Kompass: An Agentic Retrieval-Augmented Generation System for Personalized German University Admissions
The German higher education system offers tuition-free, world-class education, yet Indian students face a labyrinth of bureaucratic complexity—including the APS certificate, Uni-Assist, Blocked Accounts, and Anabin degree recognition—that is poorly served by generic Large Language Models. This paper presents Kompass, a public-facing Agentic Retrieval-Augmented Generation (RAG) system designed as an expert advisor for German admissions. Kompass employs a Hybrid Database Architecture (relational SQL + pgvector) to ground its responses in 6,000+ university programs and 100+ verified policy documents. A deterministic ”Pre-Flight Execution” router classifies user queries and invokes specialized Python tools for math, SQL filtering, OCR-based document parsing, and live web search, eliminating LLM hallucinations. Evaluation under adversarial testing demonstrates a 0% hallucination rate on fabricated universities and a 0% UI crash rate under API stress, with end-to-end response latency under 3.5 seconds. This work serves as a proof-of-concept for low-cost, domain-specific Agentic RAG systems across sectors. Index Terms—Agentic RAG, Large Language Models, German Admissions, Hybrid Database, Hallucination Mitigation.
Authors
- Ujjwal Bisht
Institutions
- Guru Gobind Singh Indraprastha University (IN)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24265011
- Primary Topic
- Topic Modeling
- Type
- preprint