GEPA.Proposer.Reflective (GEPA v0.3.0)

Reflective mutation proposer.

Generates new candidates through reflection on execution traces. Candidate text must be proposed by the adapter, a custom proposer, or an instruction_proposal backed by a reflection LLM.

With LLM-based Instruction Proposal

llm = GEPA.LLM.req_llm(:openai)
instruction_proposal = GEPA.Proposer.InstructionProposal.new(llm: llm)

proposer = Reflective.new(
  adapter: my_adapter,
  trainset: trainset,
  instruction_proposal: instruction_proposal
)

Without LLM

If the adapter does not implement propose_new_texts/4 or /3, pass a custom_candidate_proposer function. GEPA no longer ships a production placeholder mutation path.

Summary

Types

t()

Functions

new(opts)

propose(proposer, state)

Propose a new candidate through reflective mutation.

Types

t()

@type t() :: %GEPA.Proposer.Reflective{
  adapter: term(),
  batch_sampler: module() | struct(),
  callbacks: [term()] | nil,
  candidate_selector: module() | struct(),
  custom_candidate_proposer: function() | nil,
  instruction_proposal: GEPA.Proposer.InstructionProposal.t() | nil,
  minibatch_size: pos_integer(),
  module_selector: module() | struct(),
  perfect_score: float(),
  skip_perfect_score: boolean(),
  trainset: GEPA.DataLoader.t()
}

Functions

new(opts)

@spec new(keyword()) :: t()

propose(proposer, state)

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

Propose a new candidate through reflective mutation.

Algorithm:

Select candidate from Pareto front
Sample minibatch from training set
Evaluate with trace capture
Check for perfect scores (optional skip)
Generate improved version:
- Adapter propose_new_texts
- Custom proposer
- LLM-backed instruction_proposal
Evaluate new candidate
Return proposal if improved