Orchestration software is designed to allow modern IT systems to scale in a managed way by following configuration files written in concise and expressively powerful languages. As the configuration task itself increases in complexity to the point of requiring specialized knowledge it becomes sensible to employ AI agents powered by Large Language Models (LLMs) to generate, modify and translate such configuration files. The many advantages of LLM-enabled workflows (inherent knowledge of correct script syntax, ability to process complex natural language instructions, modern interfaces that allow them to autonomously fetch data and interact with their environment via external tools) make them an intuitive fit for managing systems, but their well-known shortcomings (such as hallucinations and knowledge cutoffs) may require some additional measures to make the system trustworthy enough to be deployed in real-world infrastructures. We study and implement various Retrieval Augmented Generation (RAG) techniques to fetch information from an external database of documentation files to aid an agent in correctly managing its virtual environment, producing configuration documents with correct syntax and semantics which conform to provided specifications. We also leverage the agent building framework to increase the determinism of the execution flow.
I software di orchestrazione sono progettati per consentire ai sistemi IT moderni di crescere in scala in maniera controllata seguendo file di configurazione scritti in linguaggi concisi ed espressivamente potenti. Con il crescere della complessita' del compito al punto da richiedere conoscenze specialistiche, diventa ragionevole l'impiego di agenti AI che includono Large Language Model (LLM) per generare, modificare e tradurre questi file di configurazione. I molti vantaggi di questi agenti (conoscenza intrinseca della sintassi corretta degli script, abilita' di processare istruzioni complesse in linguaggio naturale, interfacce moderne che gli consentono di richiedere dati e interagire con l'ambiente in maniera autonoma tramite tool esterni) li rendono candidati promettenti per gestire sistemi, ma i loro difetti ben noti (come allucinazioni e limite del periodo utile di fonti di conoscenza) potrebbero richiedere misure aggiuntive per rendere il sistema sufficientemente affidabile per essere impiegato in infrastrutture nel mondo reale. In questo paper studiamo e implementiamo varie tecniche di Retrieval Augmented Generation (RAG) per ottenere informazioni da un database esterno composto di file di documentazione per aiutare un agente a gestire correttamente il suo ambiente virtuale, producendo documenti di configurazione con sintassi corretta e semantica che si conforma alle specifiche assegnate. Sfruttiamo anche il framework di costruzione agenti per aumentare il determinismo del flusso di esecuzione.
Implementazione di metodologie RAG per traduzioni di file di configurazione tramite IA
GIACOMIN, MARCO
2025/2026
Abstract
Orchestration software is designed to allow modern IT systems to scale in a managed way by following configuration files written in concise and expressively powerful languages. As the configuration task itself increases in complexity to the point of requiring specialized knowledge it becomes sensible to employ AI agents powered by Large Language Models (LLMs) to generate, modify and translate such configuration files. The many advantages of LLM-enabled workflows (inherent knowledge of correct script syntax, ability to process complex natural language instructions, modern interfaces that allow them to autonomously fetch data and interact with their environment via external tools) make them an intuitive fit for managing systems, but their well-known shortcomings (such as hallucinations and knowledge cutoffs) may require some additional measures to make the system trustworthy enough to be deployed in real-world infrastructures. We study and implement various Retrieval Augmented Generation (RAG) techniques to fetch information from an external database of documentation files to aid an agent in correctly managing its virtual environment, producing configuration documents with correct syntax and semantics which conform to provided specifications. We also leverage the agent building framework to increase the determinism of the execution flow.| File | Dimensione | Formato | |
|---|---|---|---|
|
Giacomin_Elettra.pdf
accesso aperto
Dimensione
2.6 MB
Formato
Adobe PDF
|
2.6 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/107659