Nace AI has launched Drex for Search, a retrieval layer designed to help agents locate evidence within documents. The system integrates with coding tools like Claude Code, Codex, Cursor and OpenCode via CLI, SDKs, MCP or REST. According to Nace’s own evaluation, it achieves 92.05% accuracy on the BrowseComp-Plus benchmark using GPT-5, requiring just 12.6 search calls per question.
Key figures
- Capacity: A hosted search engine capable of indexing over 1 million files.
- Deployment: Available on Nace’s managed cloud, dedicated clouds, your own AWS or Azure accounts, or on-premises air-gapped environments.
- Rankings: Nace claims a top-two position on three of four public leaderboards.
- VisDocAgentBench: Achieved 77.5% R@1, placing it 10 points ahead of Claude Opus 5 (67.5%).
What the release is
Nace AI treats search as a sequence of decisions. Drex for Search applies its decision models to these steps. The system scores provided options with a calibrated probability in a single forward pass. This efficiency makes it viable for every step an agent takes.
The tool targets standard RAG workflows, which Nace summarises as chunk, embed, top-k, hope.
The pipeline fails in three specific scenarios:
- The answer is located in a table on page 23.
- The answer requires two hops across different files.
- The answer is contained within a scanned PDF.
How it works inside an agent
Nace outlines three steps:
- Classify: Before any search begins, Drex determines if the query needs a quick lookup or an agent. The agent route reads documents and queries tables.
- Squeeze: Drex navigates the workspace level by level: folder, file, section, page. Inside a sheet, it narrows down to table and column. Each level retains only what can hold the answer.
- Answer: Only the selected pages and tables reach the LLM. The response cites the file and page.
Depth is fixed, so Nace states that a million files take as many steps as a thousand.
Impact on coding agents
Nace AI tested nine agent and model pairs on OfficeQA Pro, a 133-question benchmark over U.S. Treasury Bulletins. The results showed 6% more correct answers, 62% fewer steps, 56% fewer tokens, and a 2.3x speed increase.
| Harness | Model | Accuracy as shipped | Accuracy with Drex for Search | Agent turns as shipped | Agent turns with Drex for Search |
|---|---|---|---|---|---|
| Gemini CLI | Gemini 3.1 Pro | 54.9% | 61.7% | 22.9 | 10.4 |
| Codex | GPT-6 sol | 66.2% | 70.7% | 14.3 | 8.6 |
| Hermes Agent | Opus 5.5 | 64.3% | 68.4% | 8.5 | 3.7 |
| Hermes Agent | Qwen3.8-27B | 56.4% | 60.2% | 20.1 | 6.5 |
| Claude Code | Opus 5.5 | 66.5% | 69.9% | 18.5 | 9.1 |
| pi | Opus 5.5 | 65.4% | 68.4% | 7.7 | 3.7 |
| OpenCode | Qwen3.8-27B | 51.1% | 53.4% | 23.7 | 10.0 |
| pi | Qwen3.8-27B | 43.6% | 45.9% | 24.3 | 11.1 |
| Codex | GPT-6 luna | 54.9% | 53.4% | 16.8 | 11.3 |
Workspace contents
- Files: PDFs and scans, xlsx, csv and Parquet, mp3 and mp4, docx, pptx, md, json and txt.
- Citations: Every answer names the file and the page, row or minute it came from.
- Workspaces: Isolated environments with their own index and knowledge graph, shareable with invited members.
- Storage: Upload files or folders, or keep workspaces in your own S3 bucket or Azure Blob container.
- Zero retention: Available when using your own storage.
- Domains: Tax, M&A, SEC 10-K, Contracts, or your own.
Performance on BrowseComp-Plus
BrowseComp-Plus asks a deep-research agent 830 multi-hop questions over about 100K fixed web documents. With GPT-5, Nace AI reports:
- 92.05% accuracy, #2 of 86 entries on the leaderboard.
- 12.61 search calls per question, 16.6x fewer than the only entry ahead (209.27 calls).
- 87.32% in a single pass with 8.93 search calls, the best of any entry under 10 calls.
- +20.4 points over the Qwen3-Embed-8B baseline, with the same GPT-5 model.
Document benchmark scores
On VisDocAgentBench, an agent finds 1 target page among 2,375 arXiv pages. Drex for Search scores 77.5% R@1 with an open-weight Qwen 27B planner. The best published agent, Claude Opus 5, scores 67.5%. R@1 is 97.5% on direct questions, 77.5% on one bridge, and 57.5% on two.
On MADQA, agents answer over 800 PDFs and cite pages. Nace reports 94.7% accuracy with Gemini 3.5 Flash and a page-attribution F1 of 88.2. Its effort calibration (Kuiper) is 6.66. Nace ranks it #1, based on the leaderboard as read on October 20. With open-weight Qwen 27B, Nace reports 93.4%.
On EnterpriseRAG-Bench, with 500,000+ internal company documents, Nace reports an overall score of 88.22, which it ranks #3. That is 27.2 points above OpenAI File Search (61.03).
Comparison with published results
Ranks and scores are as published by Nace AI, from leaderboards read October 20, 2026.
| Benchmark (Drex rank, per Nace) | System | Organization | Answer model | Score |
|---|---|---|---|---|
| BrowseComp-Plus, accuracy and search calls per question (#2 of 86) | Hybrid Qwen2-7B + Reason-ModernColBERT | AI21 Labs | GPT-5 | 95.18%, 209.27 calls |
| Drex for Search | Nace AI | GPT-5 | 92.05%, 12.61 calls | |
| Mixedbread Search (v3) | Mixedbread | GPT-5 | 90.48%, 11.53 calls | |
| Reason-ModernColBERT | LightOn | GPT-5 | 87.59%, 13.27 calls | |
| VisDocAgentBench, R@1 (#1) | Drex for Search | Nace AI | Qwen 27B (planner) | 77.5% |
| Tool-using agent | Anthropic | Claude Opus 5 | 67.5%Source Read original →Related articles   |