A Procedural-aware Layout Chunking Approach for SOP Question Answering in Retrieval-augmented Generation Pipelines

Read full text Summarise with VIRAT
Authors: Faqih Al Suman, Deffa Rahadiyan
Cited by: 0
References: 33

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Standard operating procedures (SOPs) are often distributed as semi-structured PDFs that mix headings, tables, and multi-step procedures.Generic chunking (fixed windows or raw layout blocks) can fragment procedural steps, causing retrieval to return incomplete or misleading evidence for SOP question answering.We propose Procedural-Aware Layout Chunking (PALC), which reconstructs heading hierarchy and step boundaries from PDF layout blocks, merges contiguous steps into procedure-centric chunks, and attaches section-path metadata.On a benchmark of 20 SOP PDFs and 100 questions, PALC achieves the best step-level retrieval across K={5,10,15}.At K=10, PALC improves step-level Recall@10 from 0.204 (Fixed) and 0.177 (Layout) to 0.371, and nDCG@10 from 0.246/0.197to 0.349 (95% bootstrap CIs reported).End-to-end answer faithfulness also improves: average human faithfulness is 1.39 for PALC vs 1.22 (Fixed) and 0.74 (Layout), and Gemini-based faithfulness is 1.31 vs 1.27 and 0.77.Controlled ablations isolate the contributions of step merging, metadata, and table handling, and mechanism metrics show PALC reduces step fragmentation compared to baselines.These results suggest that procedure-aware chunk formation is a practical way to improve SOP-focused RAG pipelines.

Fields of Study & Subjects

Topic Modeling Information Retrieval and Search Behavior Handwritten Text Recognition Techniques

Publication details

Journal / source
International journal of intelligent engineering and systems
Volume
19
Issue
4
Pages
1039–1050
Publication date
23 March 2026
Language
EN