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

🔥 Direct Corpus Interaction made agentic search simple: let agents search the corpus directly with shell tools.

🤔 But what happens when the corpus grows from 100K to 10M+ documents?

Introducing DR-DCI: Scaling Direct Corpus Interaction via Dynamic Workspace Expansion.

🚀 The Idea:
Instead of forcing agents to run terminal commands over the full corpus, DR-DCI uses retrieval as an agent-callable action to dynamically expand a local workspace.

The agent first pulls potentially relevant documents into its workspace, then performs local DCI operations — search, filter, compare, inspect, and verify — within that evolving workspace.

💡 Why it works:

DR-DCI combines the best of both worlds:

🔍 Retriever-level recall for scalable exploration
🛠️ DCI-style local operations for precise evidence verification
📁 Workspace-based reasoning for stable multi-document search

📊 Results:

On BrowseComp-Plus, DR-DCI reaches 71.2% accuracy, improving over raw DCI and ablated variants by up to 8.3 points, while reducing tool usage, wall time, and estimated cost.

With workspace-preserving context reset, accuracy further improves to 73.3%.

As the corpus scale, DR-DCI remains effective from 100K to 10M documents, while raw DCI becomes unstable and BM25 performs substantially worse.

DR-DCI also scales to a 20M-scale Wiki-18 file-per-document QA setting, achieving an average score of 63.0 across six benchmarks and outperforming retrieval-based and trained search-agent baselines.

💡 Key Insight:
For large-scale agentic search, retrieval should not replace interaction.
Instead, it should steer interaction.

DR-DCI shows that retrieval-guided workspace expansion can preserve the flexibility of Direct Corpus Interaction while making it scalable to tens of millions of documents.