Dr-DCI: Scaling Direct Corpus Interaction via Dynamic Workspace Expansion

The article introduces DR-DCI, a framework that enhances Direct Corpus Interaction (DCI) by enabling dynamic workspace expansion for efficient document retrieval and interaction. DR-DCI employs a retriever-steered approach, allowing agents to pull relevant documents into a local workspace for flexible operations, resulting in improved accuracy of 73.3% on Browsecomp-Plus and effective scaling from 100K to 10M documents. This development is significant for practitioners as it combines scalable retrieval with precise local operations, addressing performance issues faced by traditional DCI methods and enhancing the efficiency of evidence resolution in large corpora.