GRADRAG: Cross-Component Prompt Adaptation for Coordinated Multi-Agent RAG

2026-07-23

AI Summary

Existing multi-agent RAG systems optimize individual components in isolation, lacking effective coordination mechanisms. This work proposes a cross-component prompt adaptation framework that models the RAG pipeline as a computational graph. Leveraging an LLM-driven evaluator, the framework generates structured feedback on both answers and evidence, which, combined with an early-stopping mechanism, enables iterative refinement of upstream modules—including the retriever, graph builder, and answer generator—toward end-to-end joint optimization. By integrating graph-augmented retrieval with query refinement, the method substantially outperforms single-step generator optimization baselines on the SQUALITY and QMSUM benchmarks, achieving a net preference gain of 12–15 percentage points in LLM pairwise comparisons, with most performance improvements realized within two iterations.

Abstract

Retrieval-Augmented Generation (RAG) systems increasingly employ multiple LLM agents. Yet, most prior work optimizes components in isolation rather than coordinating improvements across the pipeline. We introduce GRADRAG, a framework for cross-component prompt adaptation that models the RAG pipeline as a computational graph and propagates structured evaluation feedback to update upstream agents. An Evaluator critiques downstream answers and supporting evidence, producing actionable feedback that a Prompt Optimizer uses to iteratively update adaptive agents, such as retrievers, graph constructors, and answerers. The Evaluator also triggers early stopping when the output is deemed satisfactory. We evaluate GRADRAG on the SQUALITY and QMSUM benchmarks under two retrieval paradigms: flat chunk-based retrieval using IRCoT-style query refinement (Trivedi et al., 2023), and graph-based retrieval that constructs and iteratively enriches an entity-relation graph from the document. Across both settings, GRADRAG consistently outperforms one-step refinement baselines that update only the final generator, achieving a 12-15 percentage point net preference margin in LLM-judged pairwise comparisons, with most gains realized within two refinement iterations.

Problem

• Retrieval-Augmented Generation
• Multi-Agent Coordination
• Prompt Adaptation
• Cross-Component Optimization
• RAG Pipeline

Innovation

• cross-component adaptation
• computational graph
• structured feedback propagation
• multi-agent RAG
• prompt optimization