Direct Corpus Interaction: A New Paradigm for Agentic Search
A recent paper on arXiv (2605.05242) contends that traditional retrieval systems, whether lexical or semantic, fall short for agentic search. These systems reduce access to the corpus to a single top-k step prior to reasoning, creating a bottleneck when agents must address precise lexical constraints, sparse clue combinations, local context evaluations, and multi-step hypothesis adjustments. Early-filtered evidence cannot be retrieved through enhanced downstream reasoning. The authors introduce direct corpus interaction (DCI), enabling agents to directly search the raw corpus with general-purpose tools like grep, file reads, and shell commands. DCI facilitates multi-step processes, such as identifying intermediate entities, merging weak clues, and adjusting plans based on partial evidence, aiming to address the shortcomings of standard retrievers in complex agentic tasks.
Key facts
- arXiv paper 2605.05242 critiques modern retrieval systems for agentic search.
- Current systems compress corpus access into a single top-k retrieval step.
- Fixed similarity interface is a bottleneck for exact lexical constraints and multi-step reasoning.
- Evidence filtered out early cannot be recovered by downstream reasoning.
- Proposes direct corpus interaction (DCI) using terminal tools like grep and file reads.
- DCI enables agents to discover intermediate entities, combine weak clues, and revise plans.
- The paper is categorized as 'cross' and is an abstract.
- The source URL is https://arxiv.org/abs/2605.05242.
Entities
Institutions
- arXiv