Dynamic Knowledge Lifecycle Management for Academic Advising RAG Systems

ACL ARR 2026 January Submission6569 Authors

Keywords: Retrieval-Augmented Generation, Knowledge Base Management, Vector Database, Academic Advising

Abstract: Retrieval-Augmented Generation (RAG) is the standard architecture for grounding Large Language Models (LLMs) in domain-specific facts. However, current literature largely treats the underlying vector store as a static artifact, which poses significant risks in high-velocity domains like academic advising. In this paper, we present a Dynamic Knowledge Base Management System (DKBMS) evaluated on a corpus of 10,000 university documents. We formalize an end-to-end pipeline that leverages metadata filtering to ingest new documents and surgically remove outdated embeddings. While underlying vector engines support CRUD operations, we demonstrate a unified workflow that reduces knowledge update latency by 99.8% compared to the static snapshot approach often used in academic deployments. Crucially, we analyze retrieval stability to prove that these operations do not degrade the semantic integrity of the remaining knowledge base.

Paper Type: Short

Research Area: Retrieval-Augmented Language Models

Research Area Keywords: RAG, Vector Databases, Knowledge Management

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 6569