Omni Teleprompter

A meta-teleprompter that explores several optimization strategies from the same seed program, compares their outputs with one shared validation metric, and seeds a fresh continuation optimizer from the winner.

Inspired by the Omni meta-optimizer from the GEPA project, this adapts the explore, pick-best, and continue pattern described in the GEPA Omni announcement to dsprrr modules. The original program remains a candidate throughout, so a regressing explorer or continuation step cannot replace a better program.

Omni() does not impose a common budget because dsprrr teleprompters expose different native budget controls. Configure comparable budgets on the explorer objects before constructing Omni(). Common validation re-scoring of the seed, each explorer result, and the continuation result is additional evaluation work outside those native optimizer budgets.

Usage

Omni(
  metric,
  explorers,
  continuation,
  metric_threshold = NULL,
  max_errors = 5L,
  valset_ratio = 0.1,
  parallel = FALSE,
  num_workers = NULL,
  seed = NULL,
  verbose = TRUE
)

# S3 method for class 'Omni'
print(x, ...)

Arguments

metric

Metric function used to compare every candidate on the same validation set.

explorers

Named list of at least two Teleprompter objects. Every explorer starts from an independent copy of the input program.

continuation

A Teleprompter object run from the best exploration candidate.

metric_threshold

Minimum score required to be considered successful.

max_errors

Maximum number of errors allowed during evaluation.

valset_ratio

Fraction of trainset to hold out for candidate comparison when valset is not supplied.

parallel

Whether to compile exploration branches concurrently with mirai. Parallel exploration requires .llm = NULL. Each worker creates its own default chat from OPENAI_API_KEY, ANTHROPIC_API_KEY, or GOOGLE_API_KEY.

num_workers

Number of mirai workers for parallel exploration. NULL uses one worker per explorer.

seed

Optional whole-number random seed within R's integer range for reproducible splitting, sequential exploration, and mirai worker streams.

verbose

Whether to print progress messages.

x

An Omni object.

...

Additional arguments.

Examples

if (FALSE) { # \dontrun{
metric <- metric_exact_match(field = "answer")

tp <- Omni(
  metric = metric,
  explorers = list(
    bootstrap = BootstrapFewShotWithRandomSearch(metric = metric),
    gepa = GEPA(metric = metric, population_size = 4L, generations = 2L)
  ),
  continuation = GEPA(
    metric = metric,
    population_size = 4L,
    generations = 2L
  )
)

compiled <- compile(tp, qa_module, trainset, valset = valset, .llm = llm)
compiled$config$optimizer$candidate_programs
} # }