Direct Corpus Interaction (DCI) challenges traditional retrieval systems by enabling agents to query raw data using tools like grep and file reads, bypassing conventional indexing. This approach eliminates preprocessing bottlenecks, allowing agents to dynamically access metadata, text, and file structures. Early experiments show DCI outperforms dense and sparse retrieval models in multi-step search tasks.

The method excels in BrowseComp-Plus and multi-hop QA benchmarks, achieving strong accuracy without relying on semantic embeddings. By avoiding fixed retrieval APIs, DCI adapts to evolving data sources and handles complex queries requiring intermediate entity discovery. Researchers tested it on BRIGHT and BEIR datasets, demonstrating robustness across diverse information needs.

DCI’s significance lies in redefining how agents interact with corpora. Unlike black-box retrievers, it grants fine-grained control over data access, enabling agents to combine weak clues and revise strategies iteratively. This shift could accelerate progress in agentic AI by prioritizing interface flexibility over retrieval efficiency alone.

Key entities: Authors from UC Berkeley, University of Washington, and Allen Institute for AI. Primary application: Enhancing language agent capabilities in open-domain search. The work suggests future systems may prioritize direct data interaction over conventional semantic modeling.