DF-RAG: Query-Aware Diversity for Retrieval-Augmented Generation

TL;DR

Retrieval-augmented generation often fails on reasoning-intensive QA due to redundancy when using cosine-based retrieval. The authors propose DF-RAG, a training-free, test-time, query-adaptive retrieval framework that uses a geometry-based gMMR score to balance relevance and diversity, guided by a Planner–Evaluator–Generator architecture. They formalize the per-query diversity via the gMMR score, establish an Oracle upper bound that showcases possible gains, and demonstrate that DF-RAG consistently improves over strong baselines across five datasets, approaching a substantial portion of the Oracle gap while keeping retrieval efficient. The work highlights the practical value of adaptive diversity in retrieval for complex QA and provides a plug-and-play approach that can scale with large corpora and long contexts.

Abstract

Retrieval-augmented generation (RAG) is a common technique for grounding language model outputs in domain-specific information. However, RAG is often challenged by reasoning-intensive question-answering (QA), since common retrieval methods like cosine similarity maximize relevance at the cost of introducing redundant content, which can reduce information recall. To address this, we introduce Diversity-Focused Retrieval-Augmented Generation (DF-RAG), which systematically incorporates diversity into the retrieval step to improve performance on complex, reasoning-intensive QA benchmarks. DF-RAG builds upon the Maximal Marginal Relevance framework to select information chunks that are both relevant to the query and maximally dissimilar from each other. A key innovation of DF-RAG is its ability to optimize the level of diversity for each query dynamically at test time without requiring any additional fine-tuning or prior information. We show that DF-RAG improves F1 performance on reasoning-intensive QA benchmarks by 4-10 percent over vanilla RAG using cosine similarity and also outperforms other established baselines. Furthermore, we estimate an Oracle ceiling of up to 18 percent absolute F1 gains over vanilla RAG, of which DF-RAG captures up to 91.3 percent.

Figures (12)

• Figure 1: Architecture of DF-RAG: A training-free framework that dynamically selects the right level of diversity to generate answers for reasoning-intensive QA tasks. DF-RAG adaptively adjusts diversity for each query by utilizing the Planner, Evaluator and the Diversity-focused Retrieval. We illustrate this with an example query from HotpotQA yang-etal-2018-hotpotqa.
• Figure 2: $F_1$ scores of vanilla RAG, RAG with gMMR, Long Context, and Oracle across LongBench benchmarks. Vanilla RAG, RAG with gMMR and Oracle operate at 1500 word context length. Results on En.QA are not reported as its context length exceeds our context limit.
• Figure 3: Performance ($F_1$ score) curves, of Oracle, vanilla RAG, and best DF--RAG across context lengths on all five QA benchmarks. The first row shows Qwen 2.5 72B as the backbone LLM whereas the second row shows Llama 3.3 70B.
• Figure 4: Heatmap showing Jaccard index values between candidate chunk-sets sampled at different $\lambda$. Darker color represents more common chunks between the corresponding chunk-sets.
• Figure 5: Runtime comparison of DF-RAG per sample with parallelizable worker nodes.