NVIDIA’s NeMo Retriever research team has published findings demonstrating that agentic retrieval—embedding a retrieval system within a ReAct reasoning loop—significantly outperforms standard dense retrieval methods. The approach achieves an 8.7 nDCG@10 improvement on average using the same underlying embedding model, without any reconfiguration of the retriever itself.
The agentic pipeline places retrieval within a multi-step reasoning process where the model decides when to search, what to search for, and how to synthesize results. Unlike traditional RAG approaches that perform a single retrieval pass, the agentic method allows the model to make multiple search calls, refine queries based on intermediate results, and iterate until sufficient evidence is gathered. The study found that Opus 4.5 makes an average of 9.2 search calls per query compared to just 2.4 for gpt-oss-120b, suggesting more sophisticated reasoning leads to better retrieval outcomes.
Benchmark results are noteworthy: the system holds the top position on the ViDoRe v3 benchmark and ranks second on BRIGHT without any task-specific fine-tuning. The approach consumed 764,100 input tokens and 5,800 output tokens per query on average, with queries taking 107.4 seconds compared to 0.67 seconds for standard retrieval—a 160x increase in computational cost that reflects the trade-off between quality and efficiency.
The researchers also addressed implementation details: they replaced the Model Context Protocol with an in-process thread-safe retriever to reduce overhead, acknowledging that external tool calling introduces significant latency in production scenarios. This practical consideration suggests the approach is designed with real-world deployment in mind rather than purely academic benchmarking.
For enterprise RAG deployments, the findings indicate that investing in agentic retrieval logic may yield substantial quality gains for applications where query latency is less critical than result accuracy—particularly in legal research, scientific literature synthesis, and complex technical documentation systems.