Package index • dsprrr

Signatures

Declare a task’s inputs, typed outputs and instructions.

signature(): Define a module's inputs and outputs
input(): Describe one signature input
with_instructions() append_instructions(): Replace or extend signature instructions
with_reasoning(): Add a reasoning field to a signature
without_reasoning(): Remove the reasoning field from a signature
has_reasoning(): Test whether a signature has a reasoning field

Build a program

Turn a signature into a module. module() makes one model call; the other constructors answer the same signature in a different way.

module(): Create a prediction module
chain_of_thought(): Create a chain-of-thought module
multi_chain_comparison(): Compare several reasoning chains and synthesize an answer
react(): Create a tool-using ReAct module
rag_module(): Create a retrieval-augmented generation module
module_fn(): Wrap an R function as a module
flex() experimental: Flex: a module whose implementation can be optimized

Compose programs

Chain modules, retry them with a reward, or combine several answers.

`%>>%` pipeline(): Chain modules into a pipeline
step(): Configure one pipeline step
best_of_n(): Run a module up to N times and keep the best result
refine(): Retry a module with feedback until it scores well
as_reward_fn(): Turn a metric into a reward function
ensemble(): Combine several modules into one
reduce_majority(): Pick the most common ensemble output
reduce_weighted_vote(): Pick the ensemble output with the most total weight
reduce_first(): Pick the first successful ensemble output
reduce_best_by_metric(): Pick the ensemble output that scores best against a known answer

Programs that run code

Let a model write and run R code, and choose where that code runs.

program_of_thought(): Create a Program of Thought module that answers by running R code
code_act(): Create a CodeAct agent that calls tools and runs R code
rlm_module() experimental: Create an RLM module that explores its inputs with R code
rlm(): Run a recursive language model (RLM) in one call
r_code_runner(): Run R code in a separate process
mcp_repl_runner(): Run R code in an operating-system sandbox with mcp-repl
run_demo(): Run the interactive RLM demo

Run programs

Run a module on named inputs or on a data frame, in parallel, asynchronously or as a stream, and read what came back.

run(): Run a module on named inputs
run_dataset(): Run a module on each row of a data frame
predict(<Module>): Predict with a module on new data
concurrency_control(): Control how batches run in parallel
run_async(): Run a module asynchronously
run_stream(): Run a module with streaming callbacks
stream_listener(): Listen to one output field while streaming
stream_async(): Stream a module's text output asynchronously
get_output() get_metadata() get_tokens() get_cost(): Extract outputs, metadata and costs from results

Validate outputs

Check outputs and retry with feedback when a check fails.

with_assertions(): Validate a module's outputs and retry on failure
assert_output() suggest_output() assertion_set(): Define output assertions
assert_length(): Assert the length of an output
assert_contains(): Assert that an output contains a string
assert_not_contains(): Assert that an output does not contain a string
assert_matches(): Assert that an output matches a regular expression
assert_not_matches(): Assert that an output does not match a regular expression
assert_one_of(): Assert that an output is one of a set of values
assert_range(): Assert that a numeric output is within a range
assert_not_empty(): Assert that an output is not empty
assert_custom(): Assert a custom condition

Evaluate

Score a program on labeled data with a metric.

evaluate(): Score a module on a dataset
metric_exact_match(): Exact-match metric
metric_f1(): Token-overlap F1 metric
metric_contains(): Metric that checks for a pattern in the output
metric_field_match(): Metric that compares several output fields
metric_custom(): Wrap a custom metric function
metric_threshold(): Turn a numeric metric into pass/fail
metric_with_feedback(): Metric that returns a score and feedback
metric_with_trace(): Metric that also sees the execution trace
metric_model_graded_qa() metric_model_graded_fact() metric_detect_match() metric_detect_includes() metric_detect_pattern(): Metrics built on vitals scorers

Optimize with compile()

Optimizers, which dsprrr calls teleprompters, tune a program’s demonstrations, instructions or decision rules against a metric. The “Choose an optimizer” article compares them.

compile(): Optimize a program with a teleprompter
LabeledFewShot(): Labeled few-shot: add training rows as demonstrations
BootstrapFewShot(): BootstrapFewShot: keep the program's own successful outputs as demos
BootstrapFewShotWithRandomSearch(): BootstrapFewShot with random search over candidate programs
KNNFewShot(): KNN few-shot: choose demonstrations by similarity at run time
COPRO(): COPRO: refine instructions by coordinate ascent
MIPROv2(): MIPROv2: search instructions and demonstrations together
GEPA(): GEPA: reflective prompt evolution
SIMBA(): SIMBA: stochastic introspective mini-batch ascent
GridSearchTeleprompter(): Grid search over instructions and demos
BetterTogether(): BetterTogether: run several optimizers in sequence
ReAnchor() experimental: ReAnchor: calibrate decision outputs against a metric
Teleprompter(): Arguments shared by all optimizers

Experimental optimizers

Compose optimizers, or let an agent run optimization experiments.

