跳转至

文章背景与核心概要

检索增强生成（RAG）对于利用大语言模型（LLM）生成准确且无幻觉的回答至关重要。然而，传统的文本切块（Chunking）方法丢弃了语料库中丰富的全局语义结构，而标准的大语言模型在长文本检索过程中往往会受到“迷失在中间”（lost in the middle）问题的困扰。

本文介绍了 STAIR（STructure Aware Information Retriever），这是一个利用全局文档结构（如目录 ToC）来高效存储和检索信息的新型系统。为了支持该领域的后续研究，作者推出了 SearchTome，这是一个由跨越 6 个不同领域的 18 本书构建的多样化基准测试集。与现有的先进检索器和基线模型相比，STAIR 展现出了卓越的性能，并将幻觉率显著降低。

STAIR (STructure Aware Information Retriever): A Novel Dataset and LLM-Based Retriever for Document Structure Augmentation

Summary

Retrieval-Augmented Generation (RAG) is essential for generating accurate, hallucination-free answers using Large Language Models (LLMs). However, traditional chunking methods discard the rich semantic global structure of corpora, and standard LLMs often struggle with the "lost in the middle" problem during long-context retrieval.

This paper introduces STAIR (STructure Aware Information Retriever), a novel system that leverages global document structures—such as Tables of Contents (ToCs)—to efficiently store and retrieve information. To support research in this area, the authors introduce SearchTome, a diverse benchmark built from 18 books spanning 6 different domains. STAIR demonstrates superior performance and significantly lower hallucination rates compared to existing state-of-the-art retrievers and baselines.

Metadata

  • arXiv ID: arXiv:2609.03874 [cs.AI]
  • Subjects: Artificial Intelligence (cs.AI)
  • Submission Date: September 3, 2026
  • Authors:
  • Vineet Kumar
  • Meghanadh Pulivarthi
  • Vishwajeet Kumar
  • Jaydeep Sen
  • Riyaz Ahmad Bhat
  • Sachindra Joshi

Abstract

检索增强生成（RAG）是利用大语言模型（LLM）生成准确且无幻觉回答的关键组件。大模型在处理长文本方面正在不断改进，但仍然饱受“迷失在中间”问题的困扰。因此，精确而准确的检索显得尤为重要。当前的检索器将长文本切分成基于长度的可管理块——在此过程中，却丢弃了语料库中丰富且信息量巨大的全局语义结构。我们引入了一种新颖的检索系统 STAIR，它赋能大模型利用语料库中的全局结构（如目录 ToC）来从其模型参数中高效地存储和检索信息。我们通过对微调后的可微分搜索索引（DSI）系统进行深入且仔细的消融研究表明，目录有助于构建低幻觉（小于 0.05%）的生成式信息检索（IR）系统，并且能够泛化到训练样本极少的场景。为了进一步推动基于目录检索这一新方向的研究，我们发布了 SearchTome——一个由跨越 6 个不同领域的 18 本书创建的多样化基准测试集。在 SearchTome 上，STAIR 实现了高达 82.6% 的 Recall@1 得分，相比之下 DSI 为 76.9%，并且差异具有统计学显著性。STAIR 轻松击败了其他强基线，如 BM25（59.5%）、DPR（68.7%）和开箱即用的 Mistral（13.8%）。

Retrieval Augmented Generation (RAG) is a key component for generating accurate and hallucination free answers using Large Language Models (LLMs). LLMs are improving at handling long context, but still suffer from "lost in the middle" problem. Thus, precise and accurate retrieval is important. Current retrievers chunk long context into length-based manageable chunks - in the process throwing away rich and informative semantic global structure in the corpus. We introduce a novel retrieval system STAIR that empowers an LLM to exploit global structure in a corpus such as a Table of Contents (ToC) to efficiently store and retrieve information from its model parameters. Our thorough and careful ablation studies with a finetuned Differentiable Search Index (DSI) system show that ToC helps build a low hallucination (less than 0.05%) generative Information Retrieval (IR) system and can generalize to examples where very few training samples are available. To further research in this novel direction of ToC based retrieval we release SearchTome - a diverse benchmark created from 18 books across 6 diverse domains to further research in this novel direction. STAIR achieves a high Recall@1 score of 82.6% on SearchTome as compared to DSI (76.9%), where the difference is found to be statistically significant. STAIR easily beats other strong baselines such as BM25 (59.5%), DPR (68.7%) and out-of-the-box Mistral (13.8%).

Key Findings & Performance

  • 幻觉率： 借助基于目录（ToC）的增强，STAIR 在生成式信息检索（IR）设置中实现了低于 0.05% 的极低幻觉率。
  • 泛化能力： 即使在可用训练样本极少的场景下，也能表现得异常出色。
  • 在 SearchTome 基准测试上的表现：
  • STAIR： 82.6% (Recall@1)
  • DSI（可微分搜索索引）： 76.9% (Recall@1)
  • DPR（密集段落检索）： 68.7% (Recall@1)
  • BM25： 59.5% (Recall@1)
  • Mistral（开箱即用）： 13.8% (Recall@1)
  • Hallucination Rate: Built with ToC-based augmentation, STAIR achieves an exceptionally low hallucination rate of less than 0.05% in a generative Information Retrieval (IR) setup.
  • Generalization: Performs exceptionally well even in scenarios with very few training samples available.
  • Performance on SearchTome Benchmark:
  • STAIR: 82.6% (Recall@1)
  • DSI (Differentiable Search Index): 76.9% (Recall@1)
  • DPR (Dense Passage Retrieval): 68.7% (Recall@1)
  • BM25: 59.5% (Recall@1)
  • Mistral (Out-of-the-box): 13.8% (Recall@1)

Access & Resources