Standard RAG always retrieves documents before generating a response, regardless of whether that retrieval is actually needed. Adaptive RAG adds a decision layer that evaluates the incoming query first and decides whether retrieval is necessary, and if so, what kind of retrieval makes sense.
Simple queries get answered directly from the model's knowledge. Complex ones trigger deeper retrieval across multiple sources or multiple steps. This avoids wasting time on unnecessary lookups while ensuring that genuinely complex queries get the thorough treatment they need.
An employee asking about a standard leave policy needs a single retrieval from the HR knowledge base. An employee asking how the current policy compares to last year's version, and what the new labor law requires, needs a multi-step retrieval across policy documents, historical records, and regulatory content. Adaptive RAG handles both without treating every query the same way.
