Nace AI has introduced Drex for Search, a tool designed to optimize decision-making processes for AI agents. By applying Drex models to search tasks, the system evaluates potential options with calibrated probabilities in a single forward pass, reducing computational costs for each agent step.
The company tested nine agent and model pairs on the OfficeQA Pro benchmark, which consists of 133 questions based on U.S. Treasury Bulletins. Results showed a 6% increase in correct answers, a 62% reduction in steps, a 56% decrease in token usage, and a 2.3x improvement in speed compared to standard RAG approaches.
On the BrowseComp-Plus benchmark, which includes 830 multi-hop questions across 100,000 web documents, the system achieved a 92.05% score using 12.6 search calls. In the VisDocAgentBench test, Drex for Search reached 77.5% R@1 with a Qwen 27B planner, outperforming Claude Opus 5 at 67.5%.
Regarding the MADQA benchmark, Nace AI reported 94.7% accuracy using Gemini 3.5 Flash and 93.4% with Qwen 27B. On EnterpriseRAG-Bench, which covers over 500,000 internal documents, the company achieved a score of 88.22, ranking third and surpassing OpenAI File Search by 27.2 points.
