Agentni sistem za priporočanje izbire jezikovnega modela

Madon, Matjaž (author), Žitnik, Slavko (mentor)

Izbira jezikovnega modela na portalu Hugging Face je lahko časovno potratna. Portal ponuja le iskanje po ključnih besedah in filtriranje po oznakah. V tem delu predstavimo rešitev AgentPick, agentni sistem, ki vrne kratek seznam predlaganih modelov in vsakega utemelji s podatki iz kataloga. Velik jezikovni model v omejeni zanki sam izbira med tremi orodji (semantično iskanje po modelnih karticah, strukturirano filtriranje po metapodatkih in vpogled v posamezno kartico) nad lokalno bazo podatkov 1.709 modelov pridobljenih iz portala Hugging Face. Na dvajsetih vprašanjih, ki pokrivajo šest vrst uporabniških zahtev, je agentni sistem AgentPick priporočila rangiral bolje od običajnega s poizvedovanjem obogatenega generiranja in od samostojnega jezikovnega modela brez dostopa do kataloga. Bolje je rangiral tudi od sistema z enim krogom (brez zanke), vendar je bila ta razlika statistično značilna le pri eni meri (nDCG@3).

• Language: Slovenian
• Keywords: agentni sistem, s poizvedovanjem obogateno generiranje besedil, obdelava naravnega jezika, velik jezikovni model
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2026
• PID: 20.500.12556/RUL-190022
• Publication date in RUL: 20.10.2026

Secondary language

• Language: English
• Title: Agent-based system for recommending the choice of a language model

Abstract

Choosing a language model on the Hugging Face Hub can be time-consuming. The portal offers only keyword search and filtering by tags. In this work we present AgentPick, an agent-based system that returns a short list of suggested models and justifies each one with catalogue data. Within a bounded loop, a large language model picks for itself among three tools (semantic search over model cards, structured filtering over metadata and inspection of a single card) over a local database of 1,709 models collected from the Hugging Face Hub. On twenty questions covering six types of user requests, AgentPick ranked its recommendations better than the ordinary retrieval-augmented generation and a standalone language model with no access to the catalogue. It also ranked better than a single-round system (without the loop), but the difference was statistically significant for only one metric (nDCG@3).

• Keywords: agentic system, retrieval augmented generation, natural language processing, large language model