ORCID
Nick Loghmani: 0009-0002-0844-8136
Document Type
Article
Date
Spring 4-1-2026
Keywords
Retrieval-augmented generation, Semantic entanglement, Vector embeddings, Agentic AI Document preprocessing, Topic segmentation, Information architecture, Knowledge representation, Distributed cognition, Operational context
Language
English
Disciplines
Artificial Intelligence and Robotics
Description/Abstract
Retrieval-Augmented Generation (RAG) systems deployed in agentic environments depend on the geometric properties of vector representations to retrieve contextually appropriate evidence for autonomous reasoning. When source documents conflate multiple topics within contiguous text regions, standard vectorization pipelines produce embedding spaces in which semantically distinct content occupies overlapping geometric neighborhoods — a condition we term semantic entanglement. This paper formalizes semantic entanglement as a model-relative measure of cross-topic overlap, defines an Entanglement Index (EI) as a quantitative proxy, and argues that higher EI is associated with reduced attainable Top-K retrieval precision under cosine similarity retrieval. We introduce the Semantic Disentanglement Pipeline (SDP), a four-stage context-conditioned preprocessing framework grounded in distributed cognition theory and our prior work on operational context in safety-critical systems. We evaluate the SDP on a real-world enterprise healthcare knowledge base of over 2,000 documents across approximately 25 operational sub-domains. Top-K retrieval accuracy improved from approximately 32% under fixed-token chunking to approximately 82% under SDP, while the mean Entanglement Index decreased from 0.71 to 0.14. The results support the claim that semantic entanglement is a distinct preprocessing failure mode that downstream tuning cannot reliably correct.
Recommended Citation
Loghmani, N. M. (2026). Semantic entanglement in vector-based retrieval: A formal framework and context-conditioned disentanglement pipeline for agentic RAG systems [Preprint]. SURFACE, Syracuse University. https://surface.syr.edu/etd/2255 (Also available at arXiv:2604.17677, https://arxiv.org/abs/2604.17677)
Source
submission
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 International License.
