RAG (Retrieval-Augmented Generation)

A pattern where relevant documents or data are fetched from an external store at query time and injected into the context window, so the model can answer using up-to-date or domain-specific knowledge it wasn't trained on.

Example: A support agent searches a vector database of help articles for the three most relevant docs, then includes them in the prompt before answering.

In practice: RAG is effectively a long-term memory strategy. Instead of retraining the model, you retrieve knowledge on demand. It's cheap and keeps information fresh, but retrieval quality directly caps answer quality.

• depends on Vector DatabaseRAG patterns typically rely on a vector database to find semantically relevant documents at query time.
• mitigated by HallucinationGrounding the agent's responses in retrieved documents reduces the chance of the model fabricating information.