Adaptive-RAG is a novel framework that dynamically selects the most suitable retrieval-augmented strategy for large language models (LLMs) based on query complexity. The framework uses a classifier to determine the complexity level of incoming queries, enabling the system to adapt between different retrieval strategies, from no retrieval to multi-step retrieval, depending on the query's complexity. This approach balances efficiency and accuracy by selecting the most appropriate method for each query, reducing unnecessary computational overhead for simple queries and ensuring sufficient complexity for multi-step tasks. The classifier is trained using automatically generated labels from model predictions and dataset biases, allowing it to accurately classify query complexity without human annotation. The framework is validated on open-domain QA datasets, demonstrating improved performance compared to existing adaptive retrieval methods. The results show that Adaptive-RAG enhances both the accuracy and efficiency of QA systems by dynamically adjusting strategies based on query complexity. The method is efficient, scalable, and adaptable to a wide range of query types, making it a promising solution for handling diverse user queries in real-world scenarios.Adaptive-RAG is a novel framework that dynamically selects the most suitable retrieval-augmented strategy for large language models (LLMs) based on query complexity. The framework uses a classifier to determine the complexity level of incoming queries, enabling the system to adapt between different retrieval strategies, from no retrieval to multi-step retrieval, depending on the query's complexity. This approach balances efficiency and accuracy by selecting the most appropriate method for each query, reducing unnecessary computational overhead for simple queries and ensuring sufficient complexity for multi-step tasks. The classifier is trained using automatically generated labels from model predictions and dataset biases, allowing it to accurately classify query complexity without human annotation. The framework is validated on open-domain QA datasets, demonstrating improved performance compared to existing adaptive retrieval methods. The results show that Adaptive-RAG enhances both the accuracy and efficiency of QA systems by dynamically adjusting strategies based on query complexity. The method is efficient, scalable, and adaptable to a wide range of query types, making it a promising solution for handling diverse user queries in real-world scenarios.