💬 NLP

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

The paper introduces RARG, a relevance-aware search agent that leverages relevance scores to guide coarse-to-fine corpus interaction via ordered document traversal and match reranking, thereby improving both the accuracy and efficiency of complex question answering compared to existing retrieval and direct-interaction methods.

Original authors: Jiangnan Li, Yuqing Li, Mo Yu, Jinchao Zhang, Jie Zhou

Published 2026-07-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jiangnan Li, Yuqing Li, Mo Yu, Jinchao Zhang, Jie Zhou

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery, but instead of a few clue-filled notebooks, you are handed a library containing millions of books. Your goal is to find the one specific sentence that holds the answer. In the world of artificial intelligence, this is called "retrieval." For a long time, AI assistants acted like a librarian who would quickly scan the titles of the top ten books and hand them to you, hoping the answer was right there on the first page. But what if the answer was hidden deep inside a paragraph on page 400 of a book that wasn't even in the top ten? Or what if the clue was a tiny, specific phrase that only appeared when you searched for a very specific pattern?

This is where "Agentic Search" comes in. Instead of just asking a librarian for a list, an AI "agent" is given a set of tools to walk into the library itself. It can open any book, read specific pages, and use a search tool (like a digital "grep" or "find" command) to hunt for keywords. However, there's a catch: if the agent just starts searching randomly through millions of books, it might waste hours looking in the wrong aisles before it ever finds the right clue. The big question researchers are asking is: How do we teach the AI to know which books to open first, and which sentences to read, so it solves the mystery faster and with less effort?

The paper you are about to read introduces a new character in this story: RARG (Relevance-Aware RipGrep Search Agent). Think of RARG as a super-smart detective who doesn't just search randomly. Before the detective even picks up a magnifying glass, they get a "relevance map." This map doesn't give them the answer; instead, it tells them which books are most likely to contain the clues. RARG uses this map to decide the order in which it opens the books. It starts with the most promising books, ensuring that if the answer is there, the detective finds it immediately.

But RARG doesn't stop there. It has two special tricks. First, it gives the detective a "starter pack" of a few highly relevant paragraphs right at the beginning, so they don't have to waste time guessing where to start. Second, when the detective finds a bunch of matching sentences (hits) across many books, RARG acts like a strict editor. It looks at all those matches and re-ranks them, pushing the most informative sentences to the top and hiding the boring, irrelevant ones. This ensures the detective only sees the best clues, even if those clues came from a book that wasn't originally at the very top of the list.

The researchers tested this new detective against older methods. They found that by using relevance to guide the search while the agent was looking, rather than just using it to pick a list of books beforehand, the AI solved complex questions much faster. In tests with a library of 100,000 documents, the new method (RARG++) reached 84% accuracy, beating the previous best methods which hovered around 78%. When they expanded the library to 1 million documents, the new method's accuracy did drop to 79%, but it still maintained a significant lead over the older methods, which fell to 69%. Even with this drop, the new method used significantly fewer "tool calls" (steps in the search process), making it much cheaper and faster.

The paper suggests that the key to better AI search isn't just having a bigger library or a smarter search tool, but giving the agent a better sense of direction. By turning "relevance" from a simple filter into a guide for how the search happens, the AI can find the needle in the haystack without having to move the entire haystack first. The authors show that this approach works well for both finding specific answers (like a math problem) and ranking lists of information, proving that knowing where to look is just as important as knowing how to search.

Technical Summary: A New Role for Relevance in Agentic Search (RARG)

Problem Statement

Current agentic search systems face a fundamental tension between scalability and evidence localization. Traditional retrieval agents rely on a "top-k" interface, where a retriever ranks documents and exposes a fixed set of snippets to the Large Language Model (LLM). This approach conflates document-level relevance with evidence utility; a highly relevant document may hide decisive facts in spans truncated by snippet limits, or require multi-hop reasoning that a static ranked list cannot support.

Conversely, Direct Corpus Interaction (DCI) allows agents to use terminal tools (e.g., grep, cat) to search raw documents, enabling fine-grained operations like pattern matching and entity following. However, standard DCI is largely relevance-agnostic. It treats all documents as equally promising, forcing the agent to traverse the corpus blindly. This leads to "search convergence" issues: useful clues appear late in the output stream, or are lost due to truncation, causing the agent to waste interactions on unproductive search angles.

Recent hybrid approaches attempt to solve this by using retrieval to construct a bounded "working space" (e.g., RISE, DR-DCI) before interaction. While this limits the search space, these methods still treat relevance as a one-time filter for space construction. They fail to use relevance scores as an execution prior to guide how the agent traverses the space (traversal order) or which local matches are visible when output budgets are exceeded.

