KKissan Ki Pehchan
AI and Knowledge

Model and Provider Routing

Routes live speech, selected frames, research and exact controls to the least costly adequate approved capability.

BlueprintVersion 0.25 Aug 2026

Routing principle

The media plane is continuous; model input is selective. Audio is streamed to STT, while video is converted into quality-approved selected frames or explicit stills. Use the strongest model only where complexity, conflict or consequence justifies it.

Workload routing

WorkloadRoute
VAD, network adaptation, blur/exposureOn-device or realtime deterministic/compact model
Frame relevance and crop presenceLow-cost vision model or approved local model
Ordinary grounded caseBenchmarked balanced reasoning model
Difficult differential/source conflictGPT-5.6 Sol with higher reasoning effort
Web research synthesisStrong reasoning model with controlled search
Exact treatment calculationDeterministic policy service
High-risk final decisionStrongest approved model plus human/policy gate

Routing inputs

Risk, case complexity, number of hypotheses, frame quality, novelty, research need, latency budget, cost budget, provider health, data-residency class and officer availability.

Fallback

Provider fallback must preserve evidence IDs and schema contracts. If live vision is unavailable, declare the limitation and switch to guided stills rather than silently routing an unverified frame.