Nace AI Releases Drex Search Model: 92.05% on BrowseComp-Plus

Nace AI recently released Drex for Search, a document search model for AI Agents, extending its Drex decision model family into the retrieval domain. The headline figures from the official announcement are: a score of 92.05% on the BrowseComp-Plus benchmark with an average of only 12.6 search calls, along with a reduction in Agent interaction turns on OfficeQA Pro. These metrics all point in the same direction—achieving higher answer hit rates with fewer retrieval actions.

What Drex for Search Does

Drex was originally Nace AI's family of decision models, and this release migrates decision-making capability to the document search stage. In traditional retrieval-augmented generation (RAG) pipelines, Agents are often locked into a linear "retrieve first, then read, then generate" chain, where the number of searches is determined by the pipeline rather than task difficulty. Drex for Search's approach is to let the model itself decide when to search, how many times to search, and whether it has gathered enough information to stop.

This "retrieval-as-decision" positioning is directly reflected in two metrics: the 92.05% accuracy indicates result quality, while the 12.6 average search calls indicate process efficiency. For Agent tasks requiring multi-hop reasoning, the number of search calls directly determines latency and cost, so the latter is not a secondary metric.

Retrieval-augmented generation (RAG) refers to having large language models retrieve relevant documents from external knowledge bases before generating answers, injecting the retrieval results as context into the prompt to reduce hallucinations and improve factual accuracy. Classic RAG uses a fixed "single retrieval—generation" pipeline, whereas Agent-oriented RAG often evolves into multi-round iteration: the model generates queries, retrieves, reads, judges whether information is sufficient, and retrieves again if not. It is precisely in this multi-round iteration that the number of retrievals can skyrocket with task complexity, creating dual pressure on latency and token costs. The "linear pipeline" problem that Drex for Search targets is essentially about delegating retrieval counts to a fixed process rather than model judgment, leading to over-retrieval on simple tasks and under-retrieval on complex ones.

Interpreting the Benchmark Results

BrowseComp-Plus is an evaluation set focused on browsing and retrieval capabilities, emphasizing locating and integrating answers from open web information. Nace reports a 92.05% score combined with 12.6 average search calls, meaning the model converges without extensive trial-and-error retrieval on most samples.

One caveat: these numbers are self-reported by the vendor, and the source material does not disclose evaluation configuration, baseline comparisons, or statistical methodology. A 92.05% figure lacks a reference point in isolation—only comparison with similar Agent search solutions under identical settings can determine its true standing. Likewise, 12.6 calls is a mean; the distribution shape (whether a small number of samples consume a large number of calls) heavily impacts actual deployment costs, and this remains unclear.

BrowseComp-Plus is a variant of the BrowseComp series, the latter proposed by OpenAI to evaluate models' ability to perform multi-hop, multi-source information integration on the open web, with questions typically requiring clues from multiple web pages to be combined for an answer. The Plus version generally strengthens the examination of the retrieval process, requiring models to actively initiate searches and filter results rather than relying solely on static context, so it measures both retrieval strategy and reasoning integration capability. OfficeQA Pro, meanwhile, targets enterprise office document scenarios, where documents are relatively well-structured but lengthy with many tables and hierarchy levels, testing models' ability to locate answers in mixed structured and unstructured text. These two represent typical Agent retrieval environments—open web and enterprise documents, respectively—which is also why the original article emphasizes that more scenario data is needed to establish generality.

Interaction Turn Optimization on OfficeQA Pro

OfficeQA Pro evaluates question-answering capability for office documents, where retrieval targets are typically highly structured enterprise documents. Nace claims that Drex for Search reduces Agent interaction turns on this benchmark.

The value of fewer turns lies in pipeline stability: each additional Agent round-trip can introduce error accumulation and context bloat, so reducing turns effectively compresses the error space. However, the source material does not provide the specific magnitude of turn reduction, nor does it state whether accuracy remains unchanged alongside the reduced turns—these two points determine whether the optimization is genuine efficiency improvement or a trade-off of accuracy for speed.

Implications for Agent Retrieval Architecture

From a product positioning perspective, Drex for Search belongs to the "retrieval strategy layer" rather than the "retrieval engine layer." It does not promise stronger recall but instead attempts to solve the long-standing problem of "when to retrieve and to what extent"—a question that has long relied on heuristic rules. In current mainstream Agent frameworks, retrieval turn limits and threshold judgments are mostly controlled by manually set parameters, so adaptive decision-making is indeed a valuable entry point.

Another implication is the shift in cost structure. If 12.6 calls is a normal value while maintaining 92.05% accuracy, then per-task retrieval overhead becomes predictable—more attractive to teams needing to deploy Agents at scale than absolute accuracy.

The distinction between the retrieval strategy layer and the retrieval engine layer is key to understanding this product's positioning. The retrieval engine layer focuses on recall rates, ranking quality, index construction, and embedding models, determining "what can be retrieved"; the retrieval strategy layer focuses on query construction, retrieval timing, stopping conditions, and result filtering, determining "how to search and for how long." In current mainstream Agent frameworks (such as ReAct and Plan-and-Execute paradigms), the strategy layer is mostly controlled by manually set maximum turns, similarity thresholds, or fixed prompts, lacking the ability to adapt to task difficulty. Modeling this layer theoretically allows Agents to stop early when information is sufficient and proactively rewrite queries when it is not, achieving a better balance between accuracy and call costs.

Open Questions

The primary question is reproducibility. Vendor self-reported benchmark scores require third-party re-testing, especially for evaluations like BrowseComp-Plus that emphasize open retrieval, where data contamination and retrieval source differences can significantly affect results. Second is generality: OfficeQA Pro and BrowseComp-Plus represent office documents and open web scenarios respectively, and there is no data yet on the model's performance in more complex environments such as enterprise private knowledge bases, low-quality documents, or long documents.

Third is integration cost. Drex for Search needs to be embedded into existing Agent workflows, and its compatibility with mainstream frameworks and whether it supports custom retrieval backends determine the actual migration cost for teams. None of this information is covered in the release materials and needs to be verified through subsequent technical documentation and practice.

Summary

Drex for Search brings decision models into the Agent retrieval stage, proposing an efficiency-oriented retrieval solution backed by a 92.05% BrowseComp-Plus score and 12.6 average search calls. The direction is sound, but the results still require third-party verification, and generality and integration details remain to be addressed. For teams currently optimizing Agent retrieval costs, this is an option worth tracking—not yet a conclusion that warrants immediately replacing existing solutions.

One-line Summary

Nace AI releases Drex for Search, achieving higher hit rates with fewer search calls through retrieval-as-decision.

Paragraph Summary

Nace AI has released Drex for Search, a document search model for AI Agents that extends its Drex decision model family into the retrieval stage. The company claims a 92.05% score on the BrowseComp-Plus benchmark with an average of only 12.6 search calls, and reduced Agent interaction turns on OfficeQA Pro. The core idea is to let the model decide when to retrieve, how many times, and when to stop, rather than relying on a fixed linear pipeline. The article affirms the soundness of this "retrieval-as-decision" direction, noting that it targets the retrieval strategy layer long controlled by manual thresholds and could enable a predictable cost structure; however, it also points out that the results are vendor self-reported, lacking baseline comparisons and evaluation configuration, with generality and integration costs still unverified.