Document Search for AI Agents: New Drex Model Boasts 92.05% Accuracy, Only 12.6 Calls Per Task
Nace AI has launched Drex for Search, a document search module built specifically for AI agents based on its self-developed Drex decision model. The model achieves 92.05% accuracy on the complex retrieval benchmark BrowseComp-Plus, requiring an average of only 12.6 search calls per task. In office document Q&A scenarios, it also reduces AI agent interaction rounds, addressing common pain points of current AI agents: inaccurate retrieval and high operating costs.
On October 26, 2026, AI company Nace AI officially released Drex for Search, bringing its self-developed Drex decision model to the document retrieval scenario for AI agents.
Anyone who has developed AI Agents is familiar with this problem: when connecting a general search interface to an agent, issues often arise when handling long documents or complex data queries. Either critical information is not accurately captured after multiple retrievals, leading to high response error rates; or repeated calls to search tools pile up interaction rounds, which slows response times and increases token and API call costs. The root causes of irrelevant answers and long load times that ordinary users encounter when using AI assistants to query internal documents or organize public information mostly lie in the retrieval process.
The official released two sets of benchmark test results targeting these core pain points.
The first set comes from the industry’s general complex retrieval benchmark BrowseComp-Plus, which tests AI’s ability to locate specific information among a large number of web pages and documents. Drex for Search achieves an accuracy rate of 92.05%, with an average of only 12.6 search calls per single test task.
The second set comes from the office document Q&A benchmark OfficeQA Pro, which focuses on AI’s capability to handle daily office files such as spreadsheets, reports, and emails. After integrating this model, the number of interaction rounds required for AI agents to complete Q&A has decreased significantly.
For teams developing AI agents, directly integrating this search capability eliminates the need to build retrieval ranking and query path planning strategies from scratch, reducing operating costs while ensuring accuracy. For ordinary end users, the most visible changes are faster information retrieval by the AI, more stable output results, and no need for repeated follow-up corrections.
This article is compiled from a report by Asif Razzaq of MarkTechPost.
发布时间: 2026-10-27 00:46