gepa_optimizer
diapason.learning.agents.gepa_optimizer
¶
GEPA agent optimizer -- Pareto-efficient evolutionary optimization.
Uses GEPA's adapter pattern to bridge Diapason traces into GEPA's evolutionary optimization framework. Outputs TOML config updates written via AgentConfigEvolver.
DiapasonGEPAAdapter
¶
Implements GEPA's adapter protocol for Diapason agents.
Bridges trace data into GEPA's optimization framework via
the assess() and make_reflective_dataset() methods.
assess(batch, candidate, capture_traces=False)
¶
Score a candidate config against a batch of test cases.
Returns a dict with at least 'scores' (list of floats) key.
load_traces()
¶
Load traces from the store.
make_reflective_dataset(candidate, assessment_batch, components_to_update)
¶
Package trace diagnostics as Actionable Side Information for GEPA.
GEPAAgentOptimizer
¶
Optimize agent configs using GEPA evolutionary optimization.
Parameters¶
config: GEPAOptimizerConfig controlling optimization parameters.
optimize(trace_store)
¶
Run GEPA optimization on traces from the store.
- Load traces and build the GEPA adapter
- Define the search space (agent config fields)
- Run GEPA's evolutionary optimization
- Extract best candidate as TOML updates
- Write via AgentConfigEvolver if config_dir is set