Beyond Semantic Similarity: Rethinking Retrieval for Agentic Search via Direct Corpus Interaction
- Authors: Zhuofeng Li, Haoxiang Zhang, Cong Wei, Pan Lu, Ping Nie, Yi Lu, et al. (19 authors)
- Source: https://arxiv.org/abs/2605.05242
- Category: Agentic Search
- Exact lexical constraints and sparse clue conjunctions are hard to express via an off-the-shelf retriever.
- Local context checks and multi-step hypothesis refinement require interleaved retrieval and reasoning.
- Evidence filtered out early cannot be recovered by stronger downstream reasoning.
- Agentic tasks exacerbate this: agents must discover intermediate entities, combine weak clues, and revise plans after observing partial evidence.
- No embedding model, no vector index, no retrieval API
- No offline indexing required — naturally adapts to evolving local corpora
- The interface resolution (how fine-grained the model's access to the corpus is) becomes the key design variable, not just the agent's reasoning ability
- Substantially outperforms strong sparse, dense, and reranking baselines on several BRIGHT and BEIR datasets
- Attains strong accuracy on BrowseComp-Plus and multi-hop QA without relying on any conventional semantic retriever
- Agentic Information Retrieval (arXiv:2410.09713)
- AceSearcher: Bootstrapping Reasoning and Search for LLMs via RL (arXiv:2509.24193)
- Synergizing RAG and Reasoning: A Systematic Review (arXiv:2504.15909)
TL;DR
Instead of accessing a corpus through a fixed top-k similarity interface, this work lets language agents interact with the raw corpus directly using general-purpose terminal tools (grep, file reads, shell commands, lightweight scripts) — no embedding model, vector index, or retrieval API. This Direct Corpus Interaction (DCI) approach outperforms strong sparse, dense, and reranking baselines on several BRIGHT and BEIR datasets and achieves strong accuracy on BrowseComp-Plus and multi-hop QA.
Background and Motivation
Modern retrieval systems — whether lexical or semantic — compress corpus access into a single top-k retrieval step before reasoning. This abstraction is efficient, but for agentic search it becomes a bottleneck:
The DCI Approach
The proposed Direct Corpus Interaction (DCI) paradigm lets an agent search the raw corpus directly with general-purpose terminal tools:
Results
Key Insight
As language agents become stronger, retrieval quality depends not only on reasoning ability but also on the resolution of the interface through which the model interacts with the corpus. DCI opens a broader interface-design space for agentic search.
Implications for Search / Rec / Personalization
1. Architecture: Cascaded retrieve-rerank-generate remains mainstream, but agentic paradigms are making retrieval strategy itself a learnable, iterative decision. 2. Evaluation: Static nDCG-style metrics give way to task success rate, citation accuracy, and multi-hop reasoning chain completeness. 3. Engineering trade-offs: DCI removes indexing costs but shifts the cost to test-time compute — latency, token budget, and safety (e.g., corpus poisoning, adversarial tool use) need explicit consideration. 4. Limitations: Experimental scale, benchmark-vs-real-user distribution gaps, and cross-lingual generalization remain open questions.
Original Abstract
> Modern retrieval systems, whether lexical or semantic, expose a corpus through a fixed similarity interface that compresses access into a single top-k retrieval step before reasoning. This abstraction is efficient, but for agentic search, it becomes a bottleneck: exact lexical constraints, sparse clue conjunctions, local context checks, and multi-step hypothesis refinement are difficult to implement by calling a conventional off-the-shelf retriever, and evidence filtered out early cannot be recovered by stronger downstream reasoning. [...] we study direct corpus interaction (DCI), where an agent searches the raw corpus directly with general-purpose terminal tools (e.g., grep, file reads, shell commands, lightweight scripts), without any embedding model, vector index, or retrieval API. [...] Our results indicate that as language agents become stronger, retrieval quality depends not only on reasoning ability but also on the resolution of the interface through which the model interacts with the corpus.