Table 4.

Diagnostic accuracy on the GMDB set

Dataset	Model	Vanilla LLMs	Cypher RAG LLMs	Vector RAG LLMs
GMDB set selective test	GPT-3.5-turbo	59.20	65.20	71.20
	GPT-4-turbo	85.20	87.00	88.80
	GPT-4o	89.20	91.00	92.80
	Claude-3-opus	81.20	85.40	89.60
	Claude-3-sonnet	63.20	74.80	86.40
	Claude-3-haiku	71.20	77.80	84.20
	Gemini-pro	51.20	43.00	66.80
	LLaMA-70b	70.60	75.40	80.00
	Average	71.38	74.95	82.48
GMDB set non-selective test	GPT-3.5-turbo	38.40	44.60	50.80
	GPT-4-turbo	46.20	57.60	69.00
	GPT-4o	48.00	59.80	71.60
	Claude-3-opus	52.80	63.80	74.80
	Claude-3-sonnet	45.60	58.20	70.60
	Claude-3-haiku	46.20	58.00	69.80
	Gemini-pro	22.40	0.80	49.80
	LLaMA-70b	34.80	43.60	52.60
	Average	41.80	48.30	63.63

Vector RAG delivers the strongest performance, boosting average accuracy by 11.10% (selective) and 21.83% (non-selective) over vanilla LLMs, with GPT-4o achieving the highest accuracy of 92.80%. Notably, non-selective tests show a steeper performance gap, underscoring RAG’s value in challenging scenarios. The bold values represent the best performance among these three methods.