gepa¶

gepa(
    fn,
    data,
    *,
    teacher=None,
    expected=None,
    selection=None,
    budget=300,
    minibatch=4,
    metric=None,
    feedback=None,
    num_threads=8,
    seed=0,
)

An improved copy whose instruction a teacher model rewrote from the function's mistakes (GEPA; design/04-gepa.md).

Half the rows (or selection) choose and are never shown to the teacher; the other half give feedback. better.trials is the search. Its scores on the choosing rows flatter the one chosen: measure it on rows it never saw. The optimizer class GEPA takes the same options, for fn.opt(rows, optimizer=GEPA(...)) and for @modules' AI functions one at a time.

Examples¶

import functai
from functai import *

# not run: a search asks the teacher many times (Make it better runs one)
better = functai.gepa(team.using(lm="gpt-5.4-nano"), train, teacher="gpt-6-sol", expected="category")
better.instructions
functai.evaluate(better, test, expected="category")