Well-Architected Framework

FrootAI aligns every primitive, play, and protocol element to the 6 pillars of the Azure Well-Architected Framework (WAF). This isn’t optional — WAF alignment is enforced at the protocol level via fai-manifest.json.

The 6 Pillars

Security

Every FrootAI solution enforces:

• Never hardcode secrets — use Azure Managed Identity and Key Vault
• RBAC with least-privilege — Microsoft Entra ID for user authentication
• Private endpoints for all PaaS services in production
• Content safety filters on all AI endpoints
• Rate limiting AI API calls per user/tenant
• Input sanitization — validate and sanitize all prompts before sending to models

{
  "context": {
    "waf": ["security"]
  }
}

Reliability

• All external API calls must have retry logic with exponential backoff (3 retries: 1s/2s/4s with jitter)
• Every service must expose a /health endpoint verifying downstream dependencies
• If Azure OpenAI is unavailable, fall back to cached responses or static content
• HTTP client timeouts: 30s for AI endpoints, 10s for search, 5s for metadata

{
  "thresholds": {
    "coherence": 0.90,
    "groundedness": 0.95
  }
}

Cost Optimization

• Use model routing: GPT-4o-mini for simple tasks, GPT-4o for complex reasoning
• Implement token budgets per request via max_tokens in config
• Cache frequent AI responses with TTL-based semantic deduplication
• Set costPerQuery guardrails in fai-manifest.json
• Default to the smallest viable SKU — scale up based on metrics, not assumptions

{
  "model": "gpt-4o",
  "max_tokens": 4096,
  "fallback_model": "gpt-4o-mini"
}

Operational Excellence

• All deployments must go through CI/CD pipelines — no manual deployments
• Use conventional commits (feat:, fix:, docs:, chore:)
• All infrastructure must be defined in Bicep/Terraform — no portal clicks
• Structured logging with correlation IDs across all services
• Application Insights for APM, distributed tracing, and custom AI metrics

# Validate consistency before every release
npm run validate:primitives
node engine/index.js fai-manifest.json --status

Performance Efficiency

• Target: < 3s for simple queries, < 10s for complex multi-step reasoning
• Use streaming responses for AI chat interfaces
• Implement response caching for repeated queries (semantic similarity > 0.95)
• Parallelize independent AI calls (search + glossary lookup)
• Use appropriate top_k for RAG scenarios (5–10 for most use cases)

Responsible AI

• All user-facing AI responses must pass through Azure AI Content Safety
• RAG responses must cite sources — never generate unsourced claims
• Implement groundedness checks (score ≥ 0.95 on 0–1 scale)
• Always include “AI-generated” disclaimers on outputs
• Critical decisions must have human-in-the-loop validation

{
  "content_safety": {
    "hate": 0,
    "violence": 0,
    "self_harm": 0,
    "sexual": 0
  }
}

WAF in the FAI Protocol

Every primitive can declare WAF alignment in its frontmatter:

---
description: "Reviews code for OWASP LLM Top 10 vulnerabilities"
waf: ["security", "responsible-ai"]
plays: ["30-ai-security-hardening"]
---

The fai-manifest.json enforces play-level WAF pillars. The FAI Engine validates that a play’s declared pillars are covered by its primitives:

{
  "context": {
    "waf": ["security", "reliability", "cost-optimization", "responsible-ai"]
  }
}

Valid WAF Pillar Values

These are the only valid values in waf arrays:

Next Steps

• FAI Protocol — how WAF is enforced at the protocol level
• Primitives — how each primitive type declares WAF alignment
• PR Checklist — WAF validation in pull requests