--- title: Nace AI's Drex for Search: page-level retrieval for... description: Nace AI released Drex for Search, a retrieval layer that helps AI agents find cited evidence inside large file collections. Instead of returning top-k chunks... canonical: https://daily.dev/posts/nace-ai-s-drex-for-search-page-level-retrieval-for-agents-working-over-file-collections-hqcrdgatq twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Nace AI's Drex for Search: page-level retrieval for agents working over file collections | daily.dev og:description: Nace AI released Drex for Search, a retrieval layer that helps AI agents find cited evidence inside large file collections. Instead of returning top-k chunks... og:url: https://daily.dev/posts/nace-ai-s-drex-for-search-page-level-retrieval-for-agents-working-over-file-collections-hqcrdgatq og:image: https://api.daily.dev/og/posts/HqCRdgAtq.png og:image:alt: Nace AI's Drex for Search: page-level retrieval for agents working over file collections og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Nace AI's Drex for Search: page-level retrieval for agents working over file collections **[Collections](https://daily.dev/sources/collections)** · 2 min read · 1 upvotes · 0 comments ## Summary Nace AI released Drex for Search, a retrieval layer that helps AI agents find cited evidence inside large file collections. Instead of returning top-k chunks by similarity, it narrows from folder to file to page, then returns an answer with a supporting quote and page number. It reads files from AWS S3, Azure, and DigitalOcean, including scanned PDFs, spreadsheets, email, and audio, and ships as a CLI, Python and TypeScript SDKs, MCP, and a REST API for agents like Claude Code, Codex, and Hermes Agent. Nace reports 62% fewer agent steps, 56% fewer input tokens, and accuracy rising from 63.1% to 67.1% across 529 queries. On BrowseComp-Plus it scored 92.05% with GPT-5 using 12.61 search calls per question, 16.6 times fewer than the only entry ranked above it. These are the vendor's own figures, and the BrowseComp-Plus coverage was sponsored by Nace AI. ## Content Nace AI has released Drex for Search, a retrieval layer that helps AI agents find cited evidence in large file collections. Standard top-k chunk retrieval returns the passages that look closest to the query. Drex works differently: it narrows the search step by step, from the likely folder to the file to the page, then returns the answer, the supporting quote, and the page number. It stops once it can cite a source. ## What it connects to Drex can read files stored in AWS S3, Azure, and DigitalOcean. It handles scanned PDFs, spreadsheets, email, and audio. It is available through a CLI, Python and TypeScript SDKs, MCP, and a REST API. Coding agents such as Claude Code, Codex, and Hermes Agent can use it as a tool. ## Reported results Nace's numbers: - **Efficiency:** 62% fewer agent steps and 56% fewer input tokens per question. - **Accuracy:** Across 529 queries from six professional work benchmarks, including Databricks' OfficeQA, agents using Drex answered 67.1% correctly, up from 63.1%. - **BrowseComp-Plus:** 92.05% with GPT-5, using 12.61 search calls per question. That is 16.6 times fewer calls than the only entry ranked above it. - **VisDocAgentBench:** 77.5% R@1 with an open-weight Qwen 27B planner, versus 67.5% for Claude Opus 5. - **OfficeQA Pro:** Across 9 agent and model pairs, Drex cut agent turns and tokens while improving accuracy. These are the vendor's own figures, and the coverage of BrowseComp-Plus was sponsored by Nace AI. ## Why it matters @omarsar0 called it a big upgrade for agents working over large file collections, since it decides where to look and reads only the pages that matter. He pointed to Claude Code, Codex, and Hermes Agent, where poor retrieval often leads to wrong answers. The citation step is the part I find most useful. An agent that has to return a quote and a page number is much easier to check than one that hands back a confident paragraph. ## Questions this post answers ### What is Drex for Search and how does it differ from standard chunk retrieval? Drex for Search is a retrieval layer from Nace AI that narrows searches step by step from folder to file to page, rather than returning top-k similar chunks. It stops once it can cite a source, returning the answer, a supporting quote, and the page number. It works across AWS S3, Azure, and DigitalOcean, handling scanned PDFs, spreadsheets, email, and audio. _Teams evaluating retrieval approaches for