Rethinking Retrieval-Augmentation as Synthesis: A Query-Aware Context Merging Approach

2026-03-18

AI Summary

This work proposes MergeRAG, a novel framework that addresses the limitations of traditional retrieval-augmented generation (RAG) methods, which struggle to balance critical information retention and token efficiency within the constrained context windows of large language models. Conventional approaches often discard bridging evidence in long-tail retrieved passages and introduce redundancy by selecting only top-k fragments. MergeRAG reframes retrieval augmentation as a dynamic optimization problem, employing a query-aware context synthesis mechanism to reconstruct retrieved content. Its core innovations include symmetric merging to integrate weakly relevant signals, asymmetric merging guided by entropy-based anchoring for redundancy reduction, and a hierarchical parallel merging strategy to enhance both efficiency and information preservation. Experimental results demonstrate that MergeRAG significantly outperforms existing methods on standard benchmarks, achieving gains of up to 13.7 points in F1 score and 11.5 points in exact match (EM).

Technology Category

• Data Mining & Knowledge Management: Conversational Systems for Recommendation & Retrieval
• Search and Optimization: Metareasoning and Metaheuristics
• Natural Language Processing: Information Extraction

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Retrieval-Augmented Generation (RAG) enables Large Language Models (LLMs) to extend their existing knowledge by dynamically incorporating external information. However, practical deployment is fundamentally constrained by the LLM's finite context window, forcing a trade-off between information sufficiency and token consumption. Standard pipelines address this via a retrieve-then-select strategy, typically retaining only the top-k chunks based on relevance. Nevertheless, this approach is suboptimal: it inherently truncates critical bridging evidence located in the long tail of the relevance distribution, while simultaneously wasting the token budget on semantically redundant high-ranking chunks. In this paper, we rethink retrieval-augmentation as a dynamic optimization problem aimed at maximizing information density. We propose MergeRAG, a novel framework that shifts the paradigm from static filtering to query-aware synthesis. MergeRAG employs a scoring agent to restructure retrieved contexts through a dual-pathway mechanism: 1) Symmetric Merging, which consolidates weak signals to recover lost bridging evidence; 2) Asymmetric Merging, which utilizes entropy-guided anchoring to eliminate redundancy without sacrificing semantic integrity. We further introduce a Hierarchical Parallel Merging strategy that mitigates information loss while maximizing computational parallelism. Extensive experiments on standard benchmarks demonstrate that MergeRAG significantly outperforms state-of-the-art RAG baselines, achieving up to 13.7 points improvement in F1 score and 11.5 points in Exact Match (EM), respectively.

Problem

• Retrieval-Augmented Generation
• context window limitation
• information redundancy
• bridging evidence
• token budget

Innovation

• Retrieval-Augmented Generation
• Context Merging
• Information Density
• Query-Aware Synthesis
• Entropy-Guided Redundancy Reduction