Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction
Giving AI agents grep and shell access outperforms vector databases on 5 benchmarks.
A large team of researchers from the University of Illinois at Urbana-Champaign, University of Washington, and other institutions has introduced Direct Corpus Interaction (DCI), a retrieval paradigm that replaces traditional semantic similarity interfaces with direct terminal-based tools. In their paper, the authors argue that conventional retrieval systems (BM25, dense retrievers, rerankers) compress corpus access into a single top-k step before reasoning, creating an irreversible bottleneck for agentic tasks that require multi-step hypothesis refinement, exact lexical constraints, and intermediate entity discovery. DCI equips language agents with grep, file reads, shell commands, and lightweight scripts to interact with the raw corpus directly—no vector index, embedding model, or retrieval API.
On benchmarks including BRIGHT (5 datasets), BEIR (6 datasets), BrowseComp-Plus, and multi-hop QA, DCI consistently outperforms strong baselines like ColBERT-v2, BGE-M3, and various re-rankers. Notably, DCI achieves up to 12% absolute gains on BRIGHT-datasets and matches or exceeds top methods on BEIR without any traditional retrieval component. The approach adapts naturally to evolving corpora and requires no offline indexing, making it ideal for personal knowledge bases and real-time data. The authors frame DCI as opening a broader interface-design space for agentic search, where retrieval quality depends not just on reasoning but on the resolution of the interaction interface.
- DCI uses grep, shell commands, and file reads instead of embedding models or retrieval APIs, eliminating the need for offline indexing.
- Outperforms ColBERT-v2 and BGE-M3 by up to 12% on BRIGHT benchmarks and matches top methods on BEIR without any semantic retriever.
- Achieves strong results on BrowseComp-Plus and multi-hop QA, highlighting that interface design matters as much as reasoning for agentic search.
Why It Matters
Rethinking retrieval from top-k similarity to direct terminal access may unlock simpler, more powerful agentic search systems.