Nace AI Releases Drex for Search for AI Agents, Reports 92.05% on BrowseComp-Plus

Drex for Search brings Nace's Drex decision models to document search. Coding agents such as Claude Code, Codex, Cursor and OpenCode call it through CLI, SDKs, MCP or REST. Nace said the hosted system is built to search more than 1 million files and pairs with an answering model such as GPT-5, Gemini 3.5 Flash or Qwen 27B. It runs on Nace's managed cloud, a dedicated cloud, a customer's AWS or Azure account, or on-premises and air-gapped environments.

Nace frames search as a series of decisions. Drex scores given options with a calibrated probability each in one forward pass, which Nace says makes it cheap enough for every agent step. The release targets classic RAG, which Nace describes as chunk, embed, top-k and hope. Nace says that pipeline breaks when the answer sits in a table on page 23, requires two hops across files, or is inside a scanned PDF.

Inside an agent, Nace describes three steps. Classify decides whether a question needs a quick lookup or an agent; the agent route reads documents and queries tables. Squeeze walks the workspace level by level, from folder to file to section to page, and inside a sheet narrows to table and column, keeping only what can hold the answer. Answer sends only the kept pages and tables to the LLM, and the answer cites the file and page. Depth is fixed, so Nace says a million files take as many steps as a thousand.

Nace tested nine agent and model pairs on OfficeQA Pro, a 133-question benchmark over U.S. Treasury Bulletins. It reported 6% more answers right, 62% fewer steps, 56% fewer tokens and 2.3 times faster performance. For example, Nace reported that Gemini CLI with Gemini 3.1 Pro rose from 54.9% to 61.7% while agent turns fell from 22.9 to 10.4, and Codex with GPT-6 sol rose from 66.2% to 70.7% while turns fell from 14.3 to 8.6.

On BrowseComp-Plus, which asks a deep-research agent 830 multi-hop questions over about 100,000 fixed web documents, Nace reported 92.05% accuracy with GPT-5, ranking second among 86 entries. The system used 12.61 search calls per question, 16.6 times fewer than the only entry ahead, AI21 Labs' Hybrid Qwen2-7B plus Reason-ModernColBERT, which Nace lists at 95.18% and 209.27 calls. Nace also reported 87.32% in a single pass with 8.93 search calls, which it says is the best among entries using fewer than 10 calls, and a 20.4-point gain over the Qwen3-Embed-8B baseline with the same GPT-5 model.

On VisDocAgentBench, where an agent finds one target page among 2,375 arXiv pages, Drex for Search scored 77.5% R@1 with an open-weight Qwen 27B planner, against 67.5% for Claude Opus 5, according to Nace. Nace reported R@1 of 97.5% on direct questions, 77.5% on one bridge and 57.5% on two. On MADQA, over 800 PDFs with page citations, Nace reported 94.7% accuracy with Gemini 3.5 Flash, page-attribution F1 of 88.2 and effort calibration, called Kuiper, of 6.66; it ranked first based on the leaderboard as read on Oct. 20. With open-weight Qwen 27B, Nace reported 93.4%.

On EnterpriseRAG-Bench, with more than 500,000 internal company documents, Nace reported a score of 88.22 and ranked third, 27.2 points above OpenAI File Search at 61.03. Nace said the system ranks first or second on three of four public leaderboards, with results and ranks as published by Nace from leaderboards read on Oct. 20, 2026.

Workspaces can include PDFs and scans, xlsx, csv and Parquet, mp3 and mp4, docx, pptx, md, json and txt files, and answers cite the file and page, row or minute. Workspaces are isolated with their own index and knowledge graph, shareable with invited members, and can use Nace storage or a customer's S3 bucket or Azure Blob container, with zero retention available with customer storage. Domains include tax, M&A, SEC 10-K, contracts or a custom domain.

Editor's Summary Nace AI has released Drex for Search, a retrieval layer for agents and people that cites file and page. Nace reports 92.05% on BrowseComp-Plus with GPT-5 and 12.6 search calls per question, along with top rankings on several document benchmarks. The figures are Nace's published results from leaderboards it read on Oct. 20, 2026, and are compared against named systems from AI21 Labs, Anthropic and OpenAI.