Beyond Semantic Similarity: Agentic Search via Direct Corpus Interaction

Beyond Semantic Similarity

We propose Direct Corpus Interaction (DCI) for agentic search, bypassing traditional embedding models and vector indexes. Instead of relying on fixed similarity interfaces, agents use general-purpose terminal tools like grep and shell commands to search raw corpora directly. This approach eliminates offline indexing, adapts to evolving data, and significantly outperforms strong sparse, dense, and reranking baselines on BRIGHT, BEIR, and BrowseComp-Plus benchmarks.

"言語エージェントが強力になるにつれ、検索の質は推論能力だけでなく、モデルがコーパスと対話するインターフェースの解像度にも依存します。"