Editorial illustration for DR-DCI Enables Agent-Callable Retrieval to Expand Local Workspace Efficiently
DR-DCI Enables Agent-Callable Retrieval to Expand Local...
Most AI research papers sell you a new way to fail. They’ll claim they’ve cracked some fundamental tension, then quietly fudge the data. This one’s different.
DR-DCI fixes a real problem. Direct corpus interaction, or DCI, asks a model to pick through every document to answer a question. It’s precise but agonizingly slow.
Standard retrieval pulls relevant docs first, which is fast but sloppy. You either drown in noise or miss the point entirely.
This system splits the difference by letting the AI decide when to search. The agent works within a small, local set of documents. When it hits a wall, it calls a retriever to pull in more. It’s like a researcher who knows when to ask for another box of files.
The results are hard numbers, not hype. On the Browsecomp-Plus benchmark, DR-DCI hit 71.2% accuracy. That’s up to 8.3 points better than raw DCI.
It also used fewer tools and less time. A tweak called workspace-preserving context reset pushed accuracy to 73.3%.
Scale is where other methods crumble. DR-DCI held steady from 100,000 to 10 million documents. Raw DCI got unstable.
Simple BM25 retrieval fell apart. At 20 million documents across six Wiki-18 QA tests, DR-DCI averaged a score of 63.0, beating both retrieval-based systems and trained search agents.
We introduce DR-DCI, a retriever-steered DCI framework that treats retrieval as an agent-callable action for expanding a local workspace. Rather than operating directly over the full corpus, the agent dynamically pulls relevant documents into an evolving workspace and conducts DCI operations within it. This design combines retriever-level recall with DCI-style precision: retrieval keeps exploration scalable, while DCI preserves the local operations needed for effective evidence resolution.
Experiments show that DR-DCI is both effective and efficient across scales. 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\%.
In corpus-scaling experiments, DR-DCI remains effective from 100K to 10M documents, whereas raw DCI becomes unstable and BM25 performs substantially worse. DR-DCI also scales to a 20M-scale file-per-document Wiki-18 QA setting, achieving an average score of 63.0 across six benchmarks and outperforming retrieval-based and trained search-agent baselines.
The point isn’t just a higher score. It’s that the trade-off was always a false choice. You don’t need to sacrifice precision for scale, or the other way around.
The mechanism is elegant. Give the agent a small desk and a library card. Let it know when to fetch more.
The gains come from that simple, dynamic adjustment.
This feels like actual progress. Not a revolution, but a clear step past a stubborn obstacle. The best ideas often look simple once someone finally points them out.
Related Reading
Research & Benchmarks Google's FACTS benchmark shows 70% factuality ceiling across four tests
Research & Benchmarks Databricks finds multi-step agents beat single-turn RAG by 21% to 38% on STaRK
Research & Benchmarks Nvidia's DLSS 4.5 beta adds 6x Multi Frame Generation for RTX 50 GPUs
From Archives Fused kernels boost MoE training, forward and backward passes up to 1.3×
From Archives Hybrid Open-Ended Tri-Evolution Improves Deep Research for AI Agents
Common Questions Answered
What problem does DR-DCI solve between direct corpus interaction and standard retrieval methods?
DR-DCI addresses the fundamental trade-off between precision and speed in document retrieval. Direct corpus interaction (DCI) is precise but extremely slow because it requires a model to examine every document, while standard retrieval is fast but often misses relevant information or pulls in noise. DR-DCI splits the difference by enabling agents to dynamically fetch additional documents only when needed, combining both speed and accuracy.
How does the agent-callable retrieval mechanism work in DR-DCI?
The system gives the agent a small workspace and a library card, allowing it to know when to fetch more documents from the corpus. This dynamic adjustment means the agent can start with a limited set of documents and intelligently request additional relevant materials as needed to answer questions more accurately. The mechanism enables efficient expansion of the local workspace without requiring exhaustive searches through all available documents.
Why is DR-DCI considered actual progress rather than just another incremental improvement?
DR-DCI demonstrates that the traditional trade-off between precision and scale was a false choice, showing that you don't need to sacrifice one for the other. The solution is elegant and simple once understood, representing a clear step past a stubborn obstacle in information retrieval rather than claiming revolutionary breakthroughs or fudging data like many AI papers do.
What are the key advantages of using agent-callable retrieval over traditional document selection methods?
Agent-callable retrieval allows for dynamic, intelligent document fetching based on actual information needs rather than static pre-selection. This approach prevents both the drowning-in-noise problem of pulling too many documents and the missed-information problem of pulling too few, while maintaining computational efficiency by only expanding the workspace when necessary.
Further Reading
- AgentIR: Reasoning-Aware Retrieval for Deep Research Agents — arXiv
- The 2026 Shift: Moving Beyond Vector RAG to Agentic Retrieval ... — SesameDisk
- Papers with Code Benchmarks — Papers with Code
- Chatbot Arena Leaderboard — LMSYS
