Nace AI Releases Drex for Search to Aid AI Agents

In Nace AI's evaluations, Drex for Search achieved a score of 92.05% on the BrowseComp-Plus benchmark when paired with GPT-5, utilizing an average of 12.6 search calls per question. This hosted search system can manage over 1 million files and is compatible with answering models like GPT-5, Gemini 3.5 Flash, or Qwen 27B.

Drex for Search operates by framing search as a series of decisions. Before a search begins, Drex classifies whether a question requires a quick lookup or a more complex agent-based approach involving document reading and table querying. The system then "squeezes" the workspace, narrowing down relevant information level by level, from folders to specific columns within tables.

Only the most relevant pages and tables are then passed to the LLM for answering, with citations to the file and page number. Nace AI states that the depth of search is fixed, meaning a million files take the same number of steps as a thousand. The system can be deployed on Nace's managed cloud, dedicated clouds, AWS, Azure, or on-premise and air-gapped environments.

Nace AI reports that Drex for Search ranked first or second on three out of four public leaderboards. Specifically, it achieved 77.5% R@1 on VisDocAgentBench, outperforming Claude Opus 5 by 10 points. The company tested nine agent and model pairs on the OfficeQA Pro benchmark, which consists of 133 questions based on U.S. Treasury Bulletins.