RARG Turns Relevance Into an Execution Prior for ripgrep, Not Just a top-k Filter, and Tops HuggingFace Papers at 71 Upvotes

arXiv / HuggingFace Daily Papers·medium signalXBlueskyLinkedInCopy link

A five-author Tencent paper (arXiv 2607.24223, July 27) targets a specific weakness in agentic search: Direct Corpus Interaction gives agents grep-style exploration power but its relevance-agnostic traversal surfaces useful clues late and delays convergence, while classic retrieval uses relevance only to pick top-k documents it cannot localize or verify within. The Relevance-Aware RipGrep Search Agent orders documents for sequential ripgrep traversal so globally relevant clues appear earlier, seeds promising entry points with query-relevant paragraphs, and reranks individual grep matches so informative excerpts beat merely matching lines. It improves the accuracy-efficiency frontier over both retrieval-based and direct-interaction agents on browse QA and reasoning-intensive retrieval, which is directly applicable to any coding or research agent whose default search tool is grep.