agent workflows can track tool comparisons like this on daily.dev._ ### How much more efficient is Drex compared to standard agent retrieval methods? Drex reduces agent steps by 62% and input tokens by 56% per question, according to Nace AI's reported figures. Across 529 queries from six professional work benchmarks including Databricks' OfficeQA, accuracy rose from 63.1% to 67.1% for agents using Drex versus standard retrieval. _Developers weighing agent retrieval efficiency gains against vendor claims can follow benchmark coverage on daily.dev._ ### Which coding agents can use Drex for Search as a tool? Drex is available through a CLI, Python and TypeScript SDKs, MCP, and a REST API, and coding agents such as Claude Code, Codex, and Hermes Agent can use it as a tool. On BrowseComp-Plus it reached 92.05% accuracy with GPT-5 using 12.61 search calls per question, 16.6 times fewer calls than the top-ranked entry. _Developers wiring retrieval tools into Claude Code or Codex can keep up with MCP-compatible options on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 2 discussions and 47 comments across x (as of 2026-10-26). **TL;DR:** Replies generally like the folder-to-file-to-page narrowing and the citation-gated stopping rule, framing the token/step savings as the real win, but many want to see how it holds up on scanned PDFs, messy drives, and answers spread across multiple files. **Sentiment:** 45% positive · 50% mixed · 5% skeptical **The case for** - The folder-then-file-then-page approach mirrors how people actually search their own drives. - Citation-gated stopping is seen as a smart fix for agents that otherwise don't know when to stop reading. - Returning the exact page and quote lets a human verify an answer quickly before trusting it. - The input-token reduction is viewed as the more practically important metric than the accuracy bump. **The pushback** - Doubt about whether retrieval holds up on scanned PDFs and other messy, low-quality documents. - Concern that a stop rule could fire too early on questions whose answer spans multiple files, hurting recall. - Worry that picking the wrong folder or misleading folder names early on could derail the whole process. - Skepticism that a short, trusted-looking cited snippet could be more easily steered by a planted or ambiguous instruction than a longer document. **By community** - x (mixed): Interest in the folder-to-page narrowing and citation-based stopping is tempered by repeated requests to see it tested on scanned PDFs, messy drives, and multi-file answers. **Hottest debate:** Whether the citation-gated stop rule is a genuine fix for premature termination or itself a risk of stopping too early when evidence spans multiple files. **Open questions** - How does it handle scanned PDFs or messy scans where page text is unreliable? - Does it return both pages when an answer is split across two files? - Does the reported accuracy and step reduction hold on messy codebases or shared drives, not just clean document benchmarks? - Does it return raw page text or a summarized quote that an agent could still overstate? - How does it decide there is enough evidence to stop, especially when sources disagree or gaps remain? **Highlights** > @testingcatalog folder, then file, then page is how a person finds a receipt. the weak spot is step one: pick the wrong folder and every step after it is a confident trip down the wrong hallway. > — [mimies\_tech on x](https://x.com/mimies_tech/status/2114742042568093913) > @testingcatalog Page-level retrieval is a nice way to keep context small. What I'd test is a planted instruction on the exact page it hands back, since a short, trusted-looking snippet can steer an agent more easily than a long messy file. > — [sgegaed on x](https://x.com/sgegaed/status/2114746141988515876) > @omarsar0 Stopping when it can cite the source is the interesting design choice. The risk is a stop rule that fires on the first page that looks citable, so I'd want to see how often it stops early on questions whose answer spans two files. Fewer steps only matters if recall on those holds > — [renandnzsantos on x](https://x.com/renandnzsantos/status/2114762171217527234) > @omarsar0 deciding the page before any read cuts tokens, but the +4 point gain is measured on the same harnesses that already fail when the tool result wording is ambiguous. does drex return the raw page text or a summarized quote the agent can still overstate? > — [VibeCoderOfek on x](https://x.com/VibeCoderOfek/status/2114772300025204988) > @omarsar0 