Nace releases Drex for Search for agents
Nace has released Drex for Search, a document-search tool for AI agents that the company says cuts input tokens by 56% while modestly improving accuracy on professional-work benchmarks.
Nace announced that Drex for Search is now generally available. The tool is built on Drex, which the company calls its "decision model." Before an agent reads anything, Drex picks out where an answer is likely to be: the folder, file and page, and for spreadsheets, the table and column. The agent then reads only those places and returns its answer with the file and page it came from.
According to Nace, Drex works in three steps. It first classifies the question, then narrows the workspace one level at a time, then sends only the material it kept to the LLM. Because the number of levels is fixed, the company says, a million files take the same number of steps as a thousand. Nace presents the approach as something that directs BM25 or RAG retrieval, not a replacement for them.
All the benchmark figures are Nace's own. On 529 questions across six professional-work benchmarks, coding agents using Drex took 62% fewer steps and used 56% fewer input tokens. Accuracy rose from 63.1% to 67.1%, and median time per question fell from 80.8 to 35.5 seconds. Nace also says Drex ranks first on MADQA with 94.7% accuracy, but it gave no further detail.
On Databricks' OfficeQA Pro (133 questions over 697 scanned PDFs), all ten agent-and-model pairs tested took fewer turns. Accuracy improved for eight. It fell for Claude Code with Sonnet 5.5 (69.2% to 66.9%) and for Codex with GPT-6 luna (54.9% to 53.4%).
Drex for Search works with any agent that can run a command or call an MCP tool, including Claude Code, Codex, Cursor and OpenCode. Nace did not announce pricing.