A powerful language model's accuracy can plummet from 93% to just 15% when it relies on a weak retrieval system to find supporting evidence, according to new research published by The New Stack. The analysis examines how agentic AI systems—those that autonomously search for information and complete multi-step tasks—depend far more on the quality of their search infrastructure than on the sophistication of the underlying model. When retrieval fails to surface the right documents, even the most capable reasoning model can't compensate for evidence it never received.
The report uses benchmark data from BrowseComp-Plus, which tests agents against a fixed corpus of 100,195 documents and 830 research questions, each with verified supporting materials. In an "oracle" configuration where GPT-4.1 received all relevant documents directly, the model answered questions correctly 93.49% of the time. When an agent using the identical model had to rely on BM25—a traditional search method—to locate evidence before reasoning, accuracy dropped to 14.58%. A smaller model, Qwen3-32B, showed an even steeper decline: from 83.25% accuracy with oracle-supplied documents down to 3.49% when searching with BM25. The gap illustrates that retrieval quality, not just model intelligence, determines whether an agent succeeds or fails in production environments where documents may be duplicated, outdated, or described using inconsistent terminology.
The research highlights how agentic workflows amplify retrieval challenges through volume and compounding risk. A single user task can trigger hundreds or thousands of search operations within minutes, as agents decompose questions, fan out parallel searches, combine results, and issue follow-up queries based on what they've learned. When one task generates ten parallel searches and the agent must wait for all ten before proceeding, the slowest search sets the pace—making tail latency as important as average speed. The report offers an illustrative calculation: if each of five sequential retrieval steps has a 90% success rate and there's no recovery mechanism, the probability that all five succeed falls to roughly 59%. The authors note that real agents can retry and use redundant evidence, but the model demonstrates how small per-step weaknesses accumulate when later work depends on earlier retrievals. Traditional search engines evolved to handle short, underspecified queries from people who know little about the underlying corpus. Agents, by contrast, can construct filters, specify date ranges, choose ranking strategies, group results, and chain searches together—but only if the retrieval system exposes those controls and explains how to use them.
The report recommends that organizations inspect the searches their agents issue in production before upgrading models or retuning retrieval systems, since a capable model working from poor search results still has flawed material to reason over. It describes Vespa's "agent-ready schema" approach, which makes retrieval capabilities explicit by combining the search application schema with engine guidance and application-specific knowledge into a concise skill document the agent can consult when building queries. This lets the agent choose among retrieval strategies—lexical matching, semantic search, structured filters, ranking profiles—while keeping authorization and validation on the trusted server side. The bottom line: better retrieval can improve answer quality, cut unnecessary reasoning work, and make the entire agentic system more effective, but those gains require agents to understand both the controls available and when to use them. Enterprises investing in agentic systems may discover their competitive advantage lies less in the reasoning model they license and more in whether their search infrastructure can consistently surface the right internal knowledge at the right moment.