Can relevance guide grep search beyond just selecting documents?

The paper defines relevance as "a query-dependent estimate of whether a document or excerpt contains useful evidence" and argues that agentic search has used it in only one place: retrieval agents rank the corpus and expose a top-k set of documents or snippets to the language model. The introduction says this interface "conflates document relevance with evidence utility," and the abstract adds that document relevance alone "cannot localize, compose, or verify the evidence required by complex questions." The conclusion states the reframing directly: relevance "should guide corpus interaction, not merely select its inputs."

The starting point is Direct Corpus Interaction (DCI), where the agent explores with grep-style commands. DCI gives fine-grained, compositional operations, but its relevance-agnostic search "can expose useful clues late and delay convergence." The Relevance-Aware RipGrep Search Agent (RARG) keeps the grep interface and adds relevance back as "an execution prior" at coarse-to-fine resolutions. Document-level scores order the ripgrep traversal so promising documents are scanned first. Query-relevant paragraphs initialize entry points. Match-level reranking decides which excerpts stay visible "under a limited observation budget," surfacing informative excerpts "that document-level ranking may otherwise obscure." The introduction frames relevance as "a useful but imperfect prior" that indicates where evidence may be found, not whether it is sufficient, correctly localized or combined with other clues. That is why it is used to prioritize exploration and not to decide what counts as evidence.

Against the nearest notes, this is the counterweight to Can direct corpus search beat embedding-based retrieval?. That note presents DCI as an alternative to dense retrieval; this paper accepts DCI's fine-grained interaction and argues that dropping relevance entirely costs time-to-evidence. Read together, the choice is less retriever versus grep than where the relevance signal enters the loop. It also sits beside Can document count be learned instead of fixed in RAG?: both find the fixed top-k interface inadequate, but DynamicRAG repairs the selection step, while RARG moves relevance out of selection and into the ordering of interaction. The match-level reranking echoes Can verification separate structural near-misses from topical matches? in a loose way, since both keep a coarse relevance score and add a finer-grained stage after it.

The excerpt is thin on evidence. It reports that RARG "advances the accuracy–efficiency frontier" over retrieval-only agents, unrestricted DCI and retrieval-constructed interaction spaces, with higher accuracy, "substantially fewer tool calls" and graceful degradation as the corpus grows, across QA, corpus scaling and reasoning-intensive retrieval. It names no benchmarks, gives no numbers, identifies no retriever or reranker, and says nothing about the cost of computing relevance scores or which of the three mechanisms (traversal order, entry points, match reranking) carries the gain. What it supports at this strength is a design claim: a relevance score can be reused as a prior over where and what to read next, and the reported comparisons point in that direction.