Background
Cache-augmented generation reduces repeated retrieval costs, but unconditional writes retain low-quality or irrelevant content and affect later answers.
Model Design
CAG-RAG uses two-stage selective write-back. The first stage stores high-quality content; the second reassesses remaining candidates and admits only those passing a quality threshold. Experiments use SQuAD and HotpotQA to compare plain RAG, unconditional write-back and selective write-back.
Quality Gate
Evaluates content quality and relevance before each write-back.
Two-stage Selection
Stores high-quality content first, then reassesses remaining candidates.
Stable Cache
Reduces cache contamination to maintain stable outputs during prolonged service.
Results
The research examines how selective write-back affects generation quality, cache efficiency and stability in noisy environments.
research architecture
