Modules
Lazy evaluation and search pipelines for Elixir values and LLM programs.
A pure prompt format.
DSPy-style Chat format with [[ ## field ## ]] markers.
DSPy-style JSON format: marked inputs and a JSON object reply.
DSPy-style XML format with schema-directed field delimiters.
A versioned learned-value artifact with safe provenance.
A proposed value and its completed evaluation.
One normalized input and expected result.
The paired evidence for an original target and a search winner.
One demonstration: complete inputs and outputs for a signature.
A program that runs compatible member programs and reduces their outputs.
Structured errors owned by Tune.
Error class for runtime failures that retain partial results.
Application code did not follow a Tune callback contract.
A program run failed after it started.
A managed limit stopped a call before dispatch.
Error class for invalid definitions, options, and inputs.
A dataset source failed while Tune was reading its rows.
A signature, field, predictor, demonstration, or model configuration is invalid.
Program inputs do not match the declared input fields.
An option is unknown, forbidden, or has an invalid value.
An optimization could not return an eligible program.
Error class for model replies that cannot become a validated output.
Parsed output values do not match the declared output schema.
An adapter could not parse the reply into declared output fields.
A ReqLLM response is not an acceptable completed reply.
Error class for unexpected failures.
An unexpected error.
The eager result of evaluating one target over one dataset.
Per-run state threaded through application program code.
One named, typed signature field.
A program that selects nearby training examples for each input.
One explicit, fixed model configuration: a ReqLLM model, a prompt adapter, adapter options, and ReqLLM request options.
The scores and outcome for one evaluated case.
The Optimize.Anything target: GEPA-style search over one declared value.
Contrastive instruction improvement for one named predictor.
Builds demonstrations from successful teacher trajectories.
Searches example-set choices per predictor with categorical TPE.
Builds several seeded example sets and keeps the best validation result.
Coordinate instruction search over predictors.
Builds one compatible ensemble from the pipeline target and explicit members.
Reflective prompt evolution with Pareto selection and merge.
Derives natural-language rules from bootstrapped examples.
Returns a Tune.KNN program whose pool is the training set.
Uses up to k complete labeled training examples as demonstrations for each
predictor. Makes no model calls.
Joint search over instructions and example sets.
Learns a bounded list of lessons stored in a declared JSON parameter.
Typed internal predictors used by instruction, rule, and value optimizers.
The reflective evolutionary search shared by GEPA and Optimize.Anything.
Stochastic introspective mini-batch ascent.
Searches the instruction of one explicitly named predictor.
A categorical Tree-structured Parzen Estimator.
A tunable JSON value with a Zoi schema, such as a lesson list or candidate text.
The outcome of one target execution.
An immutable typed predictor: a signature and its demonstrations.
A rendered, inspectable request for one predictor call.
The contract for a typed Tune program.
The immutable result of a completed tuning run.
Runs a typed predictor or composed program on one input.
One named score declaration and its execution contract.
The aggregate for one named score.
A structured ReqLLM judge for retrieval-context precision.
A structured ReqLLM judge for grounded-response faithfulness.
Internal operations and program data for one search strategy.
An internal program-search algorithm.
Shared limits and accounting for Tune runs.
Ordered, typed input and output fields plus task instructions.
A value space used by search strategies.
A finite choice space.
A linear or logarithmic floating-point value space.
An integer value space.
A bounded subset of a finite list.
A bounded text space for simple evolutionary search.
One completed predictor call: its name, role, signature, actual inputs, validated outputs, and the unchanged ReqLLM response.
Starts the optional local Tune workbench.
The application-code boundary used by the optional Tune UI.
Scores a saved single-turn dataset with both retrieval judges.
A UI-authored single-predictor Tune recipe.
Resolves workbench recipe IDs to trusted application modules.
The optional local workbench router.
An in-memory workbench run and its current outcome.
A data-only description of one workbench study.
Mix Tasks
Starts the optional local Tune workbench.