Compile a DSPrrr Program — compile

Main user-facing function to compile/optimize a DSPrrr module using a teleprompter optimization strategy.

Usage

compile_module(
  program,
  teleprompter,
  trainset,
  valset = NULL,
  .llm = NULL,
  ...,
  .trace_context = list()
)

Arguments

program: A DSPrrr module to optimize (e.g., from module())
teleprompter: A Teleprompter object defining the optimization strategy
trainset: Training data as a data frame
valset: Optional validation set for evaluation
.llm: Optional ellmer chat object to reuse during compilation
...: Additional arguments passed to the teleprompter
.trace_context: A named, JSON-compatible list propagated to evaluations, optimizer trials, and execution traces.

Value

An optimized module with updated demonstrations and/or instructions

Examples

if (FALSE) { # \dontrun{
# Create a simple module
classifier <- signature("text -> sentiment") |>
  module(type = "predict")

# Prepare training data
trainset <- data.frame(
  text = c("I love it!", "Terrible experience"),
  sentiment = c("positive", "negative")
)

# Compile with LabeledFewShot
tp <- LabeledFewShot(k = 2)
optimized <- compile_module(classifier, tp, trainset)

# Compile with GridSearch
variants <- data.frame(
  id = c("terse", "detailed"),
  instructions_suffix = c(
    "Be concise.",
    "Provide detailed reasoning."
  )
)
tp <- GridSearchTeleprompter(
  variants = variants,
  metric = metric_exact_match(field = "sentiment")
)
optimized <- compile_module(classifier, tp, trainset)
} # }