Glossary term
Glossary term
Memory and Retrieval
Adaptive RAG is a method that dynamically adjusts how and when external information is retrieved based on the complexity of a query. Instead of using a fixed retrieval process, it determines whether retrieval is needed and selects the appropriate strategy, such as no retrieval, single-step retrieval, or multi-step retrieval. This improves efficiency, reduces unnecessary computation, and ensures better response quality across different query types.
LangGraph implementations of Adaptive RAG route simple factual questions directly to the LLM while sending complex multi-hop questions through multi-step retrieval.
The original Adaptive RAG paper by Jeong et al. demonstrated query-complexity classification on TriviaQA and HotpotQA datasets.
LlamaIndex offers an Adaptive RAG pattern that classifies queries between no-retrieval, single-step, and multi-step retrieval flows.