CoinRAG: Contextualized Information Nugget KV Cache Reuse for Long-Context RAG
CoinRAG cuts RAG prefill cost by caching smaller, query-relevant “nuggets” instead of whole chunks.
The paper proposes a two-stage retrieval method that finds semantic units inside retrieved chunks, then reuses offline-computed sliced KV caches for those units with chunk-level context. The authors frame this as a way to reduce redundancy and noise left by chunk-level cache reuse. On LongBench multi-hop QA tasks, they report lower operational cost and a new Pareto frontier under low prefill latency constraints. They also report an average 5.3% relative F1 improvement under a standard fast-prefill budget. HF Daily Papers' note
The paper proposes a two-stage retrieval method that finds semantic units inside retrieved chunks, then reuses offline-computed sliced KV caches for those units with chunk-level context. The authors frame this as a way to reduce redundancy and noise left by chunk-level cache reuse. On LongBench multi-hop QA tasks, they report lower operational cost and a new Pareto frontier under low prefill latency constraints. They also report an average 5.3% relative F1 improvement under a standard fast-prefill budget. HF Daily Papers' note
score 4