Research · Nov 1, 2026
Drex searches. Agents reason.
Nace.AI
TL;DR
An agent searching a filesystem can spend much of its effort deciding where to look. It lists files, searches for a phrase, reads a match, then discovers the wrong document and searches again. Each step requires another navigation decision from the reasoning model. Each tool result adds to its context.
We’re releasing Drex for Search. Drex is Nace’s System 1 decision model, built to make fast, accurate decisions at high frequency. It is among the top performers on the Artificial Analysis Decision Index.
With Drex for Search, finding a candidate and judging its relevance are the same step: a single Drex call does both. We organize the workspace as a tree and use recursive Drex calls to judge which branches are relevant and locate evidence within them. This replaces the agent’s loop of ls, grep, retrieval calls, and exploratory reads with precise source references, so the agent can read the relevant material directly.
Drex for Search cut the main agent’s model calls by about 53% on OfficeQA Pro and 51% on CUAD, and its input tokens by about 57% and 44%. Answer accuracy was maintained: the mean was slightly higher on both datasets. Most of the saved iterations were navigation: directory listings, searches that found nothing, and opening the wrong document. That navigation also accounted for about 80% of the input tokens saved.
How Drex for Search navigates a workspace
Drex for Search operates on a Nace Document Intelligence (NDI) workspace. Files are ingested before search to make their internal structure available for navigation. A PDF exposes sections and source pages; an Excel workbook exposes named sheets, readable content, and queryable tables. Every location remains linked to its source.
Income statement
Revenue, expenses, and earnings for the reporting period.
Revenue by month
Reported revenue organized by monthly period.
This structure gives Drex a map to navigate recursively, narrowing the search from files to the evidence the agent needs.
A cascade of small decisions
Consider a workspace containing annual reports, contracts, and a ledger. The example below follows the search for an answer to “What was total revenue for fiscal 2023?”.
Scroll across to explore the full workspace
At each level, Drex judges which branches are relevant and where to look next. It repeats this process recursively, focusing detailed inspection on promising branches until it locates page, line, or table references the agent can use. Independent decisions can run concurrently, so the main agent does not need a separate reasoning turn for every candidate.
The agent receives focused evidence and precise source references. It can read the relevant material, reason over it, and produce an answer grounded in the source.
Evaluating search on real documents
We measured main-agent workload and answer quality on OfficeQA Pro (financial reports) and CUAD (commercial contracts). We chose them because both are public benchmarks with published reference answers, built on two different kinds of real business documents: decades of scanned Treasury Bulletins dense with tables, and commercial contracts written in legal language.
For each combination of harness, model and reasoning effort, we ran the same questions three ways:
- Raw harness: the agent used its built-in file tools, such as
ls,grepand file reads. It worked on the text each dataset publishes for its PDFs, not on the PDFs themselves: OfficeQA’s official parsed text and CUAD’s full-contract text files. That text was already in place when each question started. - Raw harness + hybrid search: the agent retained those file tools and also had access to pi-knowledge-search, a third-party search extension, over the same official parsed texts. We chose pi-knowledge-search because it adds a standard hybrid retriever to an agent without changing anything else, and the same unchanged code runs in all six harnesses. We used version 1.3.8, with
text-embedding-3-smallembeddings at 512 dimensions, BM25 keyword search, and reciprocal rank fusion. The agent in each harness had access to the same extension. The search index was built once per dataset, before evaluation, and reused for every question. - Drex for Search: the PDFs were ingested into a Nace Document Intelligence workspace once per dataset, before evaluation, and that workspace was reused for every question. Drex supplied evidence locations for the agent to read or query.
OfficeQA Pro
OfficeQA Pro asks questions over 697 U.S. Treasury Bulletin documents spanning 1939–2025. For the Drex for Search arm, the scanned PDFs were ingested into the Nace Document Intelligence workspace before evaluation. Across six harnesses and ten configurations, Drex for Search used 53% fewer main-agent iterations and 57% fewer input tokens than the raw harnesses, with fewer of both in every configuration. Against hybrid search, it used 47% fewer iterations and 50% fewer input tokens, again in every configuration, with slightly higher mean accuracy.
