TopoChunker: Topology-Aware Agentic Document Chunking Framework

Xiaoyu Liu

cs.CLMar 19, 2026

TopoChunker: Topology-Aware Agentic Document Chunking Framework

Authors: Xiaoyu Liu

Organizations: Independent Researcher, Beijing, China

Abstract

Current document chunking methods for Retrieval-Augmented Generation (RAG) typically linearize text. This forced linearization strips away intrinsic topological hierarchies, creating ``semantic fragmentation'' that degrades downstream retrieval quality. In this paper, we propose TopoChunker, an agentic framework that maps heterogeneous documents onto a Structured Intermediate Representation (SIR) to explicitly preserve cross-segment dependencies. To balance structural fidelity with computational cost, TopoChunker employs a dual-agent architecture. An Inspector Agent dynamically routes documents through cost-optimized extraction paths, while a Refiner Agent performs capacity auditing and topological context disambiguation to reconstruct hierarchical lineage. Evaluated on unstructured narratives (GutenQA) and complex reports (GovReport), TopoChunker demonstrates state-of-the-art performance. It outperforms the strongest LLM-based baseline by 8.0% in absolute generation accuracy and achieves an 83.26% Recall@3, while simultaneously reducing token overhead by 23.5%, offering a scalable approach for structure-aware RAG.

Figures & tables

Figure 1: The overall architecture of TopoChunker. The dual-agent framework operates across three layers: Cognitive Perception (Inspector Agent for adaptive routing), Execution (Topological Pruner and SIR construction), and Audit (Refiner Agent for context disambiguation). These layers interact via a shared Structured Intermediate Representation (SIR) to form a closed-loop Diagnosis-Execution-Audit pipeline.

	GutenQA				GovReport
	R@3	D@3	R@5	D@5	R@3	D@3	R@5	D@5
Recursive (FC200)	8.23	7.38	8.62	7.84	32.25	28.40	36.25	29.81
Semantic Chunker	14.49	12.43	16.56	13.29	55.37	48.96	61.75	51.39
Proposition-Level	47.23	42.27	53.22	45.73	66.17	61.36	73.88	59.44
LumberChunker	62.58	56.56	66.93	58.35	81.38	74.88	85.20	75.55
MoC	63.85	57.40	68.45	58.80	82.15	75.50	86.85	76.10

Table 1: Domain-Specific RAG Performance. Performance is reported using Recall@

k

(R@

k

) and DCG@

k

(D@

k

). TopoChunker achieves highly competitive performance across both domains.

Figure 2: Generation Accuracy. A heatmap illustrating the generation accuracy of various chunking methods across the GutenQA and GovReport datasets. Darker shades indicate higher accuracy.

Generative LLM-based (DeepSeek-R1)
Method	Dataset (Topology Type)	Input Tokens	Output Tokens	CoT ( <think> )	Total Tokens
FC200 / Semantic Chunker	Both (Heuristic)	0	0	0	0
MoC (SLM 1.5B) [ 4 ]	Both (Dynamic Routing)	$\sim$ 14,500	$\sim$ 800	0 (No CoT)	$\sim$ 15,300
LumberChunker	GovReport (Structured)	18,500	1,500	22,000	42,000
TopoChunker (Ours)	GovReport (Structured $\rightarrow$ Path 1)	22,350	2,450	6,000	30,800
LumberChunker	GutenQA (Unstructured)	18,500	1,500	14,000	34,000

Table 2: Comprehensive Offline Generative LLM Token Consumption Analysis (Per 10,000 raw words

\approx

13,000 tokens). TopoChunker explicitly breaks down DeepSeek-R1’s CoT overhead.

Method	Recall@3	Acc. (%)	Token Cost
LumberChunker	81.38	79.0	1.65x
MoC (SLM Routing)	82.15	79.0	1.18x
Topo (w/o Refiner)	76.50	71.0	0.85x
w/o Inspector (Forced Static)
Rule-Only (Path 1)	77.80	73.5	0.80x
Semantic-Only (Path 2)	79.20	76.8	0.95x

Table 3: Ablation Study: Impact of Refiner and Inspector evaluated on the heterogeneous GovReport dataset.

