Getting Answers After 12.6 Searches: Nace AI’s Document Search for AI Agents Hits 92.05%

Nace AI recently released Drex for Search, specifically built to solve longstanding pain points when AI agents locate documents.

The company shared two key metrics: On the BrowseComp-Plus benchmark, Drex for Search achieved a 92.05% score with an average of 12.6 search calls. On the OfficeQA Pro benchmark for office document Q&A, agents also completed tasks in fewer interaction rounds.

Why these numbers matter starts with understanding how current AI agents operate. When an AI agent is assigned a task like “find the breach of contract clauses in this document and compare them with last year’s version,” it typically has to decide independently what to search for, which results to open, and when to stop. Many agents are inconsistent here—either they keep switching between synonyms for searches, or they give a hasty answer after scanning just a few pages. More search calls drive up latency and token costs; more rounds signal the agent is getting lost between documents.

Drex for Search inserts Nace’s in-house Drex decision model into this pipeline, letting it judge what to retrieve at each step, whether the gathered information is sufficient, and whether to continue. The company hasn’t shared extra technical details, but from evaluation results, it mainly optimizes the amount of valid information delivered per search call.

Benchmarks like BrowseComp-Plus test an agent’s ability to locate answers across open web pages and documents. A 92.05% score is already quite high for accuracy, and the 12.6 calls serve as a clear cost benchmark. The OfficeQA Pro results are more in line with real office scenarios: fewer interaction rounds usually mean shorter execution paths and more consistent outputs.

For developers, such components can be plugged into existing agent workflows, cutting down the work of fine-tuning search strategies from scratch. Regular users don’t need to remember this name right now, but the next time you use an AI assistant to look up internal docs, review a contract, or find an email—if it no longer leaves you waiting forever with off-topic responses—it’s almost certainly a decision model like this powering it.

Source: MarkTechPost, Author: Asif Razzaq