DR-DCI Scales Agentic Search Over Large Corpora with Dynamic Workspace.

Agentic search systems often struggle with scalability when interacting directly with vast document corpora, relying on retrievers that provide limited views of information. This research introduces DR-DCI, a novel framework that improves upon Direct Corpus Interaction (DCI) by integrating retriever-steered dynamic workspace expansion. Instead of agents operating on the entire corpus, DR-DCI enables them to selectively retrieve and pull relevant documents into a localized workspace, where DCI operations can then be performed efficiently. This design effectively combines the broad recall capabilities of retrievers with the precise, flexible interaction offered by DCI. The agent can dynamically expand its workspace as needed, allowing for complex operations like filtering, comparison, and constraint verification across documents without being bogged down by full-corpus commands. This hybrid approach ensures both scalability and granular control over evidence resolution. Experimental results demonstrate DR-DCI's superior effectiveness and efficiency across various scales, from 100K to 10M documents, and even a 20M-scale Wiki-18 QA setting. It significantly improves accuracy and reduces tool usage, wall time, and estimated costs compared to raw DCI and other retrieval-based baselines, proving its robustness for large-scale agentic search tasks.