Retrieval-Augmented Generation is a pattern where a language model's response is grounded in content retrieved from an external knowledge base at query time. Instead of relying solely on parametric knowledge baked into the model's weights, the system retrieves relevant passages — from a vector database, a document store, or a live API — and passes them to the model as additional context.
RAG emerged in 2020-2021 as the dominant approach to reducing hallucination and enabling domain-specific question answering on proprietary data that is not in the model's training corpus. For enterprise deployments, RAG is often the minimum-viable architecture for any use case that touches internal documents or time-sensitive information.
The rise of very-long-context models and agentic frameworks using MCP has shifted but not replaced RAG. Long-context models can sometimes ingest entire document sets directly. Agentic frameworks can retrieve on demand via tool calls. But classical RAG — embed, index, retrieve, generate — remains the most economical and debuggable architecture for most enterprise information-retrieval use cases.