzzy_matching: true
    max_distance: 1

  # Consent gating and erasure-request detection matter when personalization depends on lawful basis across multiple jurisdictions.
  gdpr-compliance:
    consent_required: true
    consent_header: X-User-Consent-Token
    data_categories: [personal, contact, behavioral, payment]

  # Consumer assistants need a moderate quality floor so answers stay useful without over-blocking support volume.
  quality-scorer:
    min_output_chars: 120
    min_sentences: 2
    thresholds: {min_aggregate: 0.78, min_relevancy: 0.80, min_accuracy: 0.76}
    weights: {relevancy: 0.6, accuracy: 0.4}
    failure_action:
      action: fallback
      fallback_message: "Sorry, I could not produce a safe and reliable customer response. Please retry or route this case to human support."
    industry_profiles: {consumer: {min_aggregate: 0.78, min_relevancy: 0.80, min_accuracy: 0.76}}

  # Keep the audit marker active so privacy, fraud, and safety decisions stay visible in downstream reviews.
  audit-logger: {}