Sampling expresses provider-neutral generation intent. A codec maps these fields to its own wire vocabulary and applies dialect-specific constraints.

API surface

type Sampling struct {
	Temperature *float64
	TopP        *float64
	MaxTokens   *int
	Stop        []string
	Effort      Effort
}

func (s Sampling) Clone() Sampling

Field	Zero value	Ownership
`Temperature`	nil, unset	Pointer is copied by `Clone`
`TopP`	nil, unset	Pointer is copied by `Clone`
`MaxTokens`	nil, unset	Pointer is copied by `Clone`
`Stop`	nil, unset	Slice is copied by `Clone`
`Effort`	`EffortNone`	OpenAI and Anthropic codecs map it differently

The model package does not validate numeric ranges or provider-specific combinations. For example, an Anthropic thinking request may have a temperature rule that belongs in the Anthropic codec.

temperature := 0.2
maxTokens := 512
defaults := model.Sampling{
	Temperature: &temperature,
	MaxTokens: &maxTokens,
	Stop: []string{"END"},
	Effort: model.EffortMedium,
}
copy := defaults.Clone()
*copy.Temperature = 0.7
copy.Stop[0] = "DONE"
// defaults still contains 0.2 and "END".

Use WithSampling(defaults) for model defaults and Request.Override for a per-call replacement.

Proof

Source: inference/model/sampling.go
Tests: inference/model/sampling_test.go, inference/model/model_test.go