smart-router

Multi-model query routing evaluates per-query features and past outcomes to choose between full deliberation and a shortcut. It uses configuration settings and pattern memory to match queries against historical performance, recommending the appropriate reasoning depth.

Confidence scoring in multi-model query routing evaluates routing certainty by analyzing query features and historical pattern memory. It produces a decision with confidence levels, uncertainty flags, and estimated cost to indicate routing reliability.

To configure multi-model query routing for always using full deliberation, set the routing configuration to always_deliberate. This forces the system to bypass heuristic shortcut routing and send every query to multi-model Council deliberation regardless of query features.

Yes, multi-model query routing works for architecture decisions and compliance security questions. It applies routed question answering across categories including architecture, compliance, code implementation, strategy, creative writing, and factual lookup, detecting compound queries.

The best way to optimize model lineup and cost control is using pattern memory to recommend and exclude models based on stored performance ranks and bias warnings. This prevents systematically underperforming reviewers and reduces unnecessary deliberation costs.

When pattern memory is insufficient for learned routing, the system falls back to heuristic routing. This ensures query routing continues to function before enough logged deliberations, requiring 20 or more, have accumulated to switch to learned routing.