Editorial illustration for DR-DCI Enables Agent-Callable Retrieval to Expand Local Workspace Efficiently
DR-DCI Enables Agent-Callable Retrieval to Expand Local...
DR-DCI Enables Agent-Callable Retrieval to Expand Local Workspace Efficiently
Agentic search over massive text collections still leans on retriever‑mediated front ends—think BM25 or ColBERT—to pull in candidate passages. Those systems rank documents well, but they hand agents only a list of results or a clipped view of each file. Here's the thing: without full access, an agent can’t shuffle snippets around, cross‑check facts, or enforce multi‑document constraints.
Direct Corpus Interaction, or DCI, tries to fix that gap by exposing shell‑style commands that let a model search, filter, compare and verify across the entire corpus. While the idea sounds flexible, the reality is that running terminal‑level operations on a growing dataset quickly turns sluggish and unstable, eroding both speed and reliability. Recent ablation work points to two components that matter most: a ranked preview that gives a quick snapshot of relevance, and the ability to perform inter‑document DCI actions.
Those pieces appear to hold the key to keeping performance up as the workspace expands.
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.
Why this matters
DR-DCI redefines how agents interact with large text collections by turning retrieval into an explicit, callable action rather than a hidden ranking step. Instead of scanning an entire corpus at once, an agent now summons documents on demand, building a mutable workspace that can be reshaped as the task evolves. This design directly tackles the bottleneck of earlier retriever‑mediated pipelines, which offered only static ranked lists or limited document excerpts, hampering an agent’s capacity to cross‑reference and enforce constraints.
For developers, the promise is a more modular system where retrieval and reasoning are cleanly separated, potentially simplifying debugging and prompting strategies. Researchers may find a fertile testbed for studying how dynamic context windows affect chain‑of‑thought reasoning. Yet, several questions linger: can the workspace grow without prohibitive memory costs?
How robust is the approach when the corpus contains noisy or contradictory sources? The paper does not yet address scalability limits or failure modes, so the practical impact remains uncertain. We will watch how the community adapts this framework to real‑world pipelines.
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
Related Reading
Research & Benchmarks Hermes Agent tops use as Nous Research’s self‑improving model leads OpenRouter
Research & Benchmarks DeepMind spinoff’s AI‑designed drugs enter human trials after AlphaFold 3
Research & Benchmarks Google AI Advisors Let Users Probe Performance with Conversational “Why” Queries
From Archives Fused kernels boost MoE training, forward and backward passes up to 1.3×
From Archives