Adaptive-RAG

“Adaptive-RAG jeong2024adaptive assesses query complexity based on changes in the correctness of an LLM's response, but can only utilize to queries that are answered correctly, overlooking the nuanced effects of retrieved documents beyond basic accuracy.”

“external classifiers in Adaptive-RAG often fail to fully leverage the internal decision-making capabilities of the language model. This leads to unnecessary additional retrieval steps, resulting in knowledge conflicts between the model's internal knowledge and externally retrieved information.”

“However, it is inherently difficult for the LLM to accurately assess the boundaries of its knowledge in the process of making discrete retrieval decisions”

“But the method remains impractical due to its inability to dynamically adjust the accuracy-cost trade-off. Specifically, it lacks user-driven flexibility, preventing fine-grained control over retrieval strategies in order to support diverse application needs.”

“While the latter approach relies heavily on input content characteristics. For instance, Jeong et al. define "simple questions" as single-hop queries (e.g., "When is Michael F. Phelps's birthday?") and "difficult questions" as multi-hop queries (e.g., "What currency is used in Bill Gates's birthplace?"). Such question-answering tasks have distinct difficulty gradients, making them relatively easy for models to differentiate. Unlike single-hop or multi-hop question answering tasks, input texts in the medical domain typically do not exhibit obvious structural patterns that can be captured, making it extremely challenging for smaller language models to understand the difficulty of answering them. Therefore, the successful experiences from this approach cannot be directly transferred to other tasks.”