Drex for Search scores 92.05% on BrowseComp-Plus

Nace AI has released Drex for Search, expanding its Drex family of decision models to document retrieval for software agents. The company reports the system achieves 92.05% accuracy on BrowseComp-Plus benchmarks while using substantially fewer search operations than competing approaches.

What changed

Nace AI unveiled Drex for Search as a retrieval layer built to help agents locate and cite evidence from document collections. According to the announcement, Drex for Search scores 92.05% on BrowseComp-Plus when paired with GPT-5, making 12.6 search calls per question. The only entry ranking higher used 209.27 search calls per question, meaning Drex for Search achieves comparable accuracy with 16.6 times fewer queries.

The system integrates with coding agents including Claude Code, Codex, Cursor and OpenCode through command-line interfaces, software development kits, model context protocol (MCP) connections and REST APIs. Nace positions the tool as a decision-making layer that operates before documents are retrieved. According to Nace, Drex classifies each question to determine whether it needs a quick lookup or agent-driven document reading. It then narrows search scope level by level through folder, file, section, page, table and column to only areas that could contain the answer. The company states that search depth remains fixed regardless of file count, so a million files require the same number of steps as a thousand.

On the VisDocAgentBench benchmark, which asks agents to find target pages within 2,375 arXiv documents, Drex for Search reports 77.5% R@1 with an open-weight Qwen 27B model. Nace reports that Claude Opus 5 scored 67.5% on the same benchmark, a 10-point gap. On MADQA, an evaluation over 500 test questions across PDFs, Nace reports 94.7% accuracy with Gemini 3.5 Flash. On EnterpriseRAG-Bench, which involves 500,000 plus internal company documents, the company claims an 88.22 overall score, ranking third. Nace says this exceeds OpenAI File Search’s reported 61.03.

Nace tested Drex for Search against nine agent and model pairs on OfficeQA Pro, a 133-question benchmark covering U.S. Treasury Bulletins. The company reports agents showed 6% higher accuracy, used 62% fewer processing steps, consumed 56% fewer tokens and completed queries 2.3 times faster compared to baseline configurations.

Why it matters

Agents that search documents currently spend tokens reading irrelevant passages and miss answers buried in tables or distributed across multiple files. For teams running coding agents over contracts, tax documents, SEC filings or internal spreadsheets, improvements in search efficiency translate to lower token costs and faster agent responses. The reduction in steps matters for agents that rely on tool calls or iterative refinement.

Since Drex makes classification and narrowing decisions upfront rather than post-retrieval filtering, agents theoretically reduce hallucination risk by limiting the LLM’s view to genuinely relevant sections. This upfront decision approach differs from traditional retrieval-augmented generation pipelines, which chunk documents and embed them. That conventional approach loses table structure and struggles with multi-page documents or files requiring cross-references.

The system supports file types that break conventional RAG. Scanned PDFs receive optical character recognition while preserving layout, so tables inside them remain structured. Spreadsheets and Parquet files stay as tables, meaning answers can point to specific rows and cells rather than page ranges. Audio and video files receive support with answers citing the timestamp. This addresses real gaps in classical RAG, which typically chunks documents and embeds them uniformly.

Deployment flexibility matters for regulated industries. Drex for Search runs on Nace’s managed cloud, a dedicated cloud tier with private links, your own AWS or Azure account, or on-premises and air-gapped environments. The company commits to no training on customer content, encryption in transit and at rest, and SOC 2 Type II certification at every tier.

What to test

Before adopting Drex for Search in production, teams should validate several vendor claims:

• Benchmark reproducibility: The leaderboard scores cited (92.05% on BrowseComp-Plus, 77.5% on VisDocAgentBench, 94.7% on MADQA) should be independently verified. Confirm the exact evaluation setup, model pairs and dataset versions match Nace’s methodology.
• Cost per query: Compare token consumption and API call costs against your current RAG pipeline. Nace reports fewer tokens but does not publish per-query pricing. Request estimates for your typical document volumes and query patterns.
• Latency under load: Test end-to-end query latency with your agent framework and file volumes. Nace claims depth is fixed regardless of file count, but real-world performance depends on table and metadata complexity.
• Citation accuracy: Verify that citations reliably point to the correct page, row or timestamp. Spot-check outputs on your domain-specific documents.
• File type support: Confirm OCR quality on scanned documents and table extraction fidelity on your spreadsheet formats. Test with redacted or image-heavy PDFs.
• Workspace isolation: If you have multi-tenant use cases, verify that workspaces are truly isolated and that search scope does not leak across projects.
• Integration friction: Evaluate the CLI, SDK and MCP onboarding. Test whether agent frameworks like Cursor or Claude Code pick up the MCP server without manual configuration.

Common questions

How does Drex for Search differ from embedding-based RAG?

Drex applies its decision model to classify questions and hierarchically narrow search scope before retrieval. Classical RAG embeds document chunks and returns top-k similar passages. Drex’s approach avoids post-retrieval filtering by making structural decisions upfront. This helps when answers sit in tables, require multi-hop reasoning across files, or are inside scanned PDFs.

Can I run Drex for Search on my own infrastructure?

Yes. Nace offers four deployment modes: managed cloud, dedicated cloud, your AWS or Azure account, and on-premises or air-gapped. The company states that the same models and API are available at every tier. Data stays inside your chosen boundary, and the company enforces SOC 2 Type II certification at each deployment option.

What happens to my files after I delete a workspace?

Files remain in the workspace until you delete the workspace or your account. Nace commits to no training on customer content. If you upload files to a workspace, they are indexed but not retained after deletion. For maximum control, you can store workspaces in your own S3 bucket or Azure Blob container, keeping zero files on Nace’s infrastructure.

How does Drex for Search handle spreadsheets and tables differently from PDFs?

Spreadsheets and Parquet files are kept as tables, so answers point to the exact row and cell. PDFs and scans are OCR’d and laid out, with tables inside them preserved as table structures. Audio and video answers cite the minute. This structured approach allows answers to be precise rather than paragraph-level citations.

The conclusion

Drex for Search represents a refinement to retrieval for agent workflows. The benchmarks show strong performance on multi-hop and document-centric tasks, with substantial reductions in token consumption and query count compared to traditional RAG. Support for scanned documents, tables, audio and video addresses real limitations of embedding-based retrieval.

The main uncertainties are whether benchmark improvements translate to your specific document domain and whether the cost savings outweigh integration effort. The system requires agents to call it as a tool, which adds latency compared to inline embedding lookups. For teams already running agent workflows and decision models, adoption is straightforward. For teams on simpler keyword or BM25 search, the ROI depends on how much token waste and latency matter.

We recommend starting with a free trial on a representative corpus of a few dozen documents and queries. Measure baseline agent performance on your tasks, then run the same queries through Drex for Search and compare accuracy, token count and wall-clock time. Confirm that citations point to the right places. If latency and accuracy both improve, move to a paid tier that matches your data residency requirements.

Watch for updates to the Drex family itself. Nace continues to refine decision models across use cases, and improvements to the core model will likely flow into Drex for Search without requiring changes to your integration.

Related reading

• Carbon-X: 14.8B open DNA model scores 66.7% on benchmarks
• Exa 3.0 search API: Tops Perplexity and Brave on 8 benchmarks
• Drex Max: $0.06/M tokens, 67.59 on Decision Index