Methodology: Relevance-Aware RipGrep Search Agent (RARG)

The authors propose RARG, a search agent that integrates relevance estimation directly into the execution flow of corpus interaction. RARG transforms relevance from a static selection mechanism into a dynamic guide for the ripgrep (rg) tool. The system operates on three levels of coarse-to-fine guidance:

1. Document-Level Relevance Insertion (RARG)

Instead of returning document content, RARG uses an embedding retriever to generate a scope file (/tmp/scope_N.txt) containing the paths of the top-ranked documents (capped at 10,000) relative to an agent-generated query.

Execution Prior: The agent is instructed to traverse this scope file sequentially using rg.
Sequential Traversal: Standard rg runs multi-threaded, destroying path order. RARG injects a -j1 flag to force single-threaded, sequential scanning. This ensures that matches from globally relevant documents surface earlier in the output stream, exposing promising clues before less relevant ones.
Tooling: The agent uses embed_recall(query) to generate the scope and cat scope | xargs rg to search within it.

2. Entry Point Initialization (RARG+)

To prevent the agent from wasting steps finding a starting point, RARG+ appends a small set of query-relevant paragraphs (top-10 from the scope) to the initial scope mapping. These paragraphs serve as an "entry point" for the LLM to formulate its first precise search pattern, rather than acting as the final answer.

3. Match-Level Relevance Insertion (RARG++)

Even within a relevant document, rg output can be truncated, potentially hiding informative excerpts buried in noise. RARG++ introduces a reranking mechanism for the matches themselves:

Constructed Query: The system constructs a query combining the global scope query and the local rg pattern keywords.
Reranking: An embedding model reranks a broader pool of rg matches (up to $M=500$ ) based on this constructed query.
Selection: Only the top $m$ matches are presented to the LLM. This ensures that locally informative excerpts, even if they reside in documents with lower global relevance scores, are visible to the model.

Key Contributions

Reconceptualizing Relevance: The paper identifies that relevance should not merely select inputs or manage a workspace but should act as an execution prior for corpus interaction, guiding traversal order and match visibility.
RARG Architecture: The proposal of a relevance-aware rg agent that preserves document rankings during traversal and introduces entry-point initialization and match-level reranking to provide coarse-to-fine guidance.
Empirical Validation: Demonstration that relevance-aware interaction achieves a superior accuracy-efficiency trade-off compared to retrieval-only agents, unrestricted DCI, and space-constructed interaction agents across challenging QA and reasoning-intensive retrieval tasks.

Experimental Results

The authors evaluated RARG on BrowseComp-Plus (challenging QA with 100K–1M documents) and BRIGHT (reasoning-intensive retrieval across four domains).

Accuracy vs. Cost (BrowseComp-Plus):
- On a 100K-document corpus with GPT-5.4-mini, RARG++ achieved 84% accuracy, outperforming RISE (78%) and DCI (78%).
- Crucially, RARG++ required only 23.9 tool calls on average, compared to 28.7 for RISE and 99.1 for DCI.
- With GPT-5.4, RARG++ reached 91% accuracy (a 9-point gain over RISE) with significantly fewer turns.
Corpus Scaling:
- When scaling the corpus to 1M documents, RARG++ maintained 79% accuracy, while RISE-BM25 dropped to 69%. This suggests relevance-guided traversal is more robust to increased noise and distractors than pure lexical expansion.
Reasoning-Intensive Retrieval (BRIGHT):
- On BRIGHT, RARG+ achieved the highest average nDCG@10 (53.36), outperforming the retrieval-specialized NeMo agent (52.89) and DCI.
- Interestingly, for BRIGHT (which favors breadth/recall), RARG+ (initialization only) outperformed RARG++ (reranking), suggesting that for tasks requiring broad candidate recall, aggressive match-level filtering may be less optimal than for QA tasks requiring rapid convergence.

Significance and Claims

The paper claims that relevance-aware interaction enables faster and more reliable search convergence. By turning relevance into an execution prior, RARG allows agents to:

Expose globally relevant clues earlier by ordering document traversal.
Surface informative excerpts that might otherwise be obscured by document-level ranking or output truncation.
Reduce tool usage significantly while maintaining or improving accuracy, effectively advancing the accuracy-efficiency frontier.

The authors emphasize that this approach does not replace the need for retrieval but integrates it deeply into the interaction loop, preserving the high-resolution, compositional interface of DCI while mitigating its tendency toward unproductive exploration. The results suggest that relevance is a critical, yet underutilized, signal for guiding the execution of agentic search, not just the selection of its inputs.