Intelligent Model Routing (Complexity Routing)

AgentWatch can route model: "auto" requests to the most appropriate upstream per request, based on an in-process, CPU-only classification of each prompt (task type · domain · complexity 1–5 · language) via heuristics → a fine-tuned ONNX classifier → an optional embedding-RAG tier. No prompt content leaves the box to make the decision. Off by default; opt-in per tenant.

Capabilities: task-type and complexity-tier routing (coding:5, coding:*); multiple upstreams per tier with failover/round-robin/weighted/random strategies; a context-length override (large inputs → big-context model regardless of complexity); per-tenant configuration layered over a platform default; full observability (route header, decisions table, analytics panel) and a dry-run classify endpoint. Configured via GET/PUT /admin/router/config and the super-admin Complexity Routing tab. No SQL schema change (state in system_settings). Authoritative reference: AgentWatch_Complexity_Routing.md.