跳到主内容
@wquguru
精选75MarkTechPost(RSS)产品发布/更新

Nace AI发布Drex for Search检索层

Nace AI Releases Drex for Search for AI Agents: 92.05% on BrowseComp-Plus With 12.6 Search Calls

原文
发到 X

Nace AI has released Drex for Search, which brings its Drex decision models to document search. Drex for Search is a retrieval layer that helps people and agents find evidence. Coding agents such as Claude Code, Codex, Cursor and OpenCode call it through CLI, SDKs, MCP or REST. In Nace’s evaluation, it scores 92.05% on BrowseComp-Plus with GPT-5, using 12.6 search calls per question.

Nace AI 发布了用于搜索的 Drex,将其 Drex 决策模型引入文档搜索。Drex for Search 是一个检索层,帮助人类和智能体查找证据。Claude Code、Codex、Cursor 和 OpenCode 等编码智能体通过 CLI、SDK、MCP 或 REST 调用它。在 Nace 的评估中,使用 GPT-5 时,它在 BrowseComp-Plus 上得分 92.05%,每问题平均使用 12.6 次搜索调用。

TL;DR

TL;DR(太长不看)

  • Size: a hosted search system built to search over 1M+ files. It pairs with an answering model such as GPT-5, Gemini 3.5 Flash or Qwen 27B.
  • Runs on: Nace’s managed cloud, a dedicated cloud, your own AWS or Azure account, or on-prem and air-gapped.
  • Performance: ranked #1 or #2 on 3 of 4 public leaderboards, per Nace.
  • Best: 77.5% R@1 on VisDocAgentBench, 10 points above Claude Opus 5 (67.5%).
  • 规模:一个托管搜索系统,旨在搜索超过 100 万份文件。它与 GPT-5、Gemini 3.5 Flash 或 Qwen 27B 等回答模型配合使用。
  • 运行环境:Nace 托管云、专用云、您自己的 AWS 或 Azure 账户,或本地部署及空气隔离环境。
  • 性能:根据 Nace 的数据,在 4 个公开排行榜中的 3 个上排名第一或第二。
  • 最佳表现:在 VisDocAgentBench 上 R@1 达到 77.5%,比 Claude Opus 5(67.5%)高出 10 个百分点。

What did Nace AI release?

Nace AI 发布了什么?

Nace AI frames search as a series of decisions. Drex for Search applies its Drex models to those decisions. Drex scores given options with a calibrated probability each, in 1 forward pass. That makes it cheap enough for every agent step.

Nace AI 将搜索视为一系列决策过程。Drex for Search 将其 Drex 模型应用于这些决策。Drex 在一次前向传播中为每个选项给出校准后的概率评分。这使得其成本足够低,可用于每个智能体步骤。

The release targets classic RAG, which Nace sums up as chunk, embed, top-k, hope.

此次发布针对经典 RAG(检索增强生成),Nace 将其概括为:分块、嵌入、取 top-k、然后碰运气。

The pipeline breaks in 3 cases:

该流水线在以下三种情况下会失效:

  • the answer sits in a table on page 23
  • the answer needs 2 hops across files
  • the answer is inside a scanned PDF
  • 答案位于第 23 页的表格中
  • 答案需要跨文件进行两次跳转才能找到
  • 答案位于扫描版 PDF 内部

How does Drex for Search work inside an agent?

Drex for Search 如何在智能体内部工作?

Nace describes 3 steps:

Nace 描述了三个步骤:

  • Classify: Before any search runs, Drex decides if the question needs a quick lookup or an agent. The agent route reads documents and queries tables.
  • Squeeze: Drex walks the workspace level by level: folder, file, section, page. Inside a sheet, it narrows to table and column. Each level keeps only what can hold the answer.
  • Answer: Only the kept pages and tables reach the LLM. The answer cites the file and page.
  • 分类:在任何搜索运行之前,Drex 决定问题是只需快速查找还是需要智能体介入。智能体路径会读取文档并查询表格。
  • 压缩:Drex 逐层遍历工作区:文件夹、文件、章节、页面。在电子表格内,它会进一步缩小到表格和列。每一层只保留可能包含答案的内容。
  • 回答:只有被保留的页面和表格会被发送给大语言模型。回答会引用文件和页码。

Depth is fixed, so Nace says a million files take as many steps as a thousand.

深度是固定的,因此 Nace 表示处理一百万份文件所需的步骤数与处理一千份文件相同。

How much does it help coding agents?

它对编码智能体有多大帮助?

Nace AI tested 9 agent and model pairs on OfficeQA Pro, a 133-question benchmark over U.S. Treasury Bulletins. Key highlights: 6% more answers right, 62% fewer steps, 56% fewer tokens, and 2.3x faster.

