Rethinking Retrieval-Augmentation as Synthesis: A Query-Aware Context Merging Approach
New framework ditches simple 'top-k' retrieval for a synthesis approach that recovers lost evidence.
A research team led by Jiarui Guo has introduced MergeRAG, a novel framework that fundamentally shifts the paradigm for Retrieval-Augmented Generation (RAG). Standard RAG pipelines are constrained by an LLM's context window, forcing a trade-off where only the top-k most relevant text chunks are used. This often discards crucial 'bridging evidence' found in less relevant passages while wasting tokens on redundant high-ranking information. MergeRAG redefines the problem as one of maximizing information density through dynamic, query-aware synthesis.
The framework employs a scoring agent that restructures retrieved contexts via two key mechanisms. First, 'Symmetric Merging' consolidates multiple weak signals to recover that lost bridging evidence. Second, 'Asymmetric Merging' uses entropy-guided anchoring to eliminate semantic redundancy without losing key information. To make this efficient, the team also developed a Hierarchical Parallel Merging strategy.
Extensive testing on standard benchmarks shows the power of this approach. MergeRAG significantly outperforms state-of-the-art RAG baselines, achieving improvements of up to 13.7 points in F1 score and 11.5 points in Exact Match (EM). This demonstrates that intelligently synthesizing context, rather than just filtering it, is a superior path for grounding LLMs in external knowledge.
- Proposes MergeRAG, shifting RAG from static 'top-k' filtering to query-aware context synthesis.
- Uses Symmetric Merging to recover lost 'bridging evidence' and Asymmetric Merging to cut redundancy.
- Achieves up to a 13.7-point F1 score improvement over standard RAG methods on benchmarks.
Why It Matters
Enables more accurate, evidence-rich AI responses by making far better use of an LLM's limited context window.