Summary

This PR introduces global, user-configurable controls for LLM generation retry effort via RunConfig, and removes prior module-level hardcoding inside generation tasks (notably llm_completion generators and LLM judge). This is a necessary change so users can explicitly bound/scale effort spent per sample (latency/cost/stability) across all tasks that rely on ModelFacade.generate().

What changed

• Added new RunConfig fields (single source of truth):
  • max_conversation_restarts: int (default 5, ge=0)
  • max_conversation_correction_steps: int (default 0, ge=0)
• Removed generator-local hardcoded retry defaults in llm_completion generators
  • Generators now read retry settings from self.resource_provider.run_config
• Updated LLM judge behavior
  • Previously, LLMJudgeCellGenerator used a 2× restart multiplier relative to the module default.
  • Now, LLM judge uses the same global RunConfig values as other LLM completion generators.

Why this is needed (user impact)

LLM generation effort per sample was previously implicitly determined by internal defaults (and judge had an additional multiplier), making it difficult for users to:

• Cap worst-case latency for dataset generation
• Cap cost (token usage) per sample
• Tune quality vs throughput consistently

With this change, users can control these tradeoffs in one place.

Behavioral impact / compatibility notes

• Defaults preserve current non-judge behavior: RunConfig.max_conversation_restarts=5, max_conversation_correction_steps=0 matches the previous LLM generator defaults.
• LLM judge behavior changes: judge no longer gets extra restart budget. Users who relied on higher judge persistence can now set higher global values explicitly via RunConfig.

Plumbing details

• DataDesigner already threads RunConfig into the engine via ResourceProvider.run_config.
• LLM completion generators (including judge) now pass run_config values into:
  • ModelFacade.generate(max_correction_steps=..., max_conversation_restarts=...)

Example usage

from data_designer.essentials import DataDesigner, RunConfig

dd = DataDesigner()
dd.set_run_config(
    RunConfig(
        max_conversation_restarts=2,
        max_conversation_correction_steps=1,
    )
)

Files changed

• src/data_designer/config/run_config.py
• src/data_designer/engine/column_generators/generators/llm_completion.py
• tests/engine/column_generators/generators/test_llm_completion_generators.py
• tests/interface/test_data_designer.py

Tests

• Updated unit tests verify defaults/persistence and that generators/judge pass run_config values into model.generate(...).