GEPA¶

GEPA(
    metric=None,
    *,
    budget=300,
    minibatch=4,
    teacher=None,
    feedback=None,
    num_threads=8,
    seed=0,
)

Rewrite the instruction from the function's mistakes: GEPA (Agrawal et al., 2025), with functai's changes (design/04-gepa.md).

Half the rows (or valset) select, and are never shown; the other half give feedback. A pool of instructions is scored row by row on the selection rows; a parent is picked from its Pareto frontier, run on minibatch feedback rows, and a teacher model reads its answers with feedback in words and writes a new instruction. A child that does better on the minibatch is scored on the selection rows and joins the pool. Every fourth step, two frontier candidates that win different rows are combined. The reflection sees what was tried and failed; a proposal that copies an input is dropped; ties go to the shorter instruction; a row is run once per instruction.

budget counts calls of the function; trials holds every candidate afterwards (its selection score is optimistic: it was chosen on those rows). feedback(row, prediction, error) -> str gives the words (default: right, or the right answer, or the error).