Integrate / Inference modes
Choose an inference mode
Start with the deterministic mode when you need a local, inspectable baseline without a model provider.
Deterministic#
The resolver uses lexical evidence and available released anchors. Small margins abstain. Scores are ranking signals, not calibrated probabilities. This mode is useful for testing product integration before a provider is configured.
Jev#
Jev selects among a bounded sense inventory through TypeSafe. It requires a server-side credential and entitled access. The adapter validates returned options and distributions rather than accepting an arbitrary generated definition.
Laya#
Laya is a separately configured local inference path. It needs an actual model runtime, checkpoint identity, and context-budget checks. A provider name alone is not an installed or validated model.
Evaluate on your own task#
Use the same passages, accepted options, and abstention criteria when comparing providers. Keep model agreement separate from human validation. A benchmark from another domain does not establish quality for your users’ language.
Download the provider engineering reference for token-budget and adapter details.