Search

Search uses immutable Experiment parameters and evidence instead of importing provider code. SearchSpace validates every discrete value through typed with_params() before execution. GridSearch enumerates every compatible combination; RandomSearch samples without replacement from the same finite space under an explicit seed. Constraints filter named public parameter paths. Invalid combined substitutions fail before an evaluator runs.

evaluate_search requires each BenchmarkResult to match its trial ID, configuration digest, profile, model, and component set. ParetoFrontier accepts explicit min/max directions and compares only records with identical scope, model, workload, environment, privacy cohort, numeric cohort, warmups, repetitions, and objective semantics. Failed or unavailable records are retained but excluded from the frontier. There is no implicit scalar score or universal “best” candidate.

Python SDK example

from pllm import Constraint, GridSearch, SearchSpace
from examples.composition import experiment

space = SearchSpace(
    experiment,
    {"pipeline__kernels__threads": [1, 2, 4]},
    constraints=[Constraint("pipeline__kernels__threads", "ge", 2)],
)
candidates = GridSearch("cpu-grid", space).candidates()
assert [candidate.experiment.pipeline.kernels.params["threads"] for candidate in candidates] == [2, 4]

Finding a candidate is not proof that it works. A selected experiment must run in its stated environment and produce matching benchmark and assurance records. Bayesian search and the CLI benchmark search/compare commands remain unavailable.

API: pllm.search