Omni(): Omni: explore with several optimizers, then continue from the best
AutoResearch(): AutoResearch: let an agent run optimization experiments
MetaHarness(): Meta-Harness: fresh proposer sessions over a candidate frontier

Inspect optimization results

See what an optimizer changed and reuse the result.

optimization_result() print(<dsprrr_optimization_result>): Inspect what an optimizer did
optimization_summary() print(<dsprrr_optimization_summary>): Summarize an optimization result
best_params(): Best parameters found by an optimizer
best_demos(): Demonstrations attached to a module
top_trials(): Highest-scoring optimization trials
config_diff(): Compare a module's settings with baseline values
apply_best_config(): Copy optimized settings to another module
export_module_code(): Export a program as standalone R code

Grid search

Try combinations of request settings, instructions and templates.

optimize_grid(): Grid search over module settings
module_trials(): Summarize grid search trials
module_metrics(): Per-trial rows from a grid search
module_parameters(): Build a tidymodels parameter set for a module

Calibrated decisions

Experimental. Yes/no and label outputs decoded locally from probability evidence, with thresholds you can tune.

decision_bool() decision_score() decision_choice() experimental: Calibrated decision outputs
with_decisions() experimental: Attach calibrated decision outputs to a module
decision_settings() experimental: Decision settings of a module
decision_evidence() experimental: Probability evidence behind decision outputs

Models and caching

Choose the chat a call uses and control the response cache.

dsp_configure(): Configure the default chat from a provider name
get_default_chat() set_default_chat() clear_default_chat(): Get, set or clear the default chat
with_lm() local_lm(): Use a chat for a block of code
is_reasoning_model(): Test whether a model name is a reasoning model
configure_cache(): Configure the response cache
cache_stats(): Report response-cache statistics
clear_cache(): Clear the response cache
dsprrr_sitrep(): Report dsprrr's configuration

Prompts, traces and costs

Inspect what was sent to the model, what came back and what it cost.

get_last_prompt(): Show the most recent prompt and response
inspect_history(): List recent model calls
clear_prompt_history(): Clear the prompt history
export_traces(): Export a module's traces as a tibble
summarize_traces(): Summarize a module's traces
clear_traces(): Clear a module's traces
session_cost(): Summarize this session's token use and cost

Save and deploy

Save programs, pin them to a board, and start a workflow project.

program_artifact() program_artifact_id() save_program() load_program(): Save and load complete programs
pin_module_config(): Pin a program to a pins board
restore_module_config(): Rebuild a program from a saved artifact
pin_trace(): Pin a module's traces to a pins board
pin_vitals_log(): Pin evaluation results to a pins board
use_dsprrr_template(): Copy a workflow template into a project
validate_workflow(): Check a module and its data before a run

Program graphs

Inspect, freeze and transform the modules inside a program.

module_graph() named_modules() named_parameters() map_modules() replace_module() freeze_modules() is_module_frozen() set_module_lm() module_children(): Inspect and transform nested programs

vitals

Evaluate modules with the vitals package.

as_vitals_solver(): Use a dsprrr module as a vitals solver
as_vitals_task(): Build a vitals Task from a module and a data set
as_vitals_samples(): Convert dsprrr traces to vitals samples
as_vitals_cost(): Report dsprrr costs in the vitals format
as_dsprrr_traces(): Convert vitals samples to dsprrr traces
summarize_traces_df(): Summarize a traces data frame
as_dsprrr_metric(): Use a vitals scorer as a dsprrr metric

Tools and retrieval

Expose modules as ellmer tools and search document stores.

as_ellmer_tool(): Turn a module into an ellmer tool
ragnar_tool(): Search a ragnar store from a tool-using module
create_search_tool(): Build a search tool from a set of documents

tidymodels

A parsnip model specification and dials parameters.

llm_predict(): LLM model specification for parsnip
register_dsprrr_engine(): Register the dsprrr engine with parsnip
temperature(): Temperature parameter for dials
top_p(): Top-p parameter for dials
reasoning_effort(): Reasoning effort parameter for dials

Build your own optimizer

Lower-level pieces the built-in optimizers are made from.

optimizer_control(): Budgets and settings for an optimizer run
eval_program(): Evaluate a program with per-example detail
sample_dataset(): Sample rows reproducibly
split_dataset(): Split data into training and validation sets
TrialLog: Record optimization trials in memory or on disk
create_trial(): Create an optimization trial record
complete_trial(): Record evaluation results on a trial
write_trials_jsonl(): Write trial records to a JSON Lines file
read_trials_jsonl(): Read trial records from a JSON Lines file
load_trial_log(): Load a saved trial log