EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

A Mathematically Grounded Metric Framework for Evaluating Semantic Chunking in RAG: Quantifying Coherence and Contextual Integrity

Raghvendra Kumar Jha, Mamta Arora, Dipali Bansal

Computer and Decision Making: An International Journal · 09.11.2026 · 3 (8S) · 413-420

DOI: 10.1922/comdem.v3i8s.19183 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Computer and Decision Making: An International Journal, online erschienen am 09.11.2026. Verfasst von Raghvendra Kumar Jha und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Retrieval-Augmented Generation (RAG) systems fundamentally rely on splitting long documents into small text chunks to enable dense and sparse evidence retrieval. Evaluating the quality of chunk partition has traditionally been confused with the outcomes of downstream generation (e.g., end-to-end LLM answer correctness or RAGAS scores). This paradigm has no direct diagnostic feedback on the placement of the partition boundaries and has a significant amount of computational overhead. In this paper, we propose a mathematical model to evaluate the intrinsic chunk partition, which consists of Geometric Partition Cohesion (GPC) and Contextual Manifold Alignment (CMA). GPC is based on the idea of Intra-Chunk Semantic Density (ISD) and Inter-Chunk Boundary Isolation (IBI) to capture vector-space geometry, whereas CMA is based on the measurement of preserving global document geometry using centered principal component subspace projection. The composite metric has been assessed using experiments on publicly available datasets such as QASPER, WikiSection, and CoheSentia. This paper defines a mathematically grounded candidate metric framework for semantic chunk coherence and document-context alignment. The experiments provide preliminary evidence that the metrics respond to partition controls and topic structure; human construct validity remains to be established.

Zitieren

APA

Jha, R. K., Arora, M., & Bansal, D. (2026). A Mathematically Grounded Metric Framework for Evaluating Semantic Chunking in RAG: Quantifying Coherence and Contextual Integrity. Computer and Decision Making: An International Journal, 3(8S), 413-420. https://doi.org/10.1922/comdem.v3i8s.19183

Vancouver

Jha RK, Arora M, Bansal D. A Mathematically Grounded Metric Framework for Evaluating Semantic Chunking in RAG: Quantifying Coherence and Contextual Integrity. Computer and Decision Making: An International Journal. 2026;3(8S):413-420. doi:10.1922/comdem.v3i8s.19183

BibTeX
@article{Jha2026,
  author = {Jha, Raghvendra Kumar and Arora, Mamta and Bansal, Dipali},
  title = {A Mathematically Grounded Metric Framework for Evaluating Semantic Chunking in RAG: Quantifying Coherence and Contextual Integrity},
  journal = {Computer and Decision Making: An International Journal},
  year = {2026},
  volume = {3},
  number = {8S},
  pages = {413--420},
  doi = {10.1922/comdem.v3i8s.19183}
}
Kostenlos starten →