Optimize a program with a teleprompter

compile() improves a program (its demos, instructions or other settings) with a teleprompter such as LabeledFewShot(), BootstrapFewShot() or MIPROv2(), using a training set. It returns a new program and leaves the input unchanged. Its argument order suits the native pipe: program |> compile(teleprompter, trainset).

Usage

compile(program, teleprompter, ...)

Arguments

program

The module or pipeline to optimize. Programs made with module_fn() cannot be compiled.

teleprompter

A teleprompter object that sets the optimization strategy. Integer settings of teleprompters need integer literals, as in LabeledFewShot(k = 3L); k = 3 is an error.

...

The training set and optimizer options:

• trainset (required, third argument): a data frame with the program's input columns and the expected outputs.

• valset: an optional validation data frame, for teleprompters that use one. It may also be given as the fourth argument.

• .llm: an ellmer Chat for the program's calls (see run()).

• .trace_context: a named, JSON-compatible list copied into the metadata and traces of the calls made while compiling.

Some teleprompters take further arguments; see their help pages.

Value

A new, compiled program of the same kind as program. Check it with compiled$is_compiled().

Details

Teleprompters that score candidates call the metric as metric(prediction, expected), where expected is the whole training row, so give built-in metrics a field, as in metric_exact_match(field = "sentiment"). Compiling a program that is already compiled works but gives a warning.

See also

Other teleprompters: AutoResearch(), BetterTogether(), BootstrapFewShot(), BootstrapFewShotWithRandomSearch(), COPRO(), GEPA(), GridSearchTeleprompter(), KNNFewShot(), LabeledFewShot(), MIPROv2(), MetaHarness(), Omni(), ReAnchor(), SIMBA(), Teleprompter()

Examples

classifier <- module(
  signature("text -> sentiment: enum('positive', 'negative')")
)
trainset <- data.frame(
  text = c("I love it!", "Terrible experience", "Works great", "Broke in a day"),
  sentiment = c("positive", "negative", "positive", "negative")
)

# LabeledFewShot copies training rows into the prompt as demos, so it
# needs no model calls
compiled <- classifier |> compile(LabeledFewShot(k = 2L), trainset)
compiled$is_compiled()
#> [1] TRUE
classifier$is_compiled()
#> [1] FALSE
compiled$demo_table
#> # A tibble: 2 × 2
#>   text           output  
#>   <chr>          <chr>   
#> 1 Works great    positive
#> 2 Broke in a day negative

if (FALSE) { # \dontrun{
# BootstrapFewShot runs the program and keeps demos that pass the metric
bootstrapped <- classifier |>
  compile(
    BootstrapFewShot(
      metric = metric_exact_match(field = "sentiment"),
      max_bootstrapped_demos = 2L
    ),
    trainset,
    .llm = ellmer::chat_openai(model = "gpt-6-luna")
  )
} # }