Nace AI 在 OfficeQA Pro 上测试了 9 对智能体和模型组合,这是一个基于美国财政部公报的 133 道题目的基准测试。主要亮点包括:正确率提高 6%,步骤减少 62%,令牌消耗减少 56%,速度提升 2.3 倍。

HarnessModelAccuracy as shippedAccuracy with Drex for SearchAgent turns as shippedAgent turns with Drex for Search
Gemini CLIGemini 3.1 Pro54.9%61.7%22.910.4
CodexGPT-6 sol66.2%70.7%14.38.6
Hermes AgentOpus 5.564.3%68.4%8.53.7
Hermes AgentQwen3.8-27B56.4%60.2%20.16.5
Claude CodeOpus 5.566.5%69.9%18.59.1
piOpus 5.565.4%68.4%7.73.7
OpenCodeQwen3.8-27B51.1%53.4%23.710.0
piQwen3.8-27B43.6%45.9%24.311.1
CodexGPT-6 luna54.9%53.4%16.811.3
工具模型出厂准确率Drex for Search 准确率出厂代理轮次Drex for Search 代理轮次
Gemini CLIGemini 3.1 Pro54.9%61.7%22.910.4
CodexGPT-6 sol66.2%70.7%14.38.6
Hermes AgentOpus 5.564.3%68.4%8.53.7
Hermes AgentQwen3.8-27B56.4%60.2%20.16.5
Claude CodeOpus 5.566.5%69.9%18.59.1
piOpus 5.565.4%68.4%7.73.7
OpenCodeQwen3.8-27B51.1%53.4%23.710.0
piQwen3.8-27B43.6%45.9%24.311.1
CodexGPT-6 luna54.9%53.4%16.811.3

What does a workspace include?

工作区包含哪些内容?

  • 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, with their own index and knowledge graph, and shareable with invited members.
  • Storage: upload files or folders, or keep workspaces in your own S3 bucket or Azure Blob container.
  • Zero retention: available with your own storage.
  • Domains: Tax, M&A, SEC 10-K, Contracts, or your own.
  • 文件:PDF 和扫描件、xlsx、csv 和 Parquet、mp3 和 mp4、docx、pptx、md、json 和 txt。
  • 引用:每个答案都会指明其来源的文件及页码、行或分钟数。
  • 工作区:隔离运行,拥有独立的索引和知识图谱,并可分享给受邀成员。
  • 存储:上传文件或文件夹,或将工作区保留在您自己的 S3 存储桶或 Azure Blob 容器中。
  • 零留存:支持使用您自己的存储。
  • 领域:税务、并购(M&A)、SEC 10-K 表格、合同,或自定义领域。

How does it perform on BrowseComp-Plus?

在 BrowseComp-Plus 上的表现如何?

BrowseComp-Plus asks a deep-research agent 830 multi-hop questions over about 100K fixed web documents. With GPT-5, Nace AI reports:

BrowseComp-Plus 要求深度研究智能体在约 10 万个固定网页文档中回答 830 个多跳问题。使用 GPT-5,Nace AI 报告称:

  • 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.
  • 准确率达 92.05%,在排行榜的 86 个条目中排名第 2。
  • 每问题平均调用搜索 12.61 次,比唯一领先条目少 16.6 倍(209.27 次)。
  • 单次尝试准确率为 87.32%,调用搜索 8.93 次,是调用次数低于 10 次的条目中最佳成绩。
  • 与 Qwen3-Embed-8B 基线相比提升 20.4 分,使用的是相同的 GPT-5 模型。

How does it score on document benchmarks?

在文档基准测试中的得分如何?

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.

在 VisDocAgentBench 上,智能体需在 2,375 个 arXiv 页面中找到 1 个目标页面。Drex for Search 在使用开源权重 Qwen 27B 规划器时,R@1 得分为 77.5%。已发布的最佳智能体 Claude Opus 5 得分为 67.5%。在直接问题上 R@1 为 97.5%,在一个桥接问题上为 77.5%,在两个桥接问题上为 57.5%。

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%.

在 MADQA 上,智能体需回答超过 800 份 PDF 并引用页码。Nace 报告称,使用 Gemini 3.5 Flash 时准确率为 94.7%,页面归属 F1 分数为 88.2。其努力校准(Kuiper)值为 6.66。Nace 将其排名 #1,基于截至 10 月 20 日读取的排行榜数据。使用开源权重 Qwen 27B 时,Nace 报告准确率为 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).

在 EnterpriseRAG-Bench 上,面对 50 万+ 企业内部文档,Nace 报告总分为 88.22,排名第 3。这比 OpenAI File Search(61.03)高出 27.2 分。

How does it compare with the best published results?

与已发布的最佳结果相比如何?

Ranks and scores are as published by Nace AI, from leaderboards read October 20, 2026.

