Researchers evaluate agentic retrieval, which combines large language model reasoning with a retriever in a ReAct-style loop, against standard dense retrieval for complex search tasks where surface-level semantic similarity falls short. The agentic approach improved nDCG@10 by 8.7 points using the same embedding model and generalized well across domains, achieving competitive results on both the ViDoRe v3 and BRIGHT benchmarks where specialized retrieval methods often struggle out-of-domain. The gains come at a steep cost, however: agentic retrieval took an average of 107.4 seconds per query versus 0.67 seconds for standard retrieval, and consumed 764,100 input tokens and 5,800 output tokens per query. The authors say the results demonstrate agentic retrieval’s effectiveness while motivating further work on more cost-efficient retrieval agents for large-scale deployment.
