Skip to contents

Attach calibrated decision outputs to a module

Source: R/decision.R
with_decisions.Rd

[Experimental]

Returns a copy of a Predict module in which the named output fields are decoded from probability evidence. Each field must be declared in the module's object-shaped signature output: a type_boolean() field for decision_bool(), or a type_enum() field for decision_score() and decision_choice().

A decision field needs a question: either the output type's description (for example type_boolean("Is the service blocked?")) or the spec's description. Without one, with_decisions() fails rather than asking an unspecified question.

Pass NULL for a field to remove its decision configuration and return that field to direct generation.

Decision modules run on the sequential execution path. Concurrent batch backends are rejected rather than returning undecoded evidence.

Usage

with_decisions(module, ...)

Arguments

module

A module created by module() or chain_of_thought().

...

Named decision_types specifications, or NULL, keyed by output field name.

Value

A modified copy of module. The original is unchanged.

See also

Other decisions: ReAnchor(), decision_evidence(), decision_settings(), decision_types

Examples

sig <- signature(
  inputs = list(input("ticket", description = "Customer report")),
  output_type = ellmer::type_object(
    urgent = ellmer::type_boolean("Is the service blocked?"),
    severity = ellmer::type_enum(
      c("minor", "disruptive", "blocking"),
      "How severe is the impact?"
    ),
    category = ellmer::type_enum(
      c("billing", "technical"),
      "Which team owns the issue?"
    )
  )
)

triage <- module(sig) |>
  with_decisions(
    urgent = decision_bool(threshold = 0.7),
    severity = decision_score(),
    category = decision_choice(
      criteria = c(billing = "Payments", technical = "Product faults")
    )
  )

decision_settings(triage)
#> # A tibble: 3 × 5
#>   field    kind   threshold cuts      weights  
#>   <chr>    <chr>      <dbl> <list>    <list>   
#> 1 urgent   bool         0.7 <NULL>    <NULL>   
#> 2 severity score       NA   <dbl [2]> <NULL>   
#> 3 category choice      NA   <NULL>    <dbl [2]>