🧭 MoC-RAG Benchmark Leaderboard

Does routed, typed context (Mixture-of-Contexts RAG) beat flat RAG for agentic memory? This leaderboard reports retrieval, context-efficiency, routing, and robustness metrics for BM25 / dense / hybrid / metadata-filtered / reranked RAG and MoC-RAG (top_experts ∈ {1,2,3,all}).

Headline (sentence-transformers): BM25 collapses on adversarial queries (−36% vs keyword), while MoC-RAG holds and overtakes BM25 by ~+15 points on the adversarial split, carrying roughly half the hard distractors of dense RAG at 95–100% routing accuracy.

Dataset: ruslanmv/moc-rag-benchmark.

Robustness by query type (the key result)


metadata_rag	100%	81%	64%	27	104	103	-36%


bm25_rag	100%	81%	64%	27	68	72	-36%
dense_rag	83%	79%	74%	64	104	103	-9%
hybrid_rag	96%	88%	71%	52	101	91	-25%
metadata_rag	92%	88%	71%	63	107	103	-21%
reranked_rag	83%	86%	78%	56	96	97	-5%
moc_rag_e1	100%	84%	69%	22	40	42	-31%
moc_rag_e2	96%	86%	78%	26	48	48	-18%
moc_rag_e3	96%	89%	79%	30	64	62	-17%
moc_rag_all	98%	85%	77%	34	82	80	-21%


bm25_rag	78%	19%	0.7192	0.7271	167	23560	12%	3.9686	-	83%
dense_rag	78%	19%	0.5582	0.6166	271	23539	12%	3.9721	-	83%
hybrid_rag	82%	21%	0.7232	0.7403	244	23519	12%	4.2094	-	84%
metadata_rag	82%	20%	0.7031	0.7214	273	23364	13%	4.1945	-	84%
reranked_rag	82%	21%	0.7171	0.7446	249	23499	12%	4.2129	-	86%
moc_rag_e1	81%	20%	0.7484	0.7598	104	23509	12%	4.1473	84%	77%
moc_rag_e2	85%	21%	0.7677	0.7811	122	23547	13%	4.3318	95%	86%
moc_rag_e3	86%	22%	0.7889	0.7963	156	23529	13%	4.3988	99%	89%
moc_rag_all	85%	21%	0.7661	0.7788	196	23568	12%	4.3067	100%	89%

🧭 MoC-RAG Benchmark Leaderboard

Robustness by query type (the key result)

Single-split leaderboard