Threshold ( $\theta$ )	Recall@3 (%)	Recall@5 (%)
250 (Aggressive)	82.50	85.90
500 (Optimal)	83.26	86.58
1000 (Diluted)	78.41	82.10
2000 (Diluted)	74.15	78.60

Table 4: Sensitivity analysis of the capacity threshold (

\theta

) evaluated on GovReport.

Document Chunking

Document chunking splits long or structured documents into coherent units for retrieval, question answering, and other language-model tasks.

See all 52 papers

QChunker: Learning Question-Aware Text Chunking for Domain RAG via Multi-Agent Debate
Mar 12, 2026Jihao Zhao, Daixuan Li, Pengfei Li +3Document Chunking Retrieval-Augmented Generation
School of Information, Renmin University of China Beijing, China · School of Smart Governance, Renmin University of China Suzhou, China · School of Information, BRAIN, Renmin University of China Beijing, China +1
Adaptive Chunking: Optimizing Chunking-Method Selection for RAG
Mar 26, 2026Paulo Roberto de Moura Júnior, Jean Lelong, Annabelle BlangeroDocument Chunking Retrieval-Augmented Generation
Ekimetrics, France
Query-Adaptive Semantic Chunking for Retrieval-Augmented Generation: A Dynamic Strategy with Contextual Window Expansion
Apr 29, 2026Mudit RastogiDocument Chunking Retrieval-Augmented Generation
Independent Researcher

Mar 12, 2026J/K move · Enter open · S save

QChunker: Learning Question-Aware Text Chunking for Domain RAG via Multi-Agent Debate

Jihao Zhao, Daixuan Li, Pengfei Li, Shuaishuai Zu+2

School of Information, Renmin University of China Beijing, China · School of Smart Governance, Renmin University of China Suzhou, China · School of Information, BRAIN, Renmin University of China Beijing, China+1

Document Chunking Retrieval-Augmented Generation RAG Evaluation

Open paper AskPDF arXiv

The effectiveness upper bound of retrieval-augmented generation (RAG) is fundamentally constrained by the semantic integrity and information granularity of text chunks in its knowledge base. To address these challenges, this paper proposes QChunker, which restructures the RAG paradigm from retrieval-augmentation to understanding-retrieval-augmentation. Firstly, QChunker models the text chunking as a composite task of text segmentation and knowledge completion to ensure the logical coherence and integrity of text chunks. Drawing inspiration from Hal Gregersen's "Questions Are the Answer" theory, we design a multi-agent debate framework comprising four specialized components: a question outline generator, text segmenter, integrity reviewer, and knowledge completer. This framework operates on the principle that questions serve as catalysts for profound insights. Through this pipeline, we successfully construct a high-quality dataset of 45K entries and transfer this capability to small language models. Additionally, to handle long evaluation chains and low efficiency in existing chunking evaluation methods, which overly rely on downstream QA tasks, we introduce a novel direct evaluation metric, ChunkScore. Both theoretical and experimental validations demonstrate that ChunkScore can directly and efficiently discriminate the quality of text chunks. Furthermore, during the text segmentation phase, we utilize document outlines for multi-path sampling to generate multiple candidate chunks and select the optimal solution employing ChunkScore. Extensive experimental results across four heterogeneous domains exhibit that QChunker effectively resolves aforementioned issues by providing RAG with more logically coherent and information-rich text chunks.

Abstract

Document Chunking

Explore similar work

QChunker: Learning Question-Aware Text Chunking for Domain RAG via Multi-Agent Debate

Adaptive Chunking: Optimizing Chunking-Method Selection for RAG

Query-Adaptive Semantic Chunking for Retrieval-Augmented Generation: A Dynamic Strategy with Contextual Window Expansion