[Paper Review] Incorporating Q&A Nuggets into Retrieval-Augmented Generation

Laura Dietz, Bryan Li+6 more|arXiv (Cornell University)|Jan 19, 2026

Biomedical Text Mining and Ontologies0 citations

TL;DR

Crucible uses a nugget-first, Q&A nugget bank to guide retrieval, extraction, and report assembly in a RAG system, achieving higher nugget recall and grounding than Ginger on the TREC NeuCLIR 2024 test.

ABSTRACT

RAGE systems integrate ideas from automatic evaluation (E) into Retrieval-augmented Generation (RAG). As one such example, we present Crucible, a Nugget-Augmented Generation System that preserves explicit citation provenance by constructing a bank of Q&A nuggets from retrieved documents and uses them to guide extraction, selection, and report generation. Reasoning on nuggets avoids repeated information through clear and interpretable Q&A semantics - instead of opaque cluster abstractions - while maintaining citation provenance throughout the entire generation process. Evaluated on the TREC NeuCLIR 2024 collection, our Crucible system substantially outperforms Ginger, a recent nugget-based RAG system, in nugget recall, density, and citation grounding.

Motivation & Objective

• Motivate the use of fine-grained Q&A nuggets as reusable information units for evaluation and generation.
• Propose Crucible, a nugget-centric RAG system that constructs and uses a nugget bank to control retrieval, extraction, and report assembly while preserving citation provenance.
• Evaluate Crucible against state-of-the-art nugget-based RAG systems on the TREC NeuCLIR 2024 dataset to demonstrate gains in nugget recall, density, and grounding.
• Show how nugget-centric design reduces redundancy without clustering and maintains explicit citation provenance throughout generation.

Proposed method

• Nugget ideation: retrieve initial documents, generate query-focused summaries per document, and produce Q&A nuggets conditioned on the user request and summary.
• Paraphrase detection and merging to create a canonical top‑20 nugget bank per request.
• Nugget ranking via a Support-Vector Classifier trained on 19 quality features plus readability and complexity measures, fused with popularity cues.
• Nugget-first retrieval and scanning: for each nugget, locate supporting passages, extract concise self-contained sentences, and record extraction likelihood as extraction confidence.
• Sentence selection: rank candidate sentences per nugget by extraction confidence and select the top one sentence per nugget (k=1).
• Assembly: concatenate selected sentences in nugget quality order, with each sentence tied to exactly one citation, ensuring citation provenance is preserved.

Experimental results

Research questions

• RQ1 Can a nugget-first RAG pipeline improve nugget recall, density, and citation grounding compared to cluster-based or end-to-end approaches?
• RQ2 How does Crucible perform relative to Ginger and other baselines on Nugget Recall, Nugget Density, Sentence Novelty, Relevant Sentences, and Citation Support on the NeuCLIR 2024 task?
• RQ3 What is the impact of using different retrievers (e.g., Milco, Qwen3) and verification steps on nugget-level and citation-grounded metrics?

Key findings

• Crucible outperforms Ginger and Ginger-LLaMA across nugget-oriented metrics on NeuCLIR 2024.
• Nugget recall improves by 42% to 65% and nugget density by 21% to 25% with Crucible.
• Crucible-Verified achieves higher scores on citation support and sentence novelty compared to Crucible-Base.
• Crucible demonstrates robust performance under multiple retrievers, with verification steps providing additional gains in grounding.
• The design preserves explicit citation provenance by ensuring each sentence is self-contained and tied to a single citation.