Stopping when it can cite the source is the right stop condition for regulated work. In broking we don't accept a fact from an agent unless it points to the email or document it came from. > — [dmitry\_bindary on x](https://x.com/dmitry_bindary/status/2114785068250259779) **Source threads** - [x](https://x.com/testingcatalog/status/2114736537904595412) · 0 points · 27 comments - [x](https://x.com/omarsar0/status/2114754945916526855) · 0 points · 20 comments ## Similar posts on daily.dev - [Codex is now the recommended agent in JetBrains IDEs](https://daily.dev/posts/codex-is-now-the-recommended-agent-in-jetbrains-ides-q3gxvk24g) · JetBrains · 1 upvotes · 0 comments - [Context Engineering for AI Agents Starts with Retrieval](https://daily.dev/posts/context-engineering-for-ai-agents-starts-with-retrieval-or6sfdhgx) · LangChain · 0 upvotes · 0 comments - [Agentic AI doesn’t have a model problem. It has a retrieval problem.](https://daily.dev/posts/agentic-ai-doesn-t-have-a-model-problem-it-has-a-retrieval-problem--w44jceia1) · The New Stack · 1 upvotes · 0 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#mcp](https://daily.dev/tags/mcp), [#claude-code](https://daily.dev/tags/claude-code), [#rag](https://daily.dev/tags/rag) [View this post on daily.dev](https://daily.dev/posts/nace-ai-s-drex-for-search-page-level-retrieval-for-agents-working-over-file-collections-hqcrdgatq) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Nace AI's Drex for Search: page-level retrieval for agents working over file collections","url":"https://daily.dev/posts/nace-ai-s-drex-for-search-page-level-retrieval-for-agents-working-over-file-collections-hqcrdgatq","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/nace-ai-s-drex-for-search-page-level-retrieval-for-agents-working-over-file-collections-hqcrdgatq"},"datePublished":"2026-10-26T16:52:20.022Z","dateModified":"2026-10-26T19:53:19.846Z","description":"Nace AI released Drex for Search, a retrieval layer that helps AI agents find cited evidence inside large file collections. Instead of returning top-k chunks...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/5fbb9f4590b77d634a465e44f7520987?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/5fbb9f4590b77d634a465e44f7520987?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/nace-ai-s-drex-for-search-page-level-retrieval-for-agents-working-over-file-collections-hqcrdgatq","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,mcp,claude-code,rag","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Nace AI's Drex for Search: page-level retrieval for agents working over file collections"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/nace-ai-s-drex-for-search-page-level-retrieval-for-agents-working-over-file-collections-hqcrdgatq#faq","mainEntity":[{"@type":"Question","name":"What is Drex for Search and how does it differ from standard chunk retrieval?","acceptedAnswer":{"@type":"Answer","text":"Drex for Search is a retrieval layer from Nace AI that narrows searches step by step from folder to file to page, rather than returning top-k similar chunks. It stops once it can cite a source, returning the answer, a supporting quote, and the page number. It works across AWS S3, Azure, and DigitalOcean, handling scanned PDFs, spreadsheets, email, and audio. Teams evaluating retrieval approaches for agent workflows can track tool comparisons like this on daily.dev."}},{"@type":"Question","name":"How much more efficient is Drex compared to standard agent retrieval methods?","acceptedAnswer":{"@type":"Answer","text":"Drex reduces agent steps by 62% and input tokens by 56% per question, according to Nace AI's reported figures. Across 529 queries from six professional work benchmarks including Databricks' OfficeQA, accuracy rose from 63.1% to 67.1% for agents using Drex versus standard retrieval. Developers weighing agent retrieval efficiency gains against vendor claims can follow benchmark coverage on daily.dev."}},{"@type":"Question","name":"Which coding agents can use Drex for Search as a tool?","acceptedAnswer":{"@type":"Answer","text":"Drex is available through a CLI, Python and TypeScript SDKs, MCP, and a REST API, and coding agents such as Claude Code, Codex, and Hermes Agent can use it as a tool. On BrowseComp-Plus it reached 92.05% accuracy with GPT-5 using 12.61 search calls per question, 16.6 times fewer calls than the top-ranked entry. Developers wiring retrieval tools into Claude Code or Codex can keep up with MCP-compatible options on daily.dev."}}]} ```