面向复杂文档RAG的结构保真与质量门控技术研究综述

谭 智杰
黑龙江大学

摘要


复杂文档中的标题层级、阅读顺序、跨页表格及图文关联既承载内容，也决定知识边界。现有检索增强生
成系统多在解析后直接切分入库，使内容识别和结构识别误差沿切分、索引、检索与生成链路放大。围绕文档解析、
结构感知切分、动态检索和端到端评价研究，梳理复杂文档进入知识库前的关键技术，提出结构保真驱动的质量门
控闭环框架。该框架以多维结构保真度作为贯穿解析诊断、选择性修复、知识准入、检索路由、证据回溯与反馈校
准的控制信号，并给出准入、降级准入、修复和拒绝四类决策。所述框架为纯文献综合所得，旨在形成可检验的工
程研究议程，为企业文档知识库的可靠构建提供统一分析视角。

关键词


复杂文档；检索增强生成；结构保真；质量门控；知识库；文档解析

全文:

PDF


参考


[1]LEWIS P, PEREZ E, PIKTUS A, et al. RetrievalAugmented Generation for Knowledge-Intensive NLP

Tasks[J]. Advances in Neural Information Processing Systems,

2020, 33: 9459-9474.

[2]WANG B, XU C, ZHAO X, et al. MinerU: An

Open-Source Solution for Precise Document Content

Extraction[EB/OL]. arXiv:2409.18839, 2024.

[3]MA X, LIN S C, LI M, et al. Unifying Multimodal

Retrieval via Document Screenshot Embedding[C]//

Proceedings of EMNLP 2024. 2024: 6492-6505.

[4]YANG C, VU D K, NGUYEN M T, et al.

SuperRAG: Beyond RAG with Layout-Aware Graph

Modeling[C]//Proceedings of NAACL 2025 Industry Track.

2025: 544-557.

[5]YU R, WANG T, LIU R, et al. SF-RAG: StructureFidelity Retrieval-Augmented Generation for Academic

Question Answering[EB/OL]. arXiv:2602.13647, 2026.

[6]SOURATI Z, WANG Z, LIU M M, et al. LADRAG: Layout-aware Dynamic RAG for Visually-Rich

Document Understanding[C]//Proceedings of ACL 2026.

2026: 15945-15968.

[7]YU L, LIU H, WANG K, et al. ParseFixer: An

Agentic Framework for Document Parsing via Selective

Multimodal Correction[EB/OL]. arXiv:2606.11977, 2026.

[8]ZHOU B, XING H, CHEN Y, et al. MPDocBenchParse: Benchmarking Practical Multi-page Document

Parsing[EB/OL]. arXiv:2605.22100, 2026.

[9]ES S, JAMES J, ESPINOSA ANKE L, et al.

RAGAs: Automated Evaluation of Retrieval Augmented

Generation[C]//Proceedings of EACL 2024 System

Demonstrations. 2024: 150-158.

[10]MLCOMMONS. Introducing the MLPerf Endto-End RAG Inference Benchmark[EB/OL]. 2026-08-26.

https://mlcommons.org/2026/08/endtoend-inference/.


Refbacks

  • 当前没有refback。