Arguments shared by all optimizers

Teleprompter() is the parent class of every optimizer in dsprrr ("teleprompter" is DSPy's original name for an optimizer). You never compile with it directly: create one of the optimizers below and pass it to compile(). This page documents the three arguments that every optimizer accepts.

Usage

Teleprompter(metric = NULL, metric_threshold = NULL, max_errors = 5L)

Arguments

metric: A metric function called as metric(prediction, expected), such as metric_exact_match(field = "answer"), or NULL. Optimizers that score candidates require one.
metric_threshold: A score between 0 and 1, or NULL (the default). Only some optimizers use it, and each optimizer's page says how.
max_errors: Integer error budget (default 5L). The optimizers that run under optimizer_control() (the two bootstrap optimizers, COPRO, MIPROv2, SIMBA, GEPA, AutoResearch and MetaHarness) stop after this many consecutive failed evaluations when compile() is not given a control object.

Value

A Teleprompter object.

Details

Most optimizers store counts as S7 integer properties, so write 5L rather than 5. A double fails with an error such as @max_errors must be <integer>, not <double>.

Optimizers at a glance

Optimizer	What it changes
`LabeledFewShot()`	Demos: `k` training rows. No model calls.
`KNNFewShot()`	Demos: the training rows most similar to each input, chosen at run time.
`BootstrapFewShot()`	Demos: training rows plus the program's own outputs that pass the metric.
`BootstrapFewShotWithRandomSearch()`	Demos: the best of several bootstrap candidates on a validation set.
`GridSearchTeleprompter()`	Instructions, template or settings, from a table of variants.
`COPRO()`	Instructions, rewritten by a model over several rounds.
`MIPROv2()`	Instructions and demos, searched together.
`SIMBA()`	Instruction rules and demos taken from hard examples.
`GEPA()`	Instructions (and Flex source) evolved by reflecting on failures.
`ReAnchor()`	Thresholds, cuts and weights of decision outputs.
`BetterTogether()`, `Omni()`	Run other optimizers in sequence or in competition.
`AutoResearch()`, `MetaHarness()`	An agent proposes and tests program edits.

To sweep runtime settings such as reasoning_effort on a module in place, use optimize_grid() instead.

Examples

Teleprompter()
#> <dsprrr::Teleprompter>
#>  @ metric          : NULL
#>  @ metric_threshold: NULL
#>  @ max_errors      : int 5

# Counts must be integers
try(Teleprompter(max_errors = 5))
#> Error : <dsprrr::Teleprompter> object properties are invalid:
#> - @max_errors must be <integer>, not <double>