排名和分数均由 Nace AI 发布,数据来源于 2026 年 10 月 20 日读取的排行榜。

Benchmark (Drex rank, per Nace)SystemOrganizationAnswer modelScore
BrowseComp-Plus, accuracy and search calls per question (#2 of 86)Hybrid Qwen2-7B + Reason-ModernColBERTAI21 LabsGPT-595.18%, 209.27 calls
Drex for SearchNace AIGPT-592.05%, 12.61 calls
Mixedbread Search (v3)MixedbreadGPT-590.48%, 11.53 calls
Reason-ModernColBERTLightOnGPT-587.59%, 13.27 calls
VisDocAgentBench, R@1 (#1)Drex for SearchNace AIQwen 27B (planner)77.5%
Tool-using agentAnthropicClaude Opus 567.5%
Tool-using agentAnthropicClaude Fable 562.5%
Tool-using agentOpenAIGPT-5.6-sol61.67%
MADQA, accuracy (#1)Drex for SearchNace AIGemini 3.5 Flash94.7%
Agentic RAGProgress SoftwareGemini 3.5 Flash + 3.6 Flash94.3%
Agentic SearchMixedbreadGemini 3.5 Flash93.4%
PPLX Embed 2 RAGPerplexityGemini 3.5 Flash92.1%
File SearchOpenAIGPT-549.6%
EnterpriseRAG-Bench, overall score (#3 of 31)Flexor.aiFlexor.aiGPT-6.1 Sol91.48
CDLCausal Dynamics LabNot listed90.91
Drex for SearchNace AINot listed88.22
MixedbreadMixedbreadClaude Opus 586.58
OpenAI File SearchOpenAIManaged product61.03
Azure AI SearchMicrosoftManaged product48.42
Vertex AI SearchGoogleManaged product41.87
基准测试(Drex 排名,基于 Nace)系统组织答案模型得分
BrowseComp-Plus,准确性和每次问题的搜索调用次数(86 个中的第 2 名)混合 Qwen2-7B + Reason-ModernColBERTAI21 LabsGPT-595.18%,209.27 次调用
Drex for SearchNace AIGPT-592.05%,12.61 次调用
Mixedbread Search (v3)MixedbreadGPT-590.48%,11.53 次调用
Reason-ModernColBERTLightOnGPT-587.59%,13.27 次调用
VisDocAgentBench,R@1(第 1 名)Drex for SearchNace AIQwen 27B(规划器)77.5%
使用工具的代理AnthropicClaude Opus 567.5%
使用工具的代理AnthropicClaude Fable 562.5%
使用工具的代理OpenAIGPT-5.6-sol61.67%
MADQA,准确率(第 1 名)Drex for SearchNace AIGemini 3.5 Flash94.7%
智能体 RAGProgress SoftwareGemini 3.5 Flash + 3.6 Flash94.3%
智能体搜索MixedbreadGemini 3.5 Flash93.4%
PPLX Embed 2 RAGPerplexityGemini 3.5 Flash92.1%
文件搜索OpenAIGPT-549.6%
EnterpriseRAG-Bench,总分(31 个中的第 3 名)Flexor.aiFlexor.aiGPT-6.1 Sol91.48
CDLCausal Dynamics Lab未列出90.91
Drex for SearchNace AI未列出88.22
MixedbreadMixedbreadClaude Opus 586.58
OpenAI 文件搜索OpenAI托管产品61.03
Azure AI SearchMicrosoft托管产品48.42
Vertex AI SearchGoogle托管产品41.87

How can developers get access?

开发者如何获取访问权限?

Drex for Search runs in the Nace API Console, with 4 ways in on one key:

Drex for Search 在 Nace API 控制台中运行,通过一个密钥提供 4 种接入方式:

  • CLI: nace-cli on PyPI.
  • SDKs: nace-sdk for Python and nace-sdk for TypeScript.
  • MCP: nace-mcp, for Claude Code, Codex, Cursor or OpenCode.
  • REST: an OpenAPI spec.
  • CLI:PyPI 上的 nace-cli。
  • SDK:用于 Python 的 nace-sdk 和用于 TypeScript 的 nace-sdk。
  • MCP:nace-mcp,适用于 Claude Code、Codex、Cursor 或 OpenCode。
  • REST:OpenAPI 规范。

Key Takeaways

关键要点

  • Drex for Search gives agents cited pages and tables, not top-k chunks.
  • 92.05% on BrowseComp-Plus with 16.6x fewer search calls than the only entry ahead.
  • 77.5% R@1 on VisDocAgentBench, 10 points above Claude Opus 5.
  • On OfficeQA Pro, agent turns fell for all 9 harness and model pairs Nace tested.
  • Drex for Search 为代理提供引用的页面和表格,而非前 k 个文本块。
  • 在 BrowseComp-Plus 上达到 92.05%,搜索调用次数比唯一领先条目少 16.6 倍。
  • 在 VisDocAgentBench 上 R@1 为 77.5%,比 Claude Opus 5 高出 10 个百分点。
  • 在 OfficeQA Pro 上,所有 9 个 Nace 测试的框架和模型组合中,代理均未能成功应对。

Check out the Drex for Search page, API Console, Docs and Hugging Face.

查看 Drex for Search 页面、API 控制台、文档和 Hugging Face。

Thanks to the Nace AI team for the thought leadership / resources for this article. This article is sponsored by Nace AI.

感谢 Nace AI 团队为本篇文章提供的思想领导力/资源。本文由 Nace AI 赞助。

更进一步:量化金融体系

看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力

进入量化体系 →

关联讨论

同一事件的更多信源

相似阅读

关联信息,但可能不是同一事件