Scholarlink

The digital transformation of Higher Education Institutions (HEIs) has led to an influx of unstructured digital documents. Administrative repositories, containing orders, resolutions, and manuals, are difficult to navigate using keyword-based engines. This is problematic for the Polish language, where morphological complexity often renders simple BM25 algorithms ineffective. This paper presents the design and evaluation of an Agentic RAG system, deployed on-premise at the Poznań University of Economics and Business (PUEB). The proposed system employs an agentic workflow capable of multi-step reasoning. A comparative study of two locally hosted Small Language Models (SLMs) - the generalist open-weights OpenAI-OSS-20B and the native Polish SpeakLeash Bielik-11B-v3.0 - reveals trade-offs between reasoning capabilities and linguistic fluency. Based on these findings, a Heterogeneous Hybrid Architecture is proposed, where the generalist model handles logic, while the native model serves as a stylistic refinement layer.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Heterogeneous Agentic RAG Architecture for Administrative Knowledge Management

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Adam Gałązkiewicz

摘要

The digital transformation of Higher Education Institutions (HEIs) has led to an influx of unstructured digital documents. Administrative repositories, containing orders, resolutions, and manuals, are difficult to navigate using keyword-based engines. This is problematic for the Polish language, where morphological complexity often renders simple BM25 algorithms ineffective. This paper presents the design and evaluation of an Agentic RAG system, deployed on-premise at the Poznań University of Economics and Business (PUEB). The proposed system employs an agentic workflow capable of multi-step reasoning. A comparative study of two locally hosted Small Language Models (SLMs) - the generalist open-weights OpenAI-OSS-20B and the native Polish SpeakLeash Bielik-11B-v3.0 - reveals trade-offs between reasoning capabilities and linguistic fluency. Based on these findings, a Heterogeneous Hybrid Architecture is proposed, where the generalist model handles logic, while the native model serves as a stylistic refinement layer.