GEPA.Proposer behaviour (GEPA v0.3.0)

Copy Markdown View Source

Behavior for candidate proposal strategies.

Proposers generate new candidate programs based on the current optimizer state. They implement the mutation logic that drives the evolutionary search.

Proposal Types

  • Reflective Mutation: Uses LLM-based reflection on execution traces
  • Merge: Combines successful programs through genealogy analysis

Example Implementation

defmodule MyProposer do
  @behaviour GEPA.Proposer

  @impl true
  def propose(state) do
    # Analyze state, generate new candidate
    # Return {:ok, proposal}, :none, or {:error, reason}
  end
end

Summary

Callbacks

propose(t)

Propose a new candidate program based on current state.

Callbacks

propose(t)

@callback propose(GEPA.State.t()) ::
  {:ok, GEPA.CandidateProposal.t()} | :none | {:error, term()}

Propose a new candidate program based on current state.

Parameters

  • state: Current GEPA optimization state

Returns

  • {:ok, proposal}: Successfully generated proposal
  • :none: No proposal could be generated (e.g., perfect scores already)
  • {:error, reason}: Failure to generate proposal

Contract

  • Proposers may evaluate candidates on subsamples
  • Proposers should update state.total_num_evals if they evaluate
  • Proposers should add trace information to state.full_program_trace
  • The engine handles full validation evaluation and acceptance