Incorporating Q&A Nuggets into Retrieval-Augmented Generation

1 Introduction

• Nuggets encode information an answer should cover and support metrics such as nugget recall.The paper argues they can guide retrieval and generation as well as evaluation.Source: §1 · p. 1
• Crucible automatically constructs a nugget bank and uses nuggets as a control signal throughout its pipeline.The system uses nuggets to control redundancy without content-based clustering while preserving citation provenance.Source: §1 · p. 2
• On TREC NeuCLIR 2024, Crucible decisively outperforms Ginger across several metrics, from nugget recall to citation support.Source: §1 · p. 2

2 Related Work

• Passage-by-passage RAG preserves citation provenance but can harm coherence and fluency by making the model attend to many independent inputs.Source: §2 · p. 2
• Joint representations improve synthesis and fluency, but compact latent representations can discard explicit source links and make citation grounding difficult.HyDE's answer-first approach retrieves supporting evidence afterward, raising trustworthiness concerns.Source: §2 · p. 2
• Ginger clusters verbatim extracted spans into topical facets, reranks the clusters, and rewrites summaries of top clusters for fluency.Source: §2 · p. 2
• Clustering summarization abstracts away from original document extractions, impacting faithful citation grounding.Source: §2 · p. 2
• Agentic RAG decomposes pipelines into subtasks; GptResearcher [7] orchestrates query generation, retrieval, extraction, and report writing, with optional verification and trace-back.Source: §2 · p. 2

3 Nugget-first RAG Approach: Crucible

• Crucible organizes its workflow around structured Q&A nuggets that guide retrieval, extraction, and report assembly, unlike Ginger's clustered verbatim spans.Source: §3 · p. 2
• For each retrieved document, Crucible generates a query-focused summary and prompts LLaMA-3.3-70B-Instruct to produce Q&A nuggets conditioned on that summary and the user request.Source: §3 · p. 3
• An LLM detects paraphrase nuggets, which are successively merged from the most confident pairs until the desired number of canonical nuggets is reached.Source: §3 · p. 3
• A feature-based SVC ranking is fused with popularity ranking to select the top 20 nuggets for each request.The SVC uses 19 quality features, including LLM-judge measures of information-need fit and vitality, plus readability and sentence-complexity features.Source: §3 · p. 3
• For each nugget, Crucible locates supporting passages, extracts self-contained candidate sentences, and selects the top k by extraction confidence; k = 1 here.Each selected sentence is tied to exactly one citation, and duplicate stopped/stemmed text is omitted during assembly.Source: §3 · pp. 3–4, Figure 1 · p. 3
• Optional verification checks nugget coverage and citation support, but Crucible-Base skips it because verification risks leaking evaluator knowledge into the system.Source: §3 · p. 4

4 Evaluation Setup

• TREC NeuCLIR 2024 requires topic-focused, citation-supported reports from a multilingual corpus whose language subsets are machine-translated to English.Topics are broad, but reports must be tailored to the specific problem statement and user background; AutoArgue uses held-out, manually curated gold-nugget banks.Source: §4 · pp. 4–5
• The comparison includes Crucible-Base and Crucible-Verified, Ginger variants, GptResearcher, and BulletPoints.Ginger's best reported variant uses Qwen3-Embedding retrieval and GPT-4o, while Ginger-Llama uses the same LLaMA model as Crucible.Source: §4 · p. 5
• AutoArgue metrics cover nugget recall and density, sentence novelty and relevance, and citation support.Source: §4 · p. 5

• +56% nugget density and +267% relevant sentences are reported gains as Crucible outperforms Ginger and Ginger-LLaMA across nugget-oriented metrics.Table 1 compares RAG systems using PlaidX on TREC NeuCLIR 2024.Source: §5 · p. 5, Table 1 · p. 6
• BulletPoints significantly improves over Crucible-Base on nugget recall, density, and relevant sentences, but exceeds it only on citation support.Source: §5 · p. 6
• Crucible performs well across explored retrieval models, with similar results for Ginger; only Crucible-Verified performs better with PlaidX.Source: §5 · p. 6
• +26% nugget density and +33% sentence novelty accompany verification evaluated against gold nuggets; citation support rises +64% to 0.932.Source: §5 · p. 6
• Crucible submissions to TREC 2025 tracks provide preliminary corroborating evidence, although many relevance judgments were unavailable at writing.Source: §5 · p. 7
• Runtime is dominated by LLM calls, including Θ(D^2) paraphrase-detection prompts, Θ(N · D) nugget-scanning calls, and Θ(S) verification prompts.The system was designed to study quality improvements, not to provide a fast deployment-ready system.Source: §5 · p. 7
• The authors warn that AutoArgue-based verification may make near-perfect citation-support scores reflect circularity with the AutoArgue evaluation system.They report results both with and without verification.Source: §5 · p. 7

• On NeuCLIR, Crucible decisively outperforms Ginger, demonstrating the practical benefits of controlling generation.Source: §6 · p. 7
• 0.9 and 0.96 citation support scores are reported for Crucible-Base and Crucible-Verified, respectively, alongside extraction and citation-provenance rationale.Crucible selects and summarizes sentences addressing canonical Q&A nuggets, with one citation per generated sentence.Source: §6 · p. 7
• Future work should explore a greater variety of RAGE systems across RAG and LLM-as-a-Judge paradigms.Source: §6 · p. 8