Context Awareness Gate for Retrieval‑Augmented Generation

Retrieval‑augmented generation systems often degrade when irrelevant context is retrieved for an input query. This paper proposes the Context Awareness Gate (CAG), a neural mechanism that determines whether the question requires external context before generating an answer. If retrieval is deemed unnecessary, the large language model answers solely from its internal knowledge; otherwise, the pipeline retrieves relevant documents. The authors also introduce a vector‑candidates method – a statistical, LLM‑independent approach for selecting candidate contexts – and show through experiments that CAG improves both faithfulness and relevance of generated answers by avoiding the incorporation of distracting information.

PDF