Deploy Content Moderation

When to Use

• Deploy Azure Content Safety resources and moderation pipeline
• Configure severity thresholds per content category
• Create and manage custom blocklists
• Set up input + output moderation for LLM applications
• Configure image and multi-modal moderation

Prerequisites

1. Azure CLI authenticated: az account show
2. Bicep CLI: az bicep version
3. Azure Content Safety resource (S0 tier for production)
4. Sample content for threshold calibration
5. Custom blocklist terms prepared (if applicable)

Step 1: Validate Infrastructure

az bicep lint -f infra/main.bicep
az bicep build -f infra/main.bicep

Verify resources:

• Azure Content Safety (S0 for production throughput)
• Azure Storage (flagged content archive)
• Azure Service Bus (async moderation queue)
• Azure Monitor (moderation metrics, alerts)
• Azure Key Vault (API keys)

Step 2: Deploy Azure Resources

az deployment group create \
  --resource-group $RESOURCE_GROUP \
  --template-file infra/main.bicep \
  --parameters infra/parameters.json

Step 3: Configure Category Thresholds

Severity levels: 0 (Safe) → 2 (Low) → 4 (Medium) → 6 (High) Configure per use case: consumer-facing apps use stricter thresholds (≥ 2), internal tools can use moderate (≥ 4).

Step 4: Create Custom Blocklists

# Create blocklist
python scripts/create_blocklist.py --name "company-terms" --terms blocklists/company-terms.txt

# Create competitor blocklist
python scripts/create_blocklist.py --name "competitor-names" --terms blocklists/competitors.txt

• Blocklist terms are exact-match (case-insensitive)
• Max 10,000 terms per blocklist, max 100 blocklists
• Use for: competitor names, internal projects, regulated terms

Step 5: Configure Moderation Pipeline

# Dual moderation: check BOTH user input AND model output
# Input → Content Safety → LLM → Content Safety → Output

# Input moderation (before sending to LLM)
input_result = content_safety_client.analyze_text(
    AnalyzeTextOptions(text=user_input, categories=[...], blocklist_names=["company-terms"])
)
if input_result.hate_result.severity >= threshold:
    return "I can't help with that request."

# Output moderation (before returning to user)
output_result = content_safety_client.analyze_text(
    AnalyzeTextOptions(text=llm_response, categories=[...])
)

Step 6: Configure Image Moderation (If Applicable)

• Enable image analysis for user-uploaded content
• Categories: same as text (Hate, Violence, SelfHarm, Sexual)
• Max image size: 4MB, supported formats: JPEG, PNG, GIF, BMP, TIFF, WEBP
• For multi-modal: moderate both text prompt AND generated image

Step 7: Set Up Monitoring & Alerts

• Alert on spike in blocked content (>10% of requests)
• Alert on false positive complaints
• Dashboard: block rate by category, severity distribution, latency
• Archive flagged content for human review (with PII redaction)

Step 8: Smoke Test

python scripts/test_moderation.py --text "Normal business query" --expect pass
python scripts/test_moderation.py --text "Harmful content example" --expect block
python scripts/test_blocklist.py --term "CompetitorName" --expect block
python scripts/test_image.py --image samples/safe_image.jpg --expect pass

Post-Deployment Verification

• [ ] All 4 categories detecting at configured thresholds
• [ ] Custom blocklists active and blocking
• [ ] Input + output dual moderation working
• [ ] Moderation latency < 200ms per request
• [ ] Monitoring alerts configured and tested
• [ ] Flagged content archive working
• [ ] False positive reporting mechanism in place

Troubleshooting