GRASP Teaches Agentic RAG When and How to Search

Introduction

Agentic retrieval-augmented generation is moving RAG beyond the classic pattern of retrieving once and generating once. Instead, a language model can reason step by step, issue new queries, inspect evidence, and then decide whether it has enough information to answer. That flexibility is powerful, but it also creates a difficult control problem: when should the agent search again, what kind of search should it use, and how much surrounding context should it read?

The paper GRASP: GRanularity-Aware Search Policy for Agentic RAG tackles this problem directly. It proposes a reinforcement learning framework that trains a retrieval agent to coordinate multiple search actions and control the granularity of context during multi-step reasoning.

Key ideas

• Search less blindly, search more appropriately: GRASP is not mainly about stuffing more retrieved text into the prompt. It allows the agent to retrieve sentence-level evidence first and expand to paragraph-level context only when needed, reducing the chance that irrelevant tokens disrupt reasoning.
• Three complementary actions: The agent can use semantic search, keyword search, and paragraph reading. Semantic search supports broad exploration, keyword search helps locate entity-specific or exact-match evidence, and paragraph reading provides local context for verification.
• A reinforcement learning policy: The training reward jointly considers answer accuracy, grounded reading, complementary use of search tools, and turn efficiency. In other words, the agent is encouraged not only to answer correctly, but also to retrieve evidence in a disciplined and efficient way.
• Designed for multi-hop reasoning: Multi-hop question answering often requires connecting several pieces of evidence. GRASP aims to learn behavior that resembles skimming for direction, scanning for specific clues, and reading locally to confirm the final chain of reasoning.

Why it matters

The significance of GRASP is that it shifts part of the Agentic RAG problem from prompt design to policy learning. Many RAG systems still rely on a fixed retrieval recipe: retrieve a set of passages, place them into context, and ask the model to answer. Agentic systems add tool use and iteration, but without a reliable policy they can over-search, repeat queries, or become distracted by overly broad context.

GRASP treats retrieval as a set of complementary skills rather than a single operation. Semantic retrieval opens the search space, lexical retrieval narrows down exact clues, and paragraph reading expands context only when the evidence demands it. The qualitative analysis described in the paper suggests that the learned agent develops interpretable skimming and scanning behavior.

For real-world applications, this direction is relevant to enterprise knowledge assistants, literature review agents, compliance research, and complex support workflows. These settings often have abundant documents but limited context budgets. A granularity-aware search policy can improve evidence recall while keeping the reasoning context cleaner and more traceable.

The provided material does not include concrete numerical results, so the exact magnitude of improvement requires reading the full paper. Still, the core message is clear: stronger RAG agents will not simply retrieve more content. They will need to learn when to search, which signal to trust, and how deeply to read.

Source: Hugging Face Daily Papers