A visual breakdown of Retrieval-Augmented Generation (RAG) versus Cache-Augmented Generation (CAG) has been making the rounds among AI engineers. Both architectures aim to improve LLM outputs but take fundamentally different approaches.

RAG works by fetching relevant documents from an external database at query time and feeding them into the model context. This gives the model access to up-to-date information beyond its training data, making it ideal for knowledge-intensive tasks where accuracy on recent data matters.

CAG, on the other hand, pre-computes and caches relevant information so the model can access it without live retrieval. This trades flexibility for speed, since the cached knowledge is static but the lookup is near-instant. The choice between them depends on whether your priority is freshness or latency.

Source