NutriQA: A Traceable RAG Question-Answering Benchmark and Empirical Study on Official Chinese Nutrition Data

Abstract Retrieval-augmented generation (RAG) faces a hard constraint in public-facing nutri-tion QA: answers must be faithful to official sources. Yet Chinese official nutrition data lacks an auditable evaluation benchmark —numeric faithfulness, refusal bound-aries, and hallucination suppression remain unquantified. We present NutriQA, a traceable RAG question-answering benchmark on official Chinese nutrition data: 200 items in a four-category taxonomy (composition 60 / guideline 50 / special 50 / re-fusal 40), a three-level gold-provenance system (official food-composition values, ver-batim guideline excerpts, gold-free refusal items), and fully auditable deterministic judgment rules (numeric ±1 tolerance + key-point coverage + fabrication detection). Full-scale experiments on three commercial models (qwen-plus, deepseek-chat, glm-5.3-flash) show: (1) hybrid retrieval (BM25 character-bigram + vector RRF, top_k=8) achieves a 160/160 hit rate, with retrieval method choice outweighing chunking gran-ularity; (2) retrieval optimization lifts RAG gain from ~0.15–0.17 to 0.20–0.225, de-coupling retrieval hit from the generator; (3) a novel boundary subclass (out-of-KB foods × numeric baits) quantifies 100% numeric fabrication by bare LLMs over out-of-KB objects with mutually exclusive cross-model intervals (3–4× apart), which RAG compresses to ~0–2.5% via“retrieved-not-covered → refuse”; (4) deterministic rules agree with human annotation at 0.949, outperforming a generic LLM judge at 0.846. Data, code, and full evidence are pinned at git tag v1.0.0 for independent verification and reproduction.