Retrieval Augmented Generation

The power of LLMs is restricted by several limitations. First, it takes a long time to train them. They are called large language models for a reason. Continuously retraining foundation models like GPT-5 and Claude Opus is impractical. At some point there is a limit to recency of the training data. If you need up to date information, it mandates a different solution. In addition, foundation models are trained on broad, public data sources. They won’t have access to private datastores behind company or enterprise intranets.

Fortunately there is a way to provide foundation models access to recent and private information. And this is called Retrieval Augmented Generation, or RAG.