Engine configuration
Sematryx exposes three tunable capability families—agentic (strategy and solver coordination), interpretable (explanations and audit artifacts), and adaptive (learning and transfer when enabled). They are API-controlled knobs, not a separate product line.
For the public story, start with Product and Performance. This page is the reference surface for how those capabilities map to configuration.
Capability overview
Agentic
Meta-policy and portfolio routing select and hand off between solvers from problem signals—landscape shape, constraints, and budgets—without requiring you to name an algorithm upfront.
- • Automatic strategy selection for heterogeneous problems
- • Fewer manual tuning cycles on non-standard formulations
- • Aligns with benchmarked multi-landscape performance
Interpretable
Configurable explanation depth: summaries, technical logs, and sensitivity-style views. Heavy paths can run asynchronously so latency-sensitive solves stay lean.
- • Audit-friendly artifacts for regulated workflows
- • Debug infeasible or fragile models with structured rationale
- • Adjustable cost vs transparency
Adaptive
Optional learning and transfer across similar problems—private to your organization when using dedicated stores—reducing evaluation counts on repeat workloads.
- • Warm-start effects on related solves
- • Tenant-scoped memory (no cross-tenant blending)
- • Can be disabled for strict reproducibility
How they compose
In a typical high-capability run: routing selects a strategy, the solve executes, explanations are emitted at the requested level, and adaptive stores update only where policy allows. Presets bundle these for common production vs research trade-offs.
Presets (cost vs capability)
Minimal latency
Fast path: narrow explanations, adaptive off where possible.
Balanced (typical production)
Agentic + interpretable at moderate depth; adaptive per policy.
Maximum insight
Full explanation depth, adaptive on, higher compute budget.