Feature Routing

Feature routing is the configuration layer that maps each specific AI use case — "ticket-suggest," "call-recap," "behavioral-analysis" — to a specific provider and model. Every AI call in the platform specifies a feature string, and the Gateway resolves it to the right model automatically.

Why Feature Routing Exists

Different AI tasks have different requirements:

• Ticket summarization needs speed and low cost, not maximum intelligence — GPT-4.1-mini is a good fit
• Attack narrative generation needs deep reasoning and long context — GPT-4.1 is better
• Call transcription requires a speech-to-text model, not a chat model
• Proposal generation benefits from creativity, so a higher temperature is appropriate

Rather than hard-coding provider and model choices into every product, all products call the Gateway with a feature string and let the routing config handle the rest. This means you can change which model handles ticket suggestions across all products with a single configuration change — no deployments required.

How Feature Routing Works

When a product calls the Gateway with feature: 'ticket-suggest':

1. Gateway looks up the routing policy for ticket-suggest from Redis (cached for 5 minutes)
2. Routing policy specifies: provider, model, temperature, max_tokens, PII mode, fallback provider/model
3. If the primary provider's circuit breaker is open, the fallback is used
4. The request is sent to the resolved provider and model

Built-In Feature Routes

These are the platform's built-in feature-to-model mappings. They apply across all products that use the AI Gateway:

Feature	Provider	Model	Notes
ticket-suggest	Azure OpenAI	GPT-4.1-mini	PSA ticket field suggestions. Fallback: Anthropic / Claude Opus
ticket-triage	Azure OpenAI	GPT-4.1-mini	Auto-triage and priority scoring
call-recap	Azure OpenAI	GPT-4o	Voice call summary generation. No fallback configured.
call-transcription	Azure AI Foundry	Whisper	Speech-to-text. Separate from chat routing.
behavioral-analysis	Azure OpenAI	GPT-4.1	Defend anomaly scoring — requires large context window
attack-narrative	Azure OpenAI	GPT-4.1	Defend investigation summaries
code-review	Anthropic	Claude Sonnet	Code Platform review suggestions
proposal-generate	Azure OpenAI	GPT-4.1	Long-form content generation
kb-generate	Azure OpenAI	GPT-4.1-mini	KB article generation from topic prompts
data-enrichment	Azure OpenAI	GPT-4.1-mini	Company and contact record enrichment
governance-scan	Azure OpenAI	GPT-4.1	AI policy compliance scanning
safety-moderate	Azure OpenAI	GPT-4o	Content moderation
assessment-eval	Azure OpenAI	GPT-4.1	AI assessment scoring and recommendations
workdna-profile	Azure OpenAI	GPT-4.1	Organizational behavioral analysis
chat-playground	Azure OpenAI	User-selected	Playground respects model dropdown selection
jarvis-assistant	Azure OpenAI	GPT-4.1	Jarvis uses the most capable model for cross-product context
embed	Azure OpenAI	text-embedding-3-large	Embedding generation for semantic search

Custom Feature Routing (Enterprise)

Enterprise tier tenants can define custom feature routes for their own Custom GPTs and workflows:

1. Navigate to Settings → Feature Routing
2. Click Add custom route
3. Fill in:
   • Feature name — A unique string identifier (e.g., my-custom-agent)
   • Provider — Azure OpenAI, Azure AI Foundry, or Anthropic
   • Model — The specific model ID
   • Temperature — 0–2 (default: 0.7)
   • Max tokens — 1–32,768
   • PII mode — scrub, warn, or allow
   • Fallback provider/model — (optional)
4. Click Save — the route is live within 5 minutes (cache TTL)

ℹ️Custom routes are only available on Enterprise tier. On lower tiers, Custom GPTs use the chat-playground route, which routes to Azure OpenAI GPT-4.1 with the user-selected model as a hint.

Overriding the Model at Runtime

When calling the AI inference endpoint directly (from the Playground or API), you can suggest a model in the request:

{
  "messages": [...],
  "model": "gpt-4.1-mini"
}

The Gateway will use the suggested model if:

1. The feature route for chat-playground allows model overrides (it does by default)
2. The requested model is in your tier's allowed model list
3. The provider serving that model is healthy

If conditions are not met, the Gateway falls back to the route's default model.

Viewing Route Usage

Navigate to Usage Analytics → By Feature to see which features consumed the most tokens last month, what models served them, and what they cost. This helps you decide which routes to optimize.

See also: Cost Tracking, Usage Analytics