Self-RAG vs CRAG in LangGraph: which corrective retrieval pattern fits production RAG?
CRAG is better when retrieval ambiguity is the problem because it adds a lightweight evaluator plus web-search fallback, while Self-RAG is better when you want the model itself to self-reflect through retrieval and support checks — but Self-RAG’s richer control logic usually costs more LLM calls, so the best choice depends on latency budget and how much correction you need.