- Lowered max async LLM processes to 4 - Enabled LLM cache for entity extraction - Reduced max parallel insert to 2