Routing & Fallback

LW AI uses intelligent routing and automatic fallback mechanisms to ensure high availability and optimal performance for every request.

Request Routing

Region-Based Routing

All requests use the single endpoint openapi.linkwo.ai. GeoDNS routes each request to the nearest region automatically:

  • American users → Hillsboro infrastructure (lowest latency)
  • Other international users → Singapore infrastructure

Model Routing (Astra V1 Pro)

When using the fusion model, requests are routed to the best underlying model based on:

FactorDescription
Task complexitySimple → Flash, Complex → Pro
DomainFinance/energy → Yantronic V1 Pro
LanguageChinese → GLM, English/Code → DeepSeek
Current loadBalances across models for throughput

Automatic Fallback

If the primary model or infrastructure is unavailable, LW AI automatically falls back to maintain service:

Model Fallback

Primary Model → Fallback Model (same capability tier)

For example, if deepseek/deepseek-v4-pro is overloaded:

  1. Request routed to deepseek/deepseek-v4-pro
  2. If unavailable, falls back to z-ai/glm-5.1 (same tier)
  3. Response returned with model_routed header indicating the actual model used

Infrastructure Fallback

Primary Region → Secondary Region

If the Hillsboro region experiences issues:

  1. Requests to openapi.linkwo.ai are proxied to Singapore
  2. Latency may increase slightly
  3. Service remains available

Fallback Behavior

ScenarioBehaviorUser Impact
Model overloadedFallback to equivalent modelMinimal — same quality tier
Model offlineFallback to equivalent modelMinimal — response includes actual model used
Region degradedCross-region proxySlight latency increase
Complete outageMaintenance pageAPI returns 503 with retry-after header

Disabling Fallback

For applications that require strict model control, you can disable fallback:

response = client.chat.completions.create(
    model="deepseek/deepseek-v4-pro",
    messages=[...],
    extra_body={"allow_fallback": False}
)

When fallback is disabled, unavailable models will return a 503 Service Unavailable error instead of routing to an alternative.

Monitoring

  • Check model_routed in response headers to see which model actually handled your request
  • Use Service Status for real-time infrastructure health
Previous
Rate Limits
Next
Performance