CUAD
CUAD, the Contract Understanding Atticus Dataset, is a public collection of commercial contracts annotated for contract review. Our evaluation asks questions about contract clauses across 200 contracts, using eight harness/model configurations. Accuracy excludes failed runs.
Drex for Search used 51% fewer main-agent iterations and 44% fewer input tokens than the raw harnesses, with fewer iterations in every configuration. Against hybrid search, it used 50% fewer iterations and 44% fewer input tokens, with slightly higher mean accuracy.
Less exploration in the agent loop.
Each configuration has a grey raw baseline connected to its hybrid-search and Drex for Search results. Points toward the upper left indicate higher answer accuracy with fewer agent iterations.
Hover or tap to highlight a configuration and compare all three search methods.
Scroll across to compare workload and answer quality
An iteration is one request to the main agent’s model, and input tokens are everything sent in those requests, cached context included. Reductions compare the per-configuration means, summed over each dataset.
Cost per question
Fewer main-agent iterations and input tokens also lowered the cost of answering a question, even with the cost of the Drex calls included:
- OfficeQA Pro: $0.742 per question with the raw harness; $0.623 with hybrid search (−16%); $0.422 with Drex for Search (−43%, and −32% against hybrid search).
- CUAD: $0.046 per question with the raw harness; $0.048 with hybrid search (+3%); $0.037 with Drex for Search (−21%, and −23% against hybrid search).
Costs are averaged over each dataset’s configurations and use the model providers’ list prices, with cached input billed at the cache rate. Drex calls are billed at official Drex Max pricing. A question took about 580 Drex calls and 1.35M Drex input tokens on OfficeQA Pro, and about 30 Drex calls and 145k Drex input tokens on CUAD.
What Drex for Search saves in the agent loop
With grep or a retriever, the agent spends one iteration writing a search and another reading the results and deciding what to open next. If a match comes from the wrong file, the cycle repeats. A better retriever returns better candidates, but the agent still judges them, and every result stays in its context.
Drex for Search does retrieval and judgment in the same Drex call. Drex uses the question and the workspace’s structure to narrow from files to the pages that hold the answer, so the agent spends its iterations reading evidence and answering.
To see how Drex for Search saves agent iterations, we classified every tool call in the OfficeQA Pro runs.
7 of the 10 iterations Drex for Search saved were navigation.
Main-agent iterations per question: raw harness 18.7; hybrid search 16.4, 2.2 fewer, about one of them navigation; Drex for Search 8.8, 9.9 fewer. Of the 10 iterations saved, 7 were navigation: 3 opening the wrong file, 2 finding the file, 1 searching a long file for the table, 1 looking for the document on the web; 3 were other work. Navigation accounts for 491k of the 604k input tokens saved per question.
Where Drex for Search’s 10 saved iterations went
Input tokens saved per question
604kEvery listing, search result and wrong file stays in the agent’s context and is sent again with each later request. That is why navigation is 70% of the saved iterations but about 80% of the saved tokens.
Hybrid search removed about one navigation iteration per question; Drex for Search removed seven. Hybrid search rarely ranked the right file first, and most of its results came from other files, so agents still listed, searched and opened files to find the right one. Drex judges which file and page the question needs, and agents reached the right file in a median of one iteration, against three with hybrid search or the raw harness.
We believe that combining retrieval and judgment within Drex for Search, and moving navigation out of the main agent’s loop, helps explain the efficiency improvements we observed on OfficeQA Pro and CUAD.
Drex searches. Agents reason.
In a typical search loop, the agent reads the candidates returned by grep or a retriever and judges what to pursue. With Drex for Search, finding a candidate and judging its relevance are the same step: a single Drex call does both. Fast, inexpensive Drex calls repeat this process through the workspace and return precise references for the agent to read.
This suggests a promising division of work for file search and filesystem agents: decision models locate the evidence; reasoning models turn it into answers. Drex searches. Agents reason.
