Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cross-Document Topic-Aligned Chunking for Retrieval-Augmented Generation

Mile Stankovic

TL;DR

This paper addresses the knowledge fragmentation problem in retrieval-augmented generation by introducing Cross-Document Topic-Aligned (CDTA) chunking, which reconstructs corpus-wide knowledge around global topics. CDTA identifies topics across documents, maps segments to topics, and synthesizes them into unified, topic-centered chunks using a six-stage pipeline (segmentation, topic extraction, deduplication, relevance mapping, aggregation, synthesis) with indexing and embedding. Empirical results on HotpotQA and UAE Legal show CDTA achieving higher faithfulness ($0.93$ on HotpotQA and $0.94$ on UAE Legal) and robust retrieval efficiency, particularly at low $k$ (e.g., $k=3$ with $0.91$ faithfulness), at the cost of substantially higher indexing time and API usage. The findings demonstrate meaningful improvements in multi-document reasoning and high-query-volume applications, with trade-offs discussed for deployment, including cost, latency, and the potential for hybrid approaches with existing intra-document methods.

Abstract

Chunking quality determines RAG system performance. Current methods partition documents individually, but complex queries need information scattered across multiple sources: the knowledge fragmentation problem. We introduce Cross-Document Topic-Aligned (CDTA) chunking, which reconstructs knowledge at the corpus level. It first identifies topics across documents, maps segments to each topic, and synthesizes them into unified chunks. On HotpotQA multi-hop reasoning, our method reached 0.93 faithfulness versus 0.83 for contextual retrieval and 0.78 for semantic chunking, a 12% improvement over current industry best practice (p < 0.05). On UAE Legal texts, it reached 0.94 faithfulness with 0.93 citation accuracy. At k = 3, it maintains 0.91 faithfulness while semantic methods drop to 0.68, with a single CDTA chunk containing information requiring multiple traditional fragments. Indexing costs are higher, but synthesis produces information-dense chunks that reduce query-time retrieval needs. For high-query-volume applications with distributed knowledge, cross-document synthesis improves measurably over within-document optimization.

Cross-Document Topic-Aligned Chunking for Retrieval-Augmented Generation

TL;DR

This paper addresses the knowledge fragmentation problem in retrieval-augmented generation by introducing Cross-Document Topic-Aligned (CDTA) chunking, which reconstructs corpus-wide knowledge around global topics. CDTA identifies topics across documents, maps segments to topics, and synthesizes them into unified, topic-centered chunks using a six-stage pipeline (segmentation, topic extraction, deduplication, relevance mapping, aggregation, synthesis) with indexing and embedding. Empirical results on HotpotQA and UAE Legal show CDTA achieving higher faithfulness ( on HotpotQA and on UAE Legal) and robust retrieval efficiency, particularly at low (e.g., with faithfulness), at the cost of substantially higher indexing time and API usage. The findings demonstrate meaningful improvements in multi-document reasoning and high-query-volume applications, with trade-offs discussed for deployment, including cost, latency, and the potential for hybrid approaches with existing intra-document methods.

Abstract

Chunking quality determines RAG system performance. Current methods partition documents individually, but complex queries need information scattered across multiple sources: the knowledge fragmentation problem. We introduce Cross-Document Topic-Aligned (CDTA) chunking, which reconstructs knowledge at the corpus level. It first identifies topics across documents, maps segments to each topic, and synthesizes them into unified chunks. On HotpotQA multi-hop reasoning, our method reached 0.93 faithfulness versus 0.83 for contextual retrieval and 0.78 for semantic chunking, a 12% improvement over current industry best practice (p < 0.05). On UAE Legal texts, it reached 0.94 faithfulness with 0.93 citation accuracy. At k = 3, it maintains 0.91 faithfulness while semantic methods drop to 0.68, with a single CDTA chunk containing information requiring multiple traditional fragments. Indexing costs are higher, but synthesis produces information-dense chunks that reduce query-time retrieval needs. For high-query-volume applications with distributed knowledge, cross-document synthesis improves measurably over within-document optimization.
Paper Structure (45 sections, 2 figures, 6 tables)

This paper contains 45 sections, 2 figures, 6 tables.

Table of Contents

  1. Introduction
  2. RAG Systems and Their Limitations
  3. The Scattered Knowledge Problem
  4. Our Approach
  5. Related Work
  6. Traditional Partitioning
  7. Contextual Chunk Augmentation
  8. Advanced Intra-Document Methods
  9. Multi-Document Summarization
  10. Advanced RAG Methods
  11. The CDTA Methodology
  12. Design Principles
  13. The Six-Stage Pipeline
  14. Stage 1: Document Segmentation
  15. Stage 2: Topic Extraction
  16. ...and 30 more sections

Figures (2)

  • Figure 1: Multi-metric comparison on HotpotQA. CDTA (blue, outer) versus Fixed-Size baseline (red, inner) shows superior performance across all dimensions: Faithfulness (0.93 vs 0.64), Hit Rate@1 (0.88 vs 0.42), Answer Relevancy (0.90 vs 0.66), Context Recall (0.97 vs 0.72), and Context Precision (0.89 vs 0.58).
  • Figure 2: Faithfulness on HotpotQA as $k$ increases from 1 to 10. CDTA maintains consistently high performance (0.90-0.93) across all $k$ values, while traditional methods require more chunks to achieve competitive results. Error bands represent standard deviation across 3 runs.