Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction
Overview
- Authors: Zhuofeng Li, Haoxiang Zhang, Cong Wei, Pan Lu, Ping Nie, Yi Lu, et al. (19 authors total)
- Published: 2026-05-03
- Source: https://arxiv.org/abs/2605.05242
- Category: Agentic Search
- Conventional retrievers (lexical or semantic) expose a corpus through a fixed similarity interface that compresses access into a single top-k step before reasoning.
- This abstraction is a bottleneck for agentic search: exact lexical constraints, sparse clue conjunctions, local context checks, and multi-step hypothesis refinement are hard to express via off-the-shelf retrievers, and early-filtered evidence cannot be recovered by downstream reasoning.
- The paper proposes direct corpus interaction (DCI): an agent searches the raw corpus directly using general-purpose terminal tools (grep, file reads, shell commands, lightweight scripts) — no embedding model, vector index, or retrieval API.
- DCI requires no offline indexing and adapts naturally to evolving local corpora.
- The setup substantially outperforms strong sparse, dense, and reranking baselines on several BRIGHT and BEIR datasets, and attains strong accuracy on BrowseComp-Plus and multi-hop QA.
- Conclusion: retrieval quality depends not only on agent reasoning ability but also on the resolution of the interface between the model and the corpus, opening a broader interface-design space for agentic search.
- Evaluated on IR benchmarks (BRIGHT, BEIR) and end-to-end agentic search tasks (BrowseComp-Plus, multi-hop QA).
- Substantially outperforms sparse (e.g., BM25-style), dense, and reranking baselines on several datasets.
- Achieves these gains without any conventional semantic retriever.
- Retrieval interfaces should be treated as a design space, not a fixed primitive: higher-resolution interaction (direct corpus access) can beat pre-computed similarity indexes when the agent is capable enough.
- Relevant for engineers building LLM agents over local or evolving corpora where indexing is costly or impractical.
- Quantitative results should be verified against the original PDF tables.
- Direct corpus interaction may trade off latency and token cost against accuracy; industrial deployment must weigh these constraints.
- A Systematic Framework for Enterprise Knowledge Retrieval (arXiv 2512.05411)
- Synergizing RAG and Reasoning: A Systematic Review (arXiv 2504.15909)
- AceSearcher: Bootstrapping Reasoning and Search for LLMs via RL (arXiv 2509.24193)
- Agentic Information Retrieval (arXiv 2410.09713)
- Original paper: Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction — https://arxiv.org/abs/2605.05242
Key Points
Motivation
Modern retrieval pipelines separate retrieval, ranking, and generation. For agentic tasks — which require discovering intermediate entities, combining weak clues, and revising plans after observing partial evidence — this single-step retrieval abstraction loses information before reasoning ever begins.
Method
DCI replaces the retriever with direct corpus access via terminal tools:
1. No embedding model, no vector index, no retrieval API — the agent greps, reads files, and runs lightweight scripts against the raw corpus. 2. No offline indexing — the corpus can change freely without re-indexing. 3. The agent orchestrates multi-step exploration: forming hypotheses, verifying lexical constraints locally, and revising its plan based on partial evidence.