Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction
Summary
The paper argues that fixed similarity interfaces in retrieval limit agentic search, and introduces Direct Corpus Interaction (DCI) which lets agents query the raw corpus with terminal tools without embeddings or indices. It reports substantial performance gains over strong baselines on multiple benchmarks and highlights the importance of the interaction interface for language agents. Implications include rethinking retrieval design for AI agents and evolving local corpora handling.