Graph-Native Optimization State (CONCEPT:ORCH-1.31)¶
Overview¶
Persists the GEPA Pareto frontier (candidates + ancestry) into the durable epistemic-graph so optimization is resumable and accumulates across sessions — a synergy unique to our graph-native architecture (neither predict-rlm nor the GEPA reference impl does this). Extends ORCH-1.13 (+KG-2.7 ingestion/persistence).
How it works¶
- Snapshot round-trip (
ParetoCandidatePool.to_snapshot/load_snapshot) — pure serialization of the frontier to/from plain dicts, preserving candidate prompts, scores, andparent_idsancestry. - Persist / resume (
GEPAOptimizer.persist_frontier(run_id)/resume_frontier(run_id)) — writes aGEPAFrontiergraph node (extending the existingOptimizationTrajectoryNode/EvaluatorFeedbackNodewrites viacreate_or_merge_node) and reads it back to seed the pool. A killed run resumes from the persisted frontier with no loss of the best candidate; prior frontiers are reusable optimization state. - Best-effort — absent a backend, persist/resume degrade quietly (return falsy), never raising.
Key files / API¶
| Piece | Location |
|---|---|
| Snapshot + persistence | rlm/gepa.py (ParetoCandidatePool.to_snapshot/load_snapshot, GEPAOptimizer.persist_frontier/resume_frontier) |
Wiring (≤3 hops)¶
graph_orchestrate(action="rlm_optimize") → optimizer → persist_frontier/resume_frontier → graph (≤3 hops).
Research provenance¶
Synthesis of GEPA (frontier/ancestry) with agent-utilities' durable epistemic-graph persistence —
verified against rlm/gepa.py (OptimizationTrajectoryNode/EvaluatorFeedbackNode).