Nace AI Releases Drex for Search to Aid AI Agents

This system is designed to be called by coding agents such as Claude Code, Codex, Cursor, and OpenCode via CLI, SDKs, MCP, or REST.

In Nace's own 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. The system is described as a hosted search solution capable of searching over one million files and can be paired with answering models like GPT-5, Gemini 3.5 Flash, or Qwen 27B.

Drex for Search can be deployed on Nace's managed cloud, dedicated clouds, or users' own AWS or Azure accounts, as well as on-premises and in air-gapped environments. Nace reports that the system ranks first or second on three out of four public leaderboards, including achieving 77.5% R@1 on VisDocAgentBench, which is 10 points higher than Claude Opus 5.

Nace AI frames search as a series of decisions, and Drex for Search applies its models to these decisions by scoring options with a calibrated probability in a single forward pass, making it cost-effective for each agent step. The system targets limitations in classic RAG pipelines, such as answers being located in tables, requiring cross-file hops, or being within scanned PDFs.

Within an agent, Drex for Search operates in three steps: Classify, where it determines if a question requires a quick lookup or an agent route; Squeeze, where it narrows down the search space by navigating through folders, files, sections, and tables; and Answer, where only relevant pages and tables are sent to the LLM, with citations to the file and page number.