From the source
Perplexity's explainer sets out what RAG architecture is and walks through the main variants, describing RAG as a process in which LLMs use external information to expand on and improve their outputs.
- It lays out the foundational template, often called simple or naive RAG, built from an information source, a retrieval component and an orchestration layer that compiles retrieved text into an expanded prompt.
- The piece splits the components into an ingestion path, which organizes data into a searchable source such as a vector database, and a retrieval path, which turns a user prompt into a query and adds the results to the LLM's context.
- It then covers hybrid RAG, which combines keyword and semantic retrieval and may rerank merged results before passing them to the model, and adaptive RAG, where a routing component picks a retrieval strategy based on the prompt.
- Corrective RAG is described as evaluating retrieved information before it reaches the model, triggering further requests when the material is incomplete or not applicable.
- The explainer cautions that treating RAG types as completely distinct architectures is misleading, since most engineers implement separate components modularly and the models overlap significantly.