AIAny
Icon for item

A New Role for Relevance: Guiding Corpus Interaction in Agentic Search

Turns document relevance into an execution prior for agentic corpus interaction: orders documents for sequential ripgrep traversal, seeds promising entry points with query-relevant paragraphs, and reranks grep matches to surface informative excerpts. Improves the accuracy–efficiency frontier on browse QA and reasoning-intensive retrieval.

Introduction

Why this matters

Many retrieval agents treat relevance as a static selection signal (top‑k documents) and then hand everything to an LLM, which struggles to localize, compose, or verify dispersed evidence. Direct Corpus Interaction (DCI) enables fine‑grained grep‑style exploration but can wander through low‑value matches and converge slowly. The core insight of this paper is simple but consequential: make relevance an execution prior that actively guides how and where the agent searches the corpus, so globally relevant clues surface earlier and LLM interaction converges faster.

Key Findings
  • Relevance-as-prior: Relevance scores are used beyond document ranking — they determine traversal order for sequential 'ripgrep' scans, initialize high‑value entry paragraphs, and rerank grep matches so that informative excerpts reach the LLM earlier. This shifts relevance from a filter to an execution policy.

  • Coarse-to-fine guidance: The proposed agent (RARG) applies relevance at multiple granularities (document ordering → paragraph seeding → excerpt reranking), which reduces wasted search and accelerates evidence discovery during interactive corpus exploration.

  • Empirical payoff: Across browse question answering and reasoning‑heavy retrieval tasks, RARG improves the accuracy–efficiency tradeoff relative to both retrieval‑first pipelines and relevance‑agnostic DCI agents, yielding faster and more reliable convergence to correct answers.

Who it’s for — and tradeoffs

Great fit if you build or evaluate agentic search systems that need robust evidence localization (e.g., browse QA, multi‑hop reasoning, or tool-using LLM agents). RARG is especially useful when document relevance signals are available and where early exposure of globally relevant clues shortens costly LLM interactions. Look elsewhere if you cannot produce reasonably calibrated document‑level relevance scores, if corpus access latency makes sequential traversal impractical, or if absolute end-to-end throughput (many concurrent queries) is prioritized over per‑query search convergence time.

Where it fits

RARG sits between retrieval‑centric RAG pipelines and pure DCI agents: it retains interactive, grep‑style capabilities for fine‑grained evidence assembly while borrowing relevance structure to guide that interaction more effectively. Methodologically, it is most relevant to systems that combine retrieval, excerpt-level reasoning, and iterative LLM prompting.

Implementation notes (conceptual)

The method requires (1) a document relevance estimator to seed priors, (2) a mechanism to order documents for sequential ripgrep traversal, and (3) a reranking step that scores candidate excerpts by informativeness before exposing them to the LLM. These steps introduce extra orchestration and reranking cost, so benefits should be weighed against added computation and pipeline complexity.

Information

  • Websitearxiv.org
  • AuthorsJiangnan Li, Yuqing Li, Mo Yu, Jinchao Zhang, Jie Zhou
  • Published date2026/07/27

More Items

Measures how agent memory systems miss implicitly associated facts by introducing InMind, a 125-task benchmark with paired controls that separate stored-vs-retrieval vs knowledge gaps. Quantifies a large retrieval-interface blind spot and points to routing as the core open problem.

Drives long‑horizon desktop agents by reading and manipulating program state (files, DOM, backends) instead of relying on screenshots. The main agent uses code for actions and structural verification while a lightweight GUI subagent handles rare screenshot-click steps, improving success rates and lowering per-task cost versus screenshot-only approaches.

Bridges the proprietary-to-open-source gap in agentic search by converting multi-step retrieval and reasoning traces into a structured, style-normalized JSON protocol and using it for joint distillation + RL. Produces denser supervision that improves student success rates